Deep learning for small-molecule drug discovery: From molecular design to clinical translation

Abstract

Recent advances in artificial intelligence (AI) are increasingly transforming drug discovery, offering new approaches to accelerate the identification and optimization of therapeutic candidates. Deep learning (DL) methods have shown strong potential for generative modeling and molecular property prediction, enabling the design of compounds with tailored pharmacological profiles. In this review, we synthesize recent AI-driven progress in drug discovery, highlighting both achievements and ongoing challenges. We discuss key technical aspects underpinning DL-based models, including the use of curated molecular databases, molecular descriptors, and standardized evaluation strategies. We examine representative architectures and illustrate how they have been applied to molecular generation, binding affinity prediction, and multi-modal integration of ligand and protein data. In addition, we provide a comprehensive analysis of AI-enabled small-molecule drugs, including discovered and repurposed molecules in clinical trials, and explore the intellectual property information and chemical structures of these compounds. We demonstrate that AI-native companies are evolving pharmaceutical pipelines by integrating these tools at various stages of development. We also address the regulatory framework of agencies such as the U.S. Food and Drug Administration (FDA) and the European Medicines Agency (EMA) as they adapt to integrate AI innovations. Finally, we outline persistent limitations, such as data inconsistency, model interpretability, and the gap between benchmark performance and real-world applicability, which remain critical barriers to widespread adoption. By consolidating technical advances and open challenges, this review aims to provide a balanced perspective on the role of AI in reshaping modern drug discovery.

Graphical abstract

Highlights

• •Overview of cutting-edge AI models for molecular design and property prediction.
• •Presentation of key data sources, descriptors, and evaluation benchmarks.
• •Description of the evolution of models for molecular property prediction.
• •Enumeration of promising AI-generated small molecules progressing in clinical trials.
• •Survey of the regulatory framework responses to AI-driven drug development tools.

1. Introduction

The pharmaceutical drug development process remains one of the most challenging endeavors in modern medicine, characterized by inefficiencies and high failure rates. Despite the substantial investment made in research, approximately 90% of drug candidates fail to advance beyond phase 1 clinical trials [1]. As a result, the discovery and development process of a drug until approval for clinical use can take up to 10–15 years with an average cost of more than $1 billion [2]. These challenges arise from the intricate complexities of identifying molecules with tailored properties essential for interacting with specific biological targets and achieving the desired therapeutic functions. Given these mounting costs and failure rates, computational methods have emerged as a natural ally.

One of the main requirements for successful drug development is the design of molecules capable of selectively interacting with a target under physiological conditions. Optimizing the pharmacokinetic properties of a drug is crucial to ensure its high bioavailability through absorption and distribution [3]. Such an optimization process should carefully consider various molecular properties, including the molecular weight, lipophilicity, and conformational flexibility [3]. In addition, the design of the molecules must consider structural elements such as hydrogen bond donors and acceptors, while maintaining the synthesis feasibility in the laboratory. Minimizing toxicity, ensuring the efficient elimination, and avoiding the formation of harmful metabolites are also essential requirements. Analysis of cases of drug development failure reveals that the primary obstacles to market approval primarily stem from insufficient clinical efficacy, excessive toxicity, and poor drug-like properties, which prevent the compounds from advancing to later stages of development [1]. These factors underscore the interdisciplinary nature of drug discovery and the need for methods capable of simultaneously optimizing multiple, often competing, molecular properties. Despite these intricate requirements, the majority of computational approaches are unable to effectively capture the complex multi-objective landscape required to systematically develop promising drug candidates. This limitation poses a significant challenge to the translation of theoretical computational methods into practical pharmaceutical innovations.

Advances in artificial intelligence (AI) have catalyzed important changes in drug discovery and development. In particular, deep learning (DL) technologies show potential to accelerate the identification of promising drug candidates and reduce some of the time and cost of traditional pipelines; however, their performance remains highly dependent on the availability of high-quality, well-curated data, which is often a limiting factor in pharmaceutical research. Nevertheless, these computational approaches enable an efficient navigation through the vast chemical space of synthesizable compounds, enabling a more targeted exploration of therapeutic opportunities. The distinct advantage of DL methods lies in their ability to extract meaningful patterns from complex, heterogeneous datasets. By seamlessly integrating different types of data, including biochemical properties, biological activities, structural features, and pharmacological profiles, these systems can develop a comprehensive understanding of the drug development task.

In this review, we critically examine the demonstrated effectiveness of DL-based methods in drug development, establishing their current practical utility in the pharmaceutical research setting. First, we describe the technical aspects crucial for implementing computational methods in drug discovery. This includes an analysis of the data sources that provide rich molecular context, the descriptors that translate chemical complexity into interpretable computational language, and the most used evaluation metrics for assessing model performance and predictive accuracy. The analysis of the methodologies will focus on the prediction of molecular properties and the generation of molecules with pharmacological potential. Each methodological advance is critically illustrated with practical application examples, highlighting key innovations and significant contributions to the current state-of-the-art in AI-driven molecular design. Second, we will explore the pipeline growth of AI-native drug discovery companies through an exhaustive analysis of the AI-generated small molecules and AI-repurposed molecules. This analysis includes detailed information about each compound, including the developing company, the underpinning AI technology platform, and their progress through clinical trials (both past and ongoing). This information offers a comprehensive view of the current state of AI applications in drug discovery protocols. Furthermore, we address the ambiguous regulatory considerations surrounding AI in drug development. We provide a comprehensive analysis of how regulatory agencies, particularly the U.S. Food and Drug Administration (FDA) and the European Medicines Agency (EMA), are addressing this emerging field and what measures are being considered to facilitate the market entry of AI-enabled drugs.

The ultimate goal of this review is to provide a transparent and comprehensive description of the critical components of DL-based models, establishing a robust foundation for their broader adoption in pharmaceutical research. Additionally, we address and explore the ongoing transformation of the drug discovery field guided by the use of AI, combining information such as the AI-generated small molecules, the AI-technology platforms used by the companies, the target classes with which those compounds interact, and their therapeutic areas. This comprehensive analysis provides an insightful perspective into the current state and potential future for AI-driven drug discovery.

2. Fundamentals of DL in drug discovery: data, descriptors, and benchmarking strategies

Over the last decade, we have witnessed the proliferation of AI, particularly DL, methodologies that have influenced many critical aspects of drug discovery, including molecular generation, property prediction, drug-target interaction (DTI) prediction, drug repurposing, and synthesis route planning [4]. This evolution stems from the concurrent advances in methodological approaches, computational power, and data availability. The fundamental distinction between AI-driven approaches and traditional drug design lies in their scope: while conventional methods typically restrict their search to existing chemical libraries, AI-guided drug design enables researchers to systematically explore the vast chemical space of synthesizable molecules [5,6]. This evolution started with the introduction of quantitative structure-activity relationship (QSAR) models and has since evolved into a powerful approach that enables efficient exploration of the vast search space for potential therapeutic solutions. While AI can be applied to all stages of drug discovery, this state-of-the-art review focuses on lead identification, which encompasses two primary tasks: prediction of molecular properties and de novo targeted molecule generation employing DL models. Lead identification is a pivotal stage because it is typically one of the most resource- and time-intensive steps in the pipeline, and AI-driven approaches have the potential to accelerate the selection of viable candidates while reducing experimental costs. Fig. 1 illustrates how AI is typically applied in drug design and outlines the topics we will discuss in this review.

Fig. 1.

Description of the components associated with the computational generation of molecules with pharmacological potential. LIT-PCBA: literature-based PubChem BioAssay; SMILES: Simplified Molecular Input Line Entry System; InChI: International Chemical Identifier; RNNs: recurrent neural networks; VAEs: variational auto-encoders; GANs: generative adversarial networks; GNNs: graph neural networks; QSAR: quantitative structure-activity relationship (SAR); SVM: Support Vector Machine; CNNs: convolutional neural networks.

2.1. Data sources

With the rise of high-throughput screening (HTS) and biological assays, a wide range of molecular data resources have become publicly available, thus building the foundation for AI-driven drug discovery. The large volume and diversity of these datasets provide the extensive labeled examples that DL models require to learn complex structure-activity relationships (SARs) and make accurate predictions.

Databases like the chemical database of bioactive molecules maintained by the European Bioinformatics Institute (EBI), part of the European Molecular Biology Laboratory (EMBL), PubChem, and ZINC provide millions of annotated compounds, including bioactivity, pharmacokinetics, and drug-likeness properties, supporting both structure-based and ligand-based modeling. While ChEMBL is manually curated, ensuring higher data quality, PubChem aggregates data from hundreds of sources without curation, requiring additional preprocessing steps [7,8]. ZINC, on the other hand, focuses on purchasable compounds with 3D structures for virtual screening [9].

Beyond these general-purpose repositories, specialized datasets such as the Collection of Open Natural Products (COCONUT), DrugBank (drug-drug interactions), DGIdb (drug-gene interactions), and Side Effect Resource (SIDER)/Observational Medical Outcomes Partnership-derived Frequency Table of Side Effects (OFFSIDES) (adverse reactions) enable targeted applications in chemogenomics and pharmacovigilance.

Importantly, ligand-target complex databases, such as PDBbind, BindingDB, and BioLiP, play a crucial role in structure-based drug design, offering experimentally determined or computationally derived binding affinity data. An interesting alternative is the literature-based PubChem BioAssay (LIT-PCBA), a curated benchmark dataset for DL and virtual screening models. LIT-PCBA includes over 400,000 inactive and 7800 active compounds, derived from confirmatory dose-response PubChem assays, to better simulate the noise and class imbalance typical of real HTS screens. By reflecting these experimental challenges, the dataset provides a more robust training ground for DL models, improving their ability to generalize beyond idealized benchmarks and enhancing transfer learning (TL) to real-world screening campaigns. This resource is particularly valuable for both ligand-based and structure-based modeling, providing high-quality ligand-target pairs linked to X-ray structures for each protein. This enables a robust framework for benchmarking AI models in conditions that closely reflect real-world screening challenges. These tools are essential for training models that predict molecular docking, binding poses, and interaction energies, further enhancing the accuracy of DTI predictions.

Complementary to compound- and ligand-target-centric resources, additional databases capture mechanistic, toxicological, genetic, and systems-level information. The Comparative Toxicogenomics Database (CTD) integrates curated chemical-gene/protein-disease interactions, enabling the identification of molecular mechanisms underlying chemical toxicity and disease etiology. The Therapeutic Target Database (TTD) provides detailed information on known and explored therapeutic protein and nucleic acid targets, their associated diseases, and corresponding drugs, serving as a valuable reference for target identification and validation. The Online Mendelian Inheritance in Man (OMIM) database offers a comprehensive, curated catalog of human genes and genetic disorders, linking genotype to phenotype with detailed clinical descriptions and molecular information. Furthermore, protein-protein interaction (PPI) databases, such as Search Tool for the Retrival of Interacting Genes/Proteins (STRING) or Biological General Repository for Interaction Datasets (BioGRID), facilitate the incorporation of molecular network data, supporting the prediction of DTIs, off-target effects, and pathway-level perturbations. Together, these complementary resources enable a multi-layered view of disease biology, strengthening the robustness and translational potential of AI-based drug discovery workflows.

Together, these diverse databases empower AI models to generalize across chemical and biological spaces, supporting multiple stages of the drug development pipeline. In addition to selecting the type of database best suited to the specific problem, it is crucial to consider the popularity and credibility of the available resources, as these factors are essential for ensuring the validity, reliability, and overall robustness of the resulting DL model.

2.2. Molecular descriptors

The transformation of molecular data from raw chemical formulas into computer-interpretable notation is a critical step in computational drug discovery. The process of selecting the most suitable molecular descriptor depends on the characteristics of the computer-aided drug design (CADD) method being used. For example, QSAR models often rely on compact but informative fingerprints to capture key chemical features, whereas graph neural networks (GNNs) can directly leverage richer structural representations from molecular graphs. Since AI-based models are built on mathematical frameworks, molecular descriptors must translate chemical information into numerical representations. These descriptors should capture the most important structural and functional characteristics of compounds without becoming overly complex to model. In practice, however, selecting the most appropriate descriptor is often a trial-and-error process rather than a strictly rule-based decision, as the optimal choice depends on both the dataset and the modeling approach. This balance between comprehensibility, complexity, and informativeness has led to the development of several types of descriptors that, as shown in Fig. 2, can be grouped according to their dimensionality.

Fig. 2.

Different types of molecular encoding grouped by dimensionality. 0D: zero-dimensional; SMILES: Simplified Molecular Input Line Entry System.

The most straightforward descriptors are zero-dimensional (0D) or constitutional, which are derived directly from the chemical formula. Think of these as molecular headcounts: they indicate what atoms are present in the structure, using metrics such as molecular weight, total atomic properties, or atom-type counts, but not how those atoms are connected. 1D representations encode crucial structural information, incorporating knowledge about sub-structural fragments and functional groups (FGs). These representations provide a concise yet informative strategy to describe molecular structures, facilitating efficient computational analysis and machine learning (ML) applications in drug discovery. Simplified Molecular Input Line Entry System (SMILES), fingerprints, and International Chemical Identifier (InChI) strings are examples of 1D representations that capture the topology and connectivity of atoms within a molecule. Fingerprint vectors indicate the presence or absence of specific substructures within a molecule. These are typically represented as binary vectors of fixed length, where each bit (0 or 1) corresponds to a particular structural feature. A prime example of this approach is the extended connectivity fingerprints (ECFPs) generated using the Morgan algorithm [10]. ECFPs capture various molecular features and have proven highly effective in various cheminformatics applications, including virtual screening and QSAR modeling. These notations allow researchers to identify important structural motifs and assess molecular similarity based on shared sub-structures. As a result, by leveraging 1D representations, computational models can efficiently process large chemical libraries, extract relevant features, and predict molecular properties or biological activities. This approach is particularly useful in virtual screening, QSAR modeling, and other computational drug discovery techniques where rapid analysis of vast chemical spaces is essential.

To provide comprehensive information about atomic connectivity and adjacency, researchers developed 2D or topological descriptors. These descriptors are typically represented by molecular graphs, where nodes represent atoms and edges represent bonds between atoms. This method allows for a direct visual representation of molecular structure, facilitating intuitive interpretation and analysis of chemical compounds. Additionally, molecular graphs can capture complex structural features and relationships, making them particularly useful for tasks such as property prediction. Hence, this notation is used to encode detailed topological data in a readily processable form.

3D descriptors become essential when the goal is to represent molecules in 3D space and capture features such as surface area, volume, and steric properties. While these descriptors are more complex to compute and analyze, they provide a richer representation of molecular structure and behavior. These descriptors can represent the molecule in a Cartesian space by assigning spatial coordinates (x, y, and z) to each atom in the connection table [11]. Alternatively, features such as bond length, bond angle, and torsion angle can be used to characterize the relative position of each atom in 3D space. However, these alternatives do not consider the electron cloud surrounding each atom, which impacts molecular properties and interactions. The descriptors that consider this aspect are molecular surfaces in which a molecule is represented as a closed surface that delimits the volume it occupies. Electrostatic properties or hydrophobicity potential at particular locations are associated with this descriptor type. This 3D representation provides crucial information about molecular shape, size, conformational flexibility, and the distribution of electrostatic and hydrophobic/hydrophilic regions. There are also 4D descriptors that involve monitoring dynamics such as molecular motions, conformational changes, or time-dependent interactions. This additional temporal dimension enables encoding complex physical properties essential for drug discovery, particularly those governing DTIs at binding sites.

Despite the wide range of descriptors available, SMILES and molecular graphs have emerged as the most effective representations for molecules in property prediction and de novo molecular design tasks.

SMILES encodes molecules as American Standard Code for Information Interchange (ASCII) strings based on intuitive rules: atoms are represented by their symbols, with implicit single bonds and explicit symbols for double (=) and triple (#) bonds. Additional syntax handles branching, rings, aromaticity, charges, and stereochemistry. However, a major limitation is that SMILES strings are not inherently unique: the same molecule can be represented by multiple valid strings, which complicates downstream tasks such as data augmentation or model training. Variants such as canonical SMILES aim to address this issue by providing a standardized representation for each molecule, while isomeric SMILES preserve stereochemical information [12]. Initially used in ML through one-hot encoding, modern models treat SMILES as sequences, making them suitable for transformer-based and sequence-to-sequence architectures that learn molecular syntax and dependencies [13]. Tools like RDKit and OpenBabel facilitate SMILES parsing, generation, and conversion, ensuring their continued relevance despite not encoding 3D structures [14,15].

In parallel, molecular graphs provide a natural and expressive representation where atoms are nodes and bonds are edges. These graphs are defined by node features (e.g., atomic number and hybridization) and edge features (e.g., bond type and stereochemistry), often complemented by an adjacency matrix capturing atomic connectivity [16]. This structure enables DL models to extract both local and global molecular features efficiently. Although computationally more demanding and requiring consistent formatting, graph-based representations preserve detailed structural information and are well-suited to GNNs and other DL architectures.

The selection of the most appropriate representation is largely influenced by the specific requirements of the computational task and by the nature of the chemical space under study. For example, descriptors that work well for small synthetic molecules may be inadequate for complex natural products, which often require richer structural or 3D-aware representations.

2.3. Evaluation metrics

The field of AI-aided drug design is evolving rapidly, with new research emerging at an unprecedented pace. A robust and rigorous evaluation is necessary to identify the most promising approaches and provide insights to guide future work. In this sense, researchers widely employ standardized metrics and benchmark platforms for this purpose. In this analysis, we focus on the evaluation of two key areas of AI-driven drug discovery: property prediction models and de novo molecular design algorithms. Table S1 provides a comprehensive overview of the evaluation metrics and benchmarks commonly used in these domains.

The standard ML benchmark for property prediction is the MoleculeNet tool, implemented by Wu et al. [17], which provides a set of curated datasets and benchmarks designed to systematically evaluate ML methods. This open-source implementation platform provides standardized pipelines for data division and processing as well as a diverse set of molecular datasets in areas such as quantum mechanics (QM7, QM7b, QM8, and QM9), physical chemistry (Estimated SOLubility (ESOL), Free Solvation Energy (FreeSolv), and lipophilicity), biophysics (PCBA and Maximum Unbiased Validation (MUV)), physiology (molecule activity against human immunodeficiency virus (HIV), inhibitors for $\mathrm{\beta }$-secretase 1 (BACE), and blood-brain barrier penetration (BBBP)) and cellular (Toxicology in the 21st Century (Tox21) and Toxicity Forecaster (ToxCast)). This framework supports several molecular representations (ECFP), coulomb matrix, or graph convolution) and provides results for several baseline ML models using evaluation metrics for regression and classification tasks. One of MoleculeNet's primary advantages is its promotion of reproducible research through established data splits for training, validation, and testing. These splits can be generated randomly or based on molecular scaffolds to ensure diversity across partitions. This standardization allows for fair comparisons between different models and approaches.

Although MoleculeNet is the cornerstone of evaluating molecular ML models, other benchmark frameworks are tailored to specific tasks. Chemprop compares traditional molecular descriptors, such as ECFPs, with descriptors learned using a neural network. By providing a unified framework for both conventional and learned descriptors, Chemprop facilitates a more comprehensive assessment of model performance across various cheminformatics applications. This toolbox implements a DL model that provides a straightforward way of predicting multiple molecular properties, reactions, and atom/bond-level properties [18].

The evaluation of generative models in molecular design focuses on the relevance of the newly generated compounds. Key evaluation criteria include chemical validity, diversity, uniqueness, novelty, drug-likeness, synthetic accessibility, and compliance with target-specific properties, such as binding affinity [19].

Validity measures the proportion of chemically plausible molecules generated, reflecting the model's understanding of fundamental chemical rules. Diversity, uniqueness, and novelty assess the breadth of chemical space exploration: diversity captures variability among molecules, uniqueness quantifies how many are distinct, and novelty evaluates how dissimilar they are from the training data, indirectly indicating the creative capacity of the model.

Crucially, drug-likeness and synthetic accessibility are central to assessing real-world applicability. The quantitative estimate of drug-likeness (QED) provides a condensed score based on physicochemical properties typical of approved drugs, offering a nuanced and continuous measure that goes beyond binary rules like Lipinski's rule of five [20]. High QED values indicate promising pharmaceutical profiles. Equally important is the synthetic accessibility score (SAS), which estimates the feasibility of synthesizing a compound in a laboratory. Developed by Ertl et al. [21], SAS integrates fragment occurrence frequencies and structural complexity penalties, making it a practical and widely used proxy for medicinal chemistry tractability. Molecules with favorable QED and SAS scores are far more likely to transition successfully from computational generation to experimental validation.

Lastly, the satisfaction of desired properties determines whether the generated molecules meet specific target criteria, such as binding affinity, solubility, or other application-specific requirements. The primary goal of molecular generative models is to design biologically active compounds that selectively target specific proteins or biological pathways. In practice, success is often assessed by predicted binding affinity for the intended target. However, these predictions are highly model-dependent and can vary with input quality and target complexity, meaning that affinity scores should be interpreted as approximations rather than absolute ground truth. This makes desired property satisfaction an important, yet inherently uncertain, evaluation metric.

A notable example of the benchmark is GuacaMol, introduced in 2019 by Brown et al. [22]. This platform offers a comprehensive framework for assessing generative models, encompassing distribution-learning tasks and goal-directed benchmarks. GuacaMol's standardized evaluation metrics enable researchers to objectively compare diverse approaches to molecular generation, fostering transparency and reproducibility. By providing a common set of challenges and evaluation criteria, GuacaMol allows for a more nuanced understanding of each model's strengths and limitations [22]. By comprehensively evaluating these aspects, researchers can determine generative models' effectiveness and practical utility in creating novel, diverse, and potentially valuable chemical compounds for drug discovery applications. Multiple metrics are used to assess the distribution of the generated molecules. These include the proportions of chemically valid molecules (validity), non-duplicate molecules (uniqueness), and molecules not present in the training set (novelty). Furthermore, the Kullback-Leibler (KL) divergence is used to compare physical property distribution proximity between generated and training set molecules. The Fréchet ChemNet distance (FCD) measures the distance between two sets of molecules using their hidden representations in “ChemNet”, a multitasking neural network designed to predict biological activities [23]. Goal-directed benchmarks evaluate a model's ability to generate molecules with specific desired properties [22]. Some examples of this individual assessment of molecules include similarity to a target compound, presence of particular sub-structures, or optimization of specific biological or physicochemical properties. A key advantage of this framework is the availability of results for the different experiments for several baseline DL-based models for different experiments. MOSES, introduced by Polykovskiy et al. [19] in 2020, is another benchmark suited for molecular generation models, focusing on drug-like molecules. Its evaluations are more centered on distributional learning, assessing metrics such as internal diversity, fragment similarity, and scaffold similarity, in addition to validity, uniqueness, novelty, and FCD. Internal diversity is calculated by averaging the Tanimoto distance between pairs of molecules. As such, the Moses benchmark effectively evaluates the overall generation of drug-like molecules. Furthermore, it provides baseline results for several state-of-the-art representative generative models such as recurrent neural network (RNN), variational autoencoder (VAE), adversarial autoencoder (AAE), Junction Tree Neural Network (JTNN)-VAE, and LatentGAN. This comprehensive set of baseline results allows researchers to compare their novel approaches against established methods and promote reproducible work, facilitating progress in the field of AI-based molecular generation [[24], [25], [26], [27], [28]].

Despite the development of various evaluation metrics and benchmark platforms, Renz et al. [29] demonstrated a significant limitation in current benchmarking approaches. They found that both distribution-learning benchmarks and goal-directed tasks can yield excellent performance in theory, yet the generated compounds may have little to no practical interest. This insight highlights the potential incongruity between theoretical performance and real-world applicability in molecular generation models [29]. In response to this challenge, complementary evaluation strategies have emerged to bridge the gap between theoretical performance and practical utility. Two notable examples of evaluation strategies include docking-based assessments, which simulate the interaction of generated molecules with target proteins, and chemical space coverage analysis, which examines how thoroughly the model explores and populates specific regions of chemical space [30]. In docking evaluations, metrics such as binding affinity scores, ligand efficiency, pose root mean square deviation, and interaction fingerprints are commonly used to quantify binding strength, structural alignment, and interaction quality with the target binding site [31]. Zhang et al. [32] further proposed evaluation based on the distribution of FGs and ring systems, offering valuable structural baselines that complement physicochemical and bioactivity-oriented assessments. These structural analyses are critical for ensuring the chemical relevance and diversity of generated compounds. Table 1 summarizes the advantages and disadvantages associated with each tool and metric mentioned above so that it is possible to obtain an overview of how the evaluation of these models works.

Table 1.

Comparison of the advantages and disadvantages of different tools and metrics used in computational drug discovery.

Tool	Focus	Advantages	Disadvantages
MoleculeNet	Property prediction (classification and regression)	•Curated datasets across multiple domains •Supports multiple molecular representations for reproducibility	•Not tailored for specific niche applications beyond included sets
MoleculeACE	Activity cliff prediction in QSAR datasets	•Focuses on a known challenge in QSAR •Highlights model weaknesses in subtle chemical changes	•Narrow scope (activity cliffs only)
Chemprop	Property prediction with learned and traditional descriptors	•Compares ECFP with learned descriptors in unified framework •Implements a DL model for molecular, reaction, and atom/bond-level predictions	•Less focused on benchmark diversity than MoleculeNet
GuacaMol	De novo molecular generation	•Covers distribution-learning and goal-directed tasks •Standardized evaluation metrics and baseline results for multiple models	•Performance may not correlate with practical applicability •Limited inclusion of biological evaluation
MOSES	De novo molecular generation	•Focus on drug-likeness •Includes distributional metrics and baseline results for multiple generative models	•Does not directly assess biological activity or target-specific properties
Validity	Generative models	•Ensures molecules obey chemical rules •Quick to compute	•Says nothing about usefulness or biological relevance
Diversity	Generative models	•Measures chemical space exploration breadth •Encourages variety in generation	•High diversity can lead to loss of focus on target properties
Uniqueness	Generative models	•Rewards generation of non-duplicate molecules	•Does not ensure novelty relative to training data
Novelty	Generative models	•Encourages exploration beyond training data	•Excessive novelty may reduce synthesizability or activity
QED	Drug-likeness	•Continuous score reflecting real-world approved drugs' physicochemical profiles	•Does not account for synthetic feasibility or specific target requirements
SAS	Synthetic accessibility	•Estimates ease of synthesis based on fragment frequencies and complexity penalties	•Approximation; may misjudge feasibility for complex but synthesizable molecules
FCD	Distribution similarity	•Compares generated vs. reference sets in learned feature space	•Dependent on ChemNet's training
KL divergence	Property distribution similarity	•Quantifies how generated property distributions match reference data	•Sensitive to outliers; assumes comparable distributions
Docking-based evaluation	Target-specific binding	•Provides structural insight and estimates pose quality	•Computationally expensive •Docking scores may not correlate perfectly with experimental binding
Chemical space coverage analysis	Exploration of chemical space	•Ensures generated molecules fill underexplored areas •Identifies functional coverage gaps	•Requires well-defined reference chemical space
FG and ring system distribution analysis	Structural diversity	•Complements physicochemical metrics with structural insight	•Limited in predicting biological activity
Wet-lab feedback	Biological relevance	•Directly tests real-world applicability •Closes the loop between design and experimental proof	•Time-consuming, costly, and resource-intensive

QSAR: quantitative structure-activity relationship (SAR); ECFP: extended-connectivity fingerprints; DL: deep learning; QED: quantitative estimate of drug-likeness; SAS: synthetic accessibility score; FCD: Fréchet ChemNet distance; KL: Kullback-Leibler; FG: functional group.

Although the availability of diverse metrics and benchmarks has greatly advanced the field, their practical use often lacks standardization. Researchers are often left without clear guidance on which metrics should be prioritized for a successful drug discovery scenario, complicating model evaluation and comparison. For instance, validity, uniqueness, and novelty should be prioritized when the aim is exploration of chemical space or the generation of entirely new scaffolds. High novelty is particularly relevant in antitumor or antiviral drug discovery, where existing scaffolds may already face resistance. However, when the goal is to ensure translational viability, QED and SAS should be prioritized, as they reflect the likelihood of a molecule being drug-like and synthetic. For example, SAS is especially critical when designing kinase inhibitors or complex anticancer agents, where chemical tractability is often a bottleneck. Conversely, when the goal is to contribute to the state of the art at the methods level, it is imperative to provide distributional comparisons of metrics such as FCD and KL-divergence against known chemical libraries, ensuring models do not merely reproduce training data but learn meaningful latent representations. Finally, target-specific properties, such as binding affinity and absorption, distribution, metabolism, excretion, and toxicity (ADMET) predictions, should be prioritized when the application is directed at drug design, such as antibiotics or specific ligands.

In each optimization task, it is necessary to choose the most appropriate metrics and consider some trade-offs. Typically, maximizing novelty often reduces validity and synthetic accessibility, as highly novel molecules may violate basic chemical constraints. Also, high drug-likeness does not guarantee novelty, since many drug-like molecules cluster around known scaffolds. Regarding models, GuacaMol and MOSES benchmarks show that transformers and diffusion models often achieve higher diversity and novelty, while GNNs tend to perform better in validity and property conditioning. The most correct practice is to use a balanced multi-metric evaluation strategy tailored to the therapeutic context and complement this computational evaluation with experimental validation to ensure the biological and pharmacological relevance of AI-generated molecules. A particularly promising direction is the integration of wet-lab experimental feedback, which can validate model predictions and guide iterative improvement. Examples include in vitro binding assays, such as surface plasmon resonance (SPR), cell viability assays, such as the colorimetric assay for assessing cell metabolic activity with 3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide (MTT), ADMET profiling, and minimum inhibitory concentration tests in the case of antimicrobial discovery. Incorporating these experimental results into the generative loop can significantly enhance the biological and pharmacological relevance of designed molecules.

3. Evolving de novo design and molecular property prediction models

The recent evolution of AI has brought a significant diversity of DL architectures and generative dynamics focused on obtaining molecules with pharmacological potential. This advancement has given rise to the field of generative chemistry, which is closely associated with the extensive progress in generative modeling. In this sense, a plethora of deep generative models have been proposed for modeling molecules, each differing in their model architecture, input features, and learning paradigm. To facilitate conceptual understanding and comparison of different approaches, it is helpful to categorize methodologies into learning paradigms, architectures, and input data types. The high-level learning paradigms discussed in this review are generative models, self-supervised learning, TL, and reinforcement learning (RL). Nevertheless, it is important to note that this division is not rigid since many approaches combine different learning paradigms or DL architectures to enhance the efficiency of their generative frameworks.

Generative models in molecular design typically focus on learning the underlying distribution of molecular structures. These methods leverage existing datasets for learning and are characterized by a clear separation between the training phase, where the model acquires knowledge, and the sampling phase, where it generates new instances. After learning the distributions, the objective is to produce new, unseen molecular structures by sampling them from the obtained distribution. Within the domain of generative models, prominent approaches include likelihood-based strategies, adversarial learning, energy-based methods, as well as flow-based and diffusion models. The implementation of these strategies can vary depending on the selected input data and architecture, with each approach potentially leveraging different neural network building blocks such as RNNs, convolutional neural networks (CNNs), GNNs, and transformers. Table 2 provides a concise overview of the main neural network building blocks discussed in this section, summarizing their core principles while highlighting the strengths and shortcomings most relevant to their application in molecular design and property prediction. Regardless of the building block, the likelihood-based models are trained to maximize the probability of observing the training data, providing a principled framework for learning the underlying distribution of molecular structures. These models learn the probability distribution p(x) from training datasets, where x represents individual molecular structures.

Table 2.

Summary of the most widely used deep learning (DL) building blocks and description of their main strengths and shortcomings.

Building block	Description	Strengths	Shortcomings
RNNs	Sequence-based models that process inputs step-by-step, maintaining an internal state to capture temporal dependencies	•Naturally suited to sequential molecular data •Captures long-range dependencies (improved with LSTM/GRU) •Proven track record in SMILES-based generation and fine-tuning for specific targets •Bidirectional RNNs improve context capture	•Suffer from vanishing/exploding gradients in long sequences •Training can be slow •Less parallelizable than transformer models •SMILES-based generation sensitive to syntax errors
VAEs	Encoder-decoder framework mapping molecules to a continuous latent space, enabling sampling and optimization. Regularized by KL divergence for smoothness.	•Enables latent space exploration and property optimization •Generation conditioned by desired molecular features •Grammar/graph variants improve chemical validity •Versatile with SMILES or graph inputs	•Vanilla VAEs may generate invalid or sparse molecules •Quality depends on latent space structure •Balancing reconstruction vs. regularization loss can be challenging
GANs	Adversarial training between a generator (produces molecules) and discriminator (distinguishes real vs. fake).	•Can produce highly novel, realistic molecules •Supports property-biased generation •Hybrid VAE-GANs combine structured latent spaces with adversarial realism	•Training instability is common •Mode collapse (low diversity) •Higher invalid molecule rate compared to VAEs/graph methods unless heavily constrained
GNNs	Models that operate directly on molecular graphs, learning from atombond connectivity and optionally 3D geometry.	•Natural representation of molecular structure •Enforces chemical validity more easily •Captures topology, substructures, and geometric features •Flow/diffusion variants allow precise property-conditioned generation	•Computationally heavy for large graphs •Flow models require invertible layers •Sequential generation approaches may still yield invalid structures if constraints are not enforced
Transformers	Attention-based models that learn global relationships in sequences or graphs. Encoderdecoder or decoder-only setups for prediction/generation.	•Capture long-range dependencies without recurrence •Highly parallelizable, scalable to large datasets •Self-supervised pre-training enables TL from unlabeled chemical data •Flexible for multi-modal inputs	•Requires large training datasets •Computationally intensive •SMILES-based models are still prone to invalid syntax unless grammar constraints are applied
RL	Agent interacts with environment to iteratively build molecules, guided by reward functions encoding desired properties.	•Direct optimization toward specific objectives (multi-objective possible) •Integrates with any generator architecture •Supports property-driven exploration of chemical space	•Reward design is critical and non-trivial •Training instability and sample inefficiency •May overfit to reward function, reducing diversity

RNNs: recurrent neural networks; LSTM: long short-term memory; GRU: gated recurrent unit; SMILES: Simplified Molecular Input Line Entry System; VAEs: variational autoencoders; KL: Kullback-Leibler; GANs: generative adversarial networks; GNNs: graph neural networks; 3D: three-dimensional; TL: transfer learning; RL: reinforcement learning.

3.1. RNNs

RNNs are a class of neural networks specifically designed to process and interpret sequential input data. Their distinctive feature is the presence of recursive layers that allow the network to maintain an internal state (memory) as it processes inputs over time. This ability to capture dependencies across a sequence made RNNs a natural early choice for molecular design, since SMILES strings represent molecules as linear sequences of characters, closely mirroring the sequential data found in natural language. By learning the probability of each token in a SMILES string conditioned on the preceding tokens, RNNs can generate chemically valid structures as they generate coherent text. This property was particularly advantageous for predicting molecular properties or generating new compounds, as it allowed models to leverage the sequential nature of SMILES notation while retaining information from earlier parts of the string. This dynamic is described mathematically in Eqs. (1), (2), where x_t is the input at time step t, h_t is the hidden state at time step t, y_t is the output at time step t, W_hh, W_xh, and W_hy are weight matrices, b_h and b_y are bias vectors, and f_h and f_y are activation functions.

$$h_{\mathit{t}}=f_{\mathit{h}}\left(W_{\mathit{hh}}{h}_{t-1}+W_{\mathit{xh}}{x}_{\mathit{t}}+b_{\mathit{h}}\right)$$ (1)

$$y_t=f_y\left(W_{hy}{h}_t+b_y\right)$$ (2)

Fig. 3 depicts the architecture of a simple RNN, highlighting its primary applications: i) as a generator for creating new data sequences and ii) as a critical component in predicting pharmacological properties. Recurrent connections correspond to an architecture organization in which stored neuronal activation from the previous time step is used. As a result, the activations of the neurons in RNNs represent the accumulating state of the network, effectively maintaining context across sequence inputs. This feature enables RNNs to capture temporal dependencies and long-term patterns in sequential data. To fully comprehend the training process of these models, it's crucial to consider the perspective of an unrolled RNN. In this unrolled view, each time step is treated as a layer in a deep feedforward network. This perspective facilitates the application of a specific type of backpropagation, known as backpropagation through time (BPTT). BPTT involves unrolling the network over time steps and computing gradients with respect to all parameters for all time steps. This process allows the network to learn from sequences of varying lengths and capture temporal dependencies in the data.

Fig. 3.

Application of recurrent neural network (RNN) architectures: (A) property prediction and (B) sequence generation.

The primary motivation for using stateful RNNs is their theoretical capacity to capture long-term dependencies in input sequences with a sufficiently large recurrent layer. However, due to the training dynamics of RNNs, this potential is rarely realized. Gradients often diminish exponentially when multiplied over numerous time steps (vanishing gradients) or, less commonly, grow excessively (exploding gradients), both of which hinder the network's ability to learn from long-term dependencies. To address these issues, long short-term memory (LSTM) and gated recurrent unit (GRU) architectures were developed. LSTMs use a gating mechanism with input, forget, and output gates along with a cell state to control information flow, enabling better retention of relevant information over time.

This architecture was applied to approximate the desired probability distribution using SMILES notation as molecular descriptor. Bjerrum et al. [33] formalized this idea using an LSTM-based DL model trained with a ZINC dataset. The authors applied the trained model to generate novel compound libraries with properties similar to those in the dataset. Other works explored the efficiency of this methodology to capture and reproduce the essential features of chemical space from a relatively small sample [34]. The next objective in the field was to learn the distribution of the data embedded in the training sets and then generate molecules with specific properties of interest. In this context, RNN models were combined with a TL strategy to fine-tune the generator, enabling the production of molecules designed to possess specific, desired properties. Segler et al. [24] implemented this strategy by retraining an LSTM model representative of the general chemical space with a smaller dataset composed of molecules biologically active against specific targets. This two-step process allowed the model to generate molecules that not only adhered to general chemical principles but also possessed desired biological activities [24]. Several studies have successfully reproduced this methodology, synthesizing promising compounds obtained by fine-tuned models. By using SMILES strings as input, these early models could capture the sequential nature of molecular structures. The RNN architecture allowed the model to learn the probability of each character in the SMILES string conditioned on the previous characters, effectively learning to generate valid molecular structures one element at a time [35,36].

In addition to the standard RNN, several works have also emerged applying biRNNs as a generating architecture for molecular sequences. The idea is to process sequences in both forward and backward directions, allowing the model to simultaneously capture context from both ends of a SMILES string. This bidirectional processing can, in principle, enhance the ability to represent long-range dependencies and generate more complex molecular structures. However, the utility of biRNNs in drug design remains debated: because SMILES are inherently directional encodings of molecular graphs, backward processing does not always add chemically meaningful information. Nevertheless, some studies have demonstrated advantages. For example, Grisoni et al. [37] introduced bidirectional SMILES-based generative model (BIMODAL), which interprets the beginning of sequences by inserting a special token into SMILES and then generates the sequence in both directions from this point. BIMODAL outperformed standard RNN models in terms of scaffold novelty, suggesting that bidirectional architectures can improve exploration of chemical space when combined with tailored training strategies [37].

3.2. CNNs

CNNs are a class of DL architectures originally designed to process data with a grid-like topology, such as images. This architecture is organized in convolutional, pooling, and fully connected layers. The rationale is to apply learnable convolutional filters that capture local patterns in the input space. In the context of molecular modeling, CNNs have been successfully adapted to process molecular representations in both 2D and 3D formats, including molecular graphs, adjacency matrices, voxelized grids of atomic densities, and pharmacophore maps. The central mechanism of a CNN is the convolution operation, which involves sliding a kernel over the input data to compute feature maps. Mathematically, for a 2D convolution, the output feature map element y_i,j is obtained as:

$$y_{i,j}=f(\sum _{m=0}^{M-1}\sum _{n=0}^{N-1}{x}_{i+m,j+n}{k}_{m,n}+b)$$ (3)

where x is the input, k is the convolutional kernel of size M × N, b is a bias term, and f is a non-linear activation function.

From the typical CNN architecture applied to molecular data, it is possible to highlight feature extraction that can be then used to predict pharmacological or physicochemical properties and integrated in generative frameworks. In CNNs, convolutional layers are often followed by pooling layers, which reduce spatial dimensions and enable translation-invariant feature extraction and fully connected layers that integrate the learned features for the final prediction or generation task.

One of the key features of CNNs that is extremely useful in cheminformatics is local connectivity. In CNNs, each node receives information from only a few local values in the input matrix, and each output is related to only certain parts of the input vector. This structure allows CNNs to efficiently detect local substructural patterns, such as FGs, ring systems, or spatial arrangements of atoms, that are often strongly correlated with biological activity [38,39]. This is particularly advantageous in property prediction tasks, where CNNs can be divided in 1D and 3D CNNs based on the detail level of the target-ligand interaction considered. Wang et al. [40] developed a binding affinity prediction method using 1D notation to represent the entire protein module, local pocket module, and SMILES module of the ligand before combining the respective features in the classification layer.

In molecular generation, CNNs are frequently integrated into hybrid architectures. For example, graph convolutional layers have been used in generative adversarial networks (GANs) to learn atom-bond connectivity patterns during the generation of novel compounds [41]. Similarly, Torng et al. [42] applied 3D CNNs to extract features from voxelized representations of pocket and ligand complexes in parallel. Those features were then concatenated to predict the putative interactions between pockets and ligands complexes, enabling the direct learning of binding patterns from spatial atomic arrangements [42].

However, despite their success, the CNN's grid dependence, inefficiency for sparse 3D data, limited interpretability, and difficulty handling long-range/dynamic interactions make them less ideal as standalone solutions. Although graph convolutional variants address irregular molecular topologies, they may still struggle to capture long-range dependencies without deeper architectures or attention mechanisms. Moreover, 3D CNNs, while powerful for spatial modeling, are computationally demanding and require careful voxelization to avoid excessive sparsity.

3.3. VAEs

VAEs are another widely used generative methodology in molecular design. These models employ an encoder to compress raw, high-dimensional molecular representations into a lower-dimensional latent vector and a decoder to reconstruct molecules from this latent space. The key advantage of this approach is that the latent space provides a continuous and structured representation of molecules, enabling the model to not only regenerate known compounds but also to adjust their latent encodings to design novel variants with potentially improved properties. To maintain this structure, VAEs incorporate a regularization term in their loss function, typically by minimizing the KL divergence, which ensures that the latent space remains smooth and compatible with a prior distribution, usually Gaussian. This allows efficient sampling of new molecules and facilitates property optimization directly in the latent space. As illustrated in Fig. 4, this framework makes VAEs particularly attractive for drug discovery applications, where identifying and refining promising chemical solutions requires both flexibility in exploration and control over molecular properties.

Fig. 4.

Variational autoencoder (VAE) architecture composed of encoder and decoder, leveraging latent space manipulation for generation and property prediction tasks.

Gómez-Bombarelli et al. [25] initially directed this potential towards the generation of molecules in the SMILES format. The authors developed a methodology to predict physical properties from the latent representation and condition the sampling of the latent space to optimize these properties in the newly generated molecules. However, this innovative strategy initially proved challenging because the resulting latent space was too sparse, often leading to the generation of invalid molecules or undesirable substructures. To address this problem, researchers explored various strategies. One notable approach was developed by Kusner and Paige [43], who introduced GrammarVAE. This model incorporates explicit information on how to construct syntactically correct SMILES strings. By doing so, the model can focus on learning molecules' semantic properties without simultaneously learning syntactic rules. This approach significantly improved the validity and quality of generated molecules, demonstrating the potential of grammar-guided generation in molecular design [43]. Building on previous work, Dai et al. [44] further emphasized the importance of semantic meaning in constructing valid SMILES representations. They introduced attribute grammar into the VAE framework, which significantly increased the percentage of valid generated molecules. This advancement demonstrated that incorporating more sophisticated linguistic structures into molecular generation models could substantially improve their performance [44,45]. Another significant development was the implementation of the generation conditioned by multiple desired properties, allowing for more precise control over the characteristics of the generated molecules [46,47]. A practical example of the application of this methodology is CogMol, which optimizes the process of identifying molecules that potentially inhibit viral targets with a VAE conditioned by the structural information of the target protein [48]. Through these application examples, it is possible to highlight the versatility and efficiency of VAE-based approaches in drug discovery as well as the continuous evolution of the methodology to adapt to the complex challenges associated with the generation of valid, diverse molecules possessing the desired properties.

Molecule generation using VAEs has indeed evolved to incorporate graph representations, combining the generative power of VAEs with the structural fidelity of graphs. This approach captures the inherent structure of molecules, leading to improved generation validity. One of the initial approaches was implemented by Simonovsky and Komodakis [49], who designed a method to generate a fully connected probabilistic graph and a standard graph-matching algorithm to align it to the ground truth. This innovative approach allowed for a more direct representation of molecular structure in the generative process. The seminal work JT-VAE formalized the task of generating molecules by assembling valid chemical substructures (e.g., rings and FGs) at nodes rather than constructing them atom by atom. By operating at the level of subgraphs, the model ensures that each building block is chemically valid before combination, which greatly reduces the chance of generating invalid structures. The main contribution of JT-VAE was the tree-structured generation of the scaffold based on these subgraph components, followed by their integration through a graph message-passing network [27]. As a result, this strategy achieved substantially higher chemical validity compared to atom-level generative approaches. Researchers have experimented with various features in graph VAEs to enhance their performance:

• •Introducing a grammar graph to encode chemical constraints in a computationally efficient manner [50];
• •Applying non-autoregressive training methodology [51];
• •Implementing learning objectives to improve diversity or the generation of graph structures from a base scaffold [52].

Recent developments in graph VAE architectures have demonstrated significant progress in molecular generation tasks. NeVAE, implemented by Samanta et al. [53], provides spatial coordinates of atoms in generated molecules. Additionally, NeVAE incorporates a gradient-based algorithm to optimize the decoder, enabling it to generate molecules that maximize specific properties of interest. It can also optimize the spatial configuration of atoms for greater stability, outperforming several state-of-the-art methods in various benchmarks [53]. Wang et al. [54] proposed a molecular substructure tree generative (MSTG) model that constructs molecules as trees of chemically meaningful substructures, unlike atom-by-atom generation methods that often struggle with chemical validity and scalability. This VAE architecture uses a multichannel substructure-graph GRU that builds molecules through a three-stage process, topology prediction, edge generation, and node generation, closely mirroring how chemists conceptually approach synthesis, by first considering the overall scaffold, then the connections between fragments, and finally the specific atoms involved.

3.4. GANs

GANs are a class of deep generative models introduced by Goodfellow et al. [55], consisting of two competing networks, a generator that produces synthetic samples from random noise and a discriminator that distinguishes real from generated data, as demonstrated in Fig. 5. Through this adversarial training process, the generator gradually learns to produce outputs that resemble the training distribution. In the context of drug discovery, this framework has been adapted to molecular design by generating novel SMILES strings or molecular graphs that mimic the distribution of bioactive compounds. Such approaches allow GANs to capture chemical diversity and propose structures with desirable pharmacological properties, though training instability and a high rate of invalid molecules remain persistent challenges.

Fig. 5.

Generative adversarial networks (GANs) standard framework to approximate a desired distribution and generate new molecules.

The objective-reinforced generative adversarial network (ORGAN) for inverse-chemistry (ORGANIC) framework, building upon earlier GAN approaches, applied an inverse-design strategy to bias specific physicochemical properties in molecules encoded with SMILES notation. This model demonstrated that GANs could be extended to the discrete domain of molecular structures, enabling controlled molecular generation. However, despite its potential, the method faced major challenges, including high rates of invalid and duplicate molecules as well as hard-to-synthesize structures, reflecting the intrinsic difficulties of training GANs effectively [56]. Macedo et al. [57] developed MedGan, a scaffold-focused generative framework that combines Wasserstein GANs with graph convolutional networks (GCNs) to generate novel quinoline-based molecular graphs. The model is specifically optimized through extensive hyperparameter tuning and trained on curated subsets of the ZINC15 and PubChem datasets. By representing molecules as graphs and learning both atom and bond-level features via GCNs, MedGAN generates chemically valid, diverse, and unique molecules while maintaining key pharmacological attributes such as chirality, atom charge, and synthetic accessibility.

Due to the inherent instability of GAN training, some works have explored hybrid models that fuse VAEs and GANs to combine the advantages of both architectures. In these frameworks, VAEs first transform discrete molecular inputs into continuous latent representations, ensuring a structured and smooth space for sampling, while GANs then refine these representations to generate more coherent and diverse molecules. This combination leverages the validity and regularization strengths of VAEs with the realism and diversity-enhancing capacity of GANs, helping to reduce mode collapse and improve chemical plausibility. Generative frameworks such as druGAN and LatentGAN have demonstrated superior performance compared to VAEs or GANs used independently, successfully optimizing desired properties while minimizing repeated and invalid molecules [26,28]. However, these hybrids also come with greater training complexity, requiring careful balance of VAE reconstruction and KL regularization losses with the adversarial training dynamics of the GAN. In practice, this makes VAE-GANs less plug-and-play and more resource-intensive to train, which can limit their accessibility despite their improved performance. The evolution of these architectures is moving toward GAN-based models with more stable training procedures, higher rates of valid molecule generation, and more precise control over molecular properties.

VAE-GAN strategies have been extended to utilize graph representations as input, broadening the scope of molecular generation techniques. Maziarka et al. [58] pioneered this approach by combining the latent representations of graphs from the JT-VAE model to implement a cycleGAN [59]. This innovative combination yielded a higher rate of valid molecules compared to previous methods. However, the model showed limited success in optimizing physicochemical properties, highlighting the ongoing challenges in balancing structural validity with desired molecular characteristics. In parallel, De Cao and Kipf [60] introduced an alternative generative approach, developing a likelihood-free model for small molecular graphs. By directly generating adjacency tensors and node features in a single step, their method avoided the computationally expensive graph matching and node ordering heuristics required in earlier graph-based VAEs, significantly reducing training and sampling overhead. This efficiency is particularly valuable in early-stage virtual screening, where speed and scalability are critical for evaluating large chemical libraries while maintaining validity and control over desired molecular properties [60].

Graph representations of molecules provide crucial information about atomic connectivity. Some advanced generative approaches leverage this information to predict not only the molecular structure but also the spatial arrangements of atoms in 3D space. Mansimov et al. [61] combined a conditional VAE architecture with graphs to generate energetically favorable molecular conformations by learning the energy function. Their model processes molecular graphs using a message-passing neural network and represents molecular conformations with 3D coordinate vectors. This approach aims to maximize the likelihood of the reference conformations for the most promising molecules. The results demonstrated that the generated conformations closely match the experimental references and that the model is more computationally efficient than traditional force field methods. Another promising generative framework integrating GANs was proposed by Bai et al. [62]. MolAICal was designed to generate 3D drug-like ligands directly within protein pockets. It integrates a Wasserstein GAN-based sequence and graph generative model, trained on FDA-approved fragments and ZINC compounds, with a fragment-based growth algorithm and Vinardo scoring for affinity optimization. MolAICal approaches 3D structure generation by anchoring learned molecular fragments in the protein binding site and iteratively expanding them using perturbation search and genetic operations, thus enabling the design of diverse and synthetically accessible molecules with high predicted binding affinities.

3.5. GNNs

Graphs provide a natural and intuitive way to represent molecular structures, accurately capturing their topology. Compared to linear notations such as SMILES, graph-based models generally achieve higher rates of chemical validity, since they more easily enforce bonding rules and constraints. However, invalid structures can still occur in practice, particularly when incorporating complex FGs or extending models to 3D geometries, meaning that additional constraints are often required for reliable molecular generation. Recognizing these benefits, researchers have developed architectures focused on graphs, such as GCNs, graph attention networks (GATs), diffusion-based models, or flow-based models.

Li et al. [63] applied a sequential graph generation strategy to iteratively refine its intermediate structure, allowing for more precise control over the generation process. Their most successful strategy treated the generation task as a Markov process, incorporating a molecule-level recurrent unit to consider the previous experience in the generation sequence. Additionally, the architecture included a GCN to extract information from intermediate graph states. The results demonstrated versatility across different drug discovery tasks. The model showed proficiency in generating molecules with desired properties, a crucial capability for targeted drug design. Furthermore, it demonstrated the ability to generate molecules from a given scaffold, which is particularly useful in lead optimization and SAR studies.

Due to the ease of manipulating molecular substructures, graphs are also suitable for fragment-based generation. Imrie et al. [64] presented a model that takes two independent fragments to design a single molecule that incorporates both. By incorporating 3D structural information about the fragments, such as inter-fragment distance and orientation, the model was successfully applied to scaffold hopping and fragment linking.

Considering the versatility of graphs, flow models are also an architecture used to generate molecules with optimized properties, as described in Fig. 6. Flow-based models are generative frameworks that focus on learning invertible transformations between simple distributions and complex data distributions. One of the initial works was implemented by Madhawa et al. [65], in which two latent representations (adjacency tensor and node label assignments) were used to characterize a molecular graph. The generation dynamics involves generating the graph structure and then the node's attributes according to the structure, which yields the exact likelihood maximization on the graph with two reversible flows. This dynamic allows researchers to explore the latent space to identify molecules with optimized properties [65]. However, the model generates graphs in a single-shot dynamic, which increases the probability of generating invalid molecules. To address the limitations of single-shot generation, other approaches have emerged that propose an iterative and sequential generation of graph representations to leverage chemical domain knowledge, such as valency checking in each step.

Fig. 6.

Description of the steps required to implement a flow-based molecular generator model.

Beyond molecular graphs, knowledge graphs (KGs) have become increasingly important for integrating heterogeneous biomedical data, including drugs, targets, diseases, side effects, and pathways. Unlike molecular graphs, which capture intramolecular topology, KGs encode relational knowledge between diverse biological entities. KGs have been applied to drug repositioning, polypharmacy prediction, and target discovery by leveraging graph embedding and message-passing techniques. Models such as RotatE and TransE-based embeddings have demonstrated that representing drugs and proteins in a relational graph framework enables the prediction of novel drug-drug interactions, drug-disease associations, and candidate therapeutic targets [66,67]. Integrating molecular-level GNNs with large-scale biomedical KGs represents a promising multi-scale approach for end-to-end drug discovery pipelines.

Drug discovery increasingly relies on systems-level representations that go beyond pairwise molecular interactions. Complex biological networks, such as PPI networks, gene regulatory networks, and metabolic pathways, provide insights into emergent properties of diseases and drug responses. These frameworks allow the identification of hub nodes as potential therapeutic targets, the study of disease modules, and the prediction of drug effects at the systems biology scale. Extending this concept, hypergraph theory enables the modeling of higher-order relationships that cannot be captured by traditional pairwise graphs, such as drug-gene-disease triplets or multi-protein complexes. Hypergraph neural networks have been applied to predict polypharmacy effects, model drug-target-disease associations, and capture synergistic or antagonistic drug actions [68,69]. Together, complex networks and hypergraphs provide complementary tools that integrate molecular-level design with systems pharmacology, enabling more holistic and accurate modeling of drug discovery processes.

Flow-based models require exact likelihood estimation through specialized invertible layers, which becomes computationally challenging in high-dimensional spaces. Diffusion models offer an alternative by gradually adding and then removing noise in a forward-reverse process. This framework enables approximate likelihood learning and facilitates property-guided molecular generation. For instance, Hoogeboom et al. [70] proposed an equivariant diffusion model that learns to denoise atomic coordinates and types under symmetry constraints, generating 3D molecular structures invariant to translation and rotation. It generates 3D molecular structures by iteratively refining molecular conformations, integrating continuous and categorical features for efficient and accurate generation. These methods exemplify how graph-based diffusion approaches are now bridging geometric DL with pharmacophore-guided design. However, such 3D diffusion models remain largely experimental, and only a few have progressed into robust drug discovery pipelines, highlighting the need for further validation and scalability before widespread adoption.

3.6. Transformer architecture

Attention mechanisms have become increasingly important in the field of drug development, as demonstrated by their ubiquity in all recent publications. The emergence of the transformer architecture was pivotal in property prediction and de novo drug design applications. At the core of these models lies the concept of attention, which allows models to relate different sequence positions and to generate context-aware representations of sequence data by weighing the importance of different tokens within a sequence [71]. This is particularly useful in SMILES representations, where chemically related atoms may appear far apart in the sequence, for example, ring closures or distant FGs. By dynamically attending to these non-adjacent tokens, attention mechanisms enable the model to capture long-range dependencies that are critical for accurately reconstructing molecular structures and predicting their properties. In the scaled dot-product self-attention, given a set of input token embeddings X ∈ R^Nxd, attention is computed as:

$$\text{Attention}\left(Q,K,V\right)=\text{softmax}\left(\left(Q{K}^T\right)/\left(\sqrt{d_k}\right)\right)\mathit{V}$$ (4)

where Q = XW^Q, K = XW^K, and V = XW^V are the query, key, and value matrices, respectively, and d_k is the dimension of the key vectors. Intuitively, each token in the input sequence (e.g., an atom or bond symbol in a SMILES string) generates a query asking “which other parts of the molecule should I pay attention to”, while the keys represent the features of all other tokens, and the values carry the information to be aggregated. This operation enables the model to dynamically focus on chemically relevant substructures, regardless of their position in the sequence.

The core of the transformer architecture consists of an encoder stack and a decoder stack, as illustrated in Fig. 7. The encoder processes the input data, creating a rich, contextualized representation of the input sequence, while the decoder generates the output sequence based on the encoder's representations and its self-attention mechanisms. At a high level, the transformer encoder uses stacked layers of multi-head self-attention and feed-forward blocks to learn rich representations of molecular sequences, which are then used for classification or regression tasks in property prediction. Meanwhile, the transformer decoder, equipped with causal attention and cross-attention to the encoder, enables controlled molecule generation token-by-token, making it ideal for tasks like scaffold-based or target-guided molecule design. Therefore, the transformer architecture serves distinct yet complementary functions through its encoder and decoder components. The encoder leverages attention mechanisms to transform each token in the input sequence into a robust, contextualized representation, while the decoder specializes in sequence generation based on both the encoded input and the previously generated context. This architectural division enables multiple applications: the encoder alone often serves as the basis for property prediction models (Fig. 7A), the decoder alone can act as a generative policy for de novo molecule design (Fig. 7B), and when used together in a sequence-to-sequence framework, the encoder-decoder pairing supports tasks such as scaffold-based molecule generation or protein-to-ligand translation, where one type of sequence must be mapped onto another.

Fig. 7.

Transformer architecture used in computational drug development: (A) property prediction and (B) sequence generation. CLS: classification token; CI: chlorine atom; MASK: mask token; SEP: separator token.

One of the first applications of the transformer architecture in molecular design tasks was implemented by Grechishnikova [72], who formulated the generation of molecules as a translation problem. The task is to use the amino acid sequence of the protein to interact with as input and obtain chemical compounds with the predicted ability to bind to the protein target [72]. The model generated promising molecules, with the advantage of not needing other known inhibitors to generate candidate molecules. Following a similar reasoning, Qian et al. [73] presented a framework that uses protein information to guide the generation of promising molecules, but in this case, composed of a modified version of the transformer model and a Monte Carlo tree search algorithm for the conditional molecular generation. The change in the model aims to make the joint embedding responsible for combining information about the protein and the candidate molecule more efficient and instructive [73].

However, protein sequences are often lengthy, and not all regions are directly involved in interactions. On that account, Yoshimori and Bajorath [74] designed a similar strategy for translating amino acid sequences into molecular representations but restricting amino acids that constitute ligand-binding regions. This approach helps reduce redundancy and noise in sequence information by excluding amino acids not critical to the binding process [74]. Another branch of research is generative pre-trained transformer (GPT), initially developed for applications in natural language processing applications, which several molecular generation frameworks have adapted. Bagal et al. [75] presented a transformer-decoder trained to predict the next token in SMILES sequences conditioned by desired molecular properties, obtaining high validity rates and unique molecules.

The potential of attention mechanisms has also been explored in combination with VAEs to harness the capabilities of powerful sequence modeling and latent spatial exploration, respectively. Kim et al. [76] demonstrated the ability of attention mechanisms to capture complex substructural representations of molecular features by implementing a transformer embedded in a conditional VAE framework. This approach creates molecules with desired properties by exploring the latent space formed by the transformer model's powerful understanding of chemical language [76]. The transformer architecture has also been integrated into hybrid strategies, where genetic algorithms are applied to select molecules with the most significant pharmacological potential. Monteiro et al. [77] implemented a fine-tuned transformer-decoder to generate molecules with multiple optimized properties by applying a non-dominated sorting genetic algorithm that selects molecular hits according to the dominance across molecular properties. Alternatively, Blanchard et al. [78] propose using an masked language model (MLM) transformer with an adaptive optimization strategy. In this approach, the model is continually trained on selected populations to adapt to new regions of chemical space that are valuable for specific optimization tasks [78]. Their results demonstrate that this adaptive approach leads to improved fitness optimization compared to a fixed pre-trained model, both for heuristic metrics (drug-likeness and synthesizability) and predicted protein binding affinity, thus enhancing the capabilities of genetic algorithm optimization for molecular design. The potential of this architecture was also demonstrated in the work developed by Rafiei et al. [79], DeepTraSynergy. This multitask DL framework predicts synergistic drug combinations by integrating multimodal biological data. It uses transformers for drug feature extraction from SMILES strings and incorporates DTIs, PPIs, and cell-protein associations to improve predictive performance. The model jointly learns three tasks: synergy prediction, DTI prediction, and toxicity estimation.

These works provide a landscape of the field's progression and highlight how transformer-based architectures can be applied to generate new compounds with desired characteristics, while also noting that their success often depends on the availability of large, high-quality pretraining datasets and comes at the cost of substantial computational resources.

3.7. RL

RL is a subfield of ML that allows an agent to learn specific behaviors in an interactive environment through trial and error based on its own actions and acquired experience. The primary objective of RL is to maximize cumulative reward, making it particularly well-suited for sequential decision-making tasks. In RL settings, each action not only affects the immediate outcome but also influences future situations and potential rewards. This characteristic makes RL especially valuable for addressing complex, multi-step processes such as constructing molecules atom-by-atom. As described in Fig. 8, RL settings include an agent that interacts with the environment and makes the decisions. The environment represents the external system providing the context for the agent's actions and determining the consequences that arise from those actions. The agent's current situation within the environment is captured through a state (s), a representation that helps the agent understand its position and circumstances. Based on this state information, the agent selects an action (a) that modifies the environment and triggers state transitions. As a result of these actions, the environment provides feedback as a reward (r), a numerical signal that guides the agent by indicating the desirability or utility of specific state-action pairs. The choice of actions is based on the agent's policy, which determines the strategy for choosing the next action, given the current state. The agent's ultimate goal is to optimize this policy through learning, thereby improving its decision-making process and achieving better outcomes over time.

Fig. 8.

Reinforcement learning (RL) paradigm for directed molecular generation.

RL has emerged as a powerful ML paradigm in de novo drug design, serving as an auxiliary tool for generating molecules with optimized properties. The versatility of RL enables seamless integration of property prediction models (environment) with novel molecular instance generators (agent), irrespective of the underlying architectures. Typically, molecular generation using RL as the learning paradigm is defined as a Markov decision process in which:

• •The possible actions are the atoms or bonds that can be integrated into the structure;
• •The set of states corresponds to the obtained molecular structures (intermediate or final).

The seminal work by Popova et al. [80] paved the way for using RL in drug design. The proposed policy-based method directly learns a policy (probability distribution over actions in a given state) for decision-making without necessarily estimating a value function. The model is based on LSTMs for both generating and predicting molecular properties. The idea is for the generator to navigate through the chemical space and the predictor to assign rewards to the sampled compounds depending on the optimization level of the desired properties. After learning through trial and error, the generator can produce optimized chemical libraries of novel compound properties [80].

The field has further evolved with the integration of policy-gradient approaches and GANs, as exemplified by ORGAN and ORGANIC. These works utilize an RL framework to approach drug design as a sequential decision-making process [56,81]. The transformer architecture has also been combined with RL policy-gradient methods to leverage RL's ability to direct the potential of attention mechanisms toward desired regions of chemical space. Liu et al. [82] implemented a transformer model for molecular structure generation, containing an encoder to receive scaffolds as input and a decoder to generate molecules as output. In their approach, molecules were represented as graphs, and the authors proposed a novel positional encoding for atoms and bonds based on an adjacency matrix. This encoding incorporated growing and connecting procedures for molecule generation, starting from a given scaffold based on fragments. RL optimization was guided by two scoring functions acting as the external environment. This method achieved 100% validity for generated molecules and demonstrated a high predicted affinity value towards the desired target [82].

Transformer-based networks have been further optimized to maximize the expected reward of objective functions by encoding various properties such as affinity towards crucial proteins of the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) virus [83] and the optimization of multiple molecular constraints using conditional transformers [84]. Zhavoronkov et al. [85] implemented a promising approach, combining RL, variational inference, and self-organizing maps (SOMs) as reward functions to guide the search for molecules with affinity towards discoidin domain receptor 1 kinase (DDR1). In this framework, SOMs served as feature-mapping tools to organize molecules by structural similarity, encouraging exploration of chemically diverse regions of chemical space rather than over-optimizing a narrow subset of scaffolds. In molecular RL, high gradient variance is a critical challenge because reward signals are typically sparse and noisy, most randomly generated molecules fail to meet activity or validity criteria, leading to unstable training and poor convergence. To address this, the authors applied a “baseline” correction by calculating the average reward for all molecules in a batch and subtracting it from individual rewards, thereby reducing gradient variance and stabilizing policy updates. Another method to mitigate gradient variation is proximal policy optimization (PPO), which is a strategy to limit overly bold policy updates that risk moving away from advantageous optimization regions. You et al. [41] demonstrated this concept by presenting a GCN for directed graph generation through RL. Their approach incorporated domain-specific rules in the external environment and utilized PPO with a clipped objective function to prevent excessively large policy updates [41]. Despite the success of these approaches, policy-gradient methods can still be unstable in terms of gradients, making it challenging to learn the correct policy.

Actor-critic methods offer an intermediate alternative between policy-based and value-based approaches, combining characteristics of both learning types. This method comprises an actor who learns and improves the policy and a critic who evaluates the policy by estimating value functions. Ståhl et al. [86] applied an actor-critic algorithm to generate novel molecules with multiple optimized properties by transforming existing lead molecules. The workflow is based on fragmenting candidate drugs and selectively replacing some molecular fragments to improve their properties [86]. Unlike full-molecule generation, which explores chemical space from scratch and can propose entirely novel scaffolds, fragment replacement starts from known lead compounds and introduces modifications within a constrained chemical framework. This narrower scope allows for more targeted multi-objective optimization, such as improving potency, solubility, or ADMET properties, while maintaining a higher likelihood of preserving drug-like features. However, this added precision comes at the cost of reduced structural novelty, since the search space is limited by the chosen precursor scaffolds.

Lastly, value-based methods estimate the expected cumulative reward from state-action pairs to derive an optimal policy. One of the most used variants of value-based methods is Q-learning. A noteworthy example was implemented by Zhou et al. [87], who presented a multi-objective optimization framework designed for molecular optimization without the need for pre-training. It formulates molecule editing as a Markov decision process and uses deep Q-networks to guide stepwise atom/bond modifications while ensuring 100% chemical validity. Unlike earlier approaches, molecular deep Q-network (MolDQN) operates directly on molecule graphs and allows for multi-objective optimization, such as maximizing QED while maintaining similarity to a lead compound [87]. Addressing the traditional limitations of property-optimization metrics, Fang et al. [88] propose QADD, a novel framework for de novo drug design that integrates multiobjective deep RL with a GNN-based molecular quality assessment module. Molecules are represented as graphs, and the evaluation model is trained using a large benchmark dataset of drug-like and non-drug-like molecules. During RL, the generator is rewarded based on multiple objectives: QED, SAS, and the predicted quality assessment module score. This approach enhances validity, success rate, and target-specific binding affinity, demonstrated through experiments on dopamine receptor D2.

Architectures have also evolved to consider other challenges such as preventing redundant exploration of specific chemical regions or consideration of the exploration/exploitation dilemma [89]. Regarding the multi-objective nature of drug design, strategies have emerged to prioritize molecules that focus on this aspect within the new chemical libraries. Examples include non-dominated sorting or pair-based comparisons [77,90]. The primary objective is to find solutions with Pareto optimality, ensuring an optimal compromise between various molecular properties to be optimized. Considering the need to standardize the knowledge developed in the field of deep RL research, Bou et al. [91] developed automated chemical exploration via generative enhancement (ACEGEN), a robust and modular framework for RL-based molecular generation. Built on top of the TorchRL library, ACEGEN supports multi-policy gradient algorithms and incorporates core techniques such as KL divergence regularization, experience replay, and reward modeling to improve stability and diversity. This tool facilitates the application of multi-objective optimization through the combination of several scoring functions within a unified reward structure and improves chemical diversity through exploration-aware metrics [91].

These advancements collectively underscore the pivotal role of RL in revolutionizing de novo drug design, offering unprecedented capabilities in generating optimized molecular structures with desired properties.

3.8. Molecular property prediction

The prediction of molecular properties is a central aspect of all CADD methodologies. These models aim to predict biological, physical, or chemical properties of compounds, taking into account their structures or a learned representation. Molecular property prediction constitutes the initial core of HTS methods. However, given the continual improvements in performance and their potential to optimize resource allocation in drug discovery, their application has expanded for areas such as bioactivity, toxicity, and solubility prediction.

Initially, bioactivity prediction relied heavily on classical QSAR models and quantum mechanics-based simulations, which were often limited by handcrafted descriptors, small datasets, and high computational cost. In recent years, however, these approaches have gradually been complemented and, in many cases, surpassed by data-driven solutions powered by ML and DL. Advances in computational power and the availability of large, heterogeneous datasets have enabled these models to rapidly learn complex SARs that were difficult to capture with traditional methods. As ML and DL models continue to reshape molecular property prediction, it is essential to understand their current capabilities and limitations, as well as likely future directions such as the integration of multi-modal data, the use of large pre-trained foundation models, and tighter coupling with experimental feedback to improve real-world applicability.

QSAR modeling forms the theoretical foundation of prediction models that establish the empirical relationship between a molecular structure and its corresponding biological activity [92]. The first data-driven approaches began to emerge with the advent of ML in cheminformatics. Methods such as Support Vector Machine (SVM), Random Forest (RF), and Gradient Boosting Machines (GBM) demonstrated the ability to predict a variety of molecular properties with promising results. The key advantage of early models like SVM or RF was their ability to introduce non-linearity into predictions. With the growth of available chemical datasets, other methodologies, such as GBM, emerged due to their performance in handling imbalanced datasets. Nevertheless, these methods remained constrained by the quality and informativeness of their input features, underscoring the need for more automated and generalizable approaches.

The transition to data-driven CADD methods was catalyzed by the availability of vast open-access databases, which provided unprecedented access to chemical structures, bioactivity, and assay data. Specialized datasets such as BindingDB, Davis, and LIT-PCBA offer curated information on drug-target binding affinities, further supporting the development of predictive models for molecular interaction and pharmacological profiling. These diverse repositories laid the groundwork for automated feature extraction and learning, with DL emerging as a natural solution due to its capacity to learn rich molecular and interaction representations directly from raw data. This advancement enabled scalable, accurate modeling of complex biochemical relationships, fueling progress in both virtual screening and structure-activity prediction.

A pivotal advancement in molecular representation came with graph-based approaches, which comprehensively capture molecular topology and connectivity patterns. GNNs and variants like message-passing neural networks became popular due to their ability to learn chemical language and generate informative descriptors [93]. These architectures process molecules as graphs, where atoms represent nodes and chemical bonds serve as edges, enabling a more natural representation of chemical structures. Duvenaud et al. [94] implemented CNNs to process molecular graphs, forming more interpretable molecular descriptors for DL architectures and with better performance in predictive tasks. To take advantage of the geometric structure information present in molecules, Wang et al. [95] combined CNNs and graph structures to obtain a convolution spatial graph embedding layer to retain the spatial connection information on molecules. The results demonstrated state-of-the-art performance in several benchmark experiments. The success of these graph-based methodologies stems from their ability to simultaneously process topological, chemical, and geometric information, providing a more thorough representation of molecular structures than traditional descriptors.

SMILES notation enabled the application of sequence-based models based on RNNs. This representation allows DL models to learn statistical patterns in molecular sequences, including dependencies between distant tokens. However, these long-range relationships are not always chemically intuitive, since SMILES are linearized and non-canonical representations of inherently graph-based structures. Chakravarti and Alla [96] applied LSTM and bi-directional LSTM layers to SMILES strings and demonstrated their ability to interpret sequences, form informative descriptors, and predict biological properties in three large and diverse datasets. In these studies, LSTM models consistently outperformed traditional fragment-based methods in predicting the activity of compounds with low similarity to the training data. This generalization is a crucial quality for real-world QSAR applications [96].

Generative models such as VAEs have emerged as powerful tools for property prediction through their ability to generate informative molecular descriptors. These models encode molecular structures into low-dimensional latent spaces, facilitating more efficient property prediction and subsequent optimization. Tevosyan et al. [97] trained a predictor using embeddings from a VAE as molecular descriptors, where the VAE was trained on chemical formula information and other complementary molecular descriptors. This approach leverages the VAE's ability to capture essential molecular characteristics in a compressed, continuous representation.

A significant limitation of traditional molecular property prediction approaches lies in their dependence on property-annotated molecular datasets, which are often limited in size and scope. This constraint challenges model robustness and generalization capabilities. One potential solution to this problem is the self-supervised learning characteristic of transformer architectures. These approaches take advantage of the large chemical space of unlabeled datasets to learn the basic rules for constructing molecules either in the form of SMILES or molecular graphs. Although this initial learning is task-agnostic, the model can later be easily adjusted to predict chemical properties. The bidirectional encoder representations from transformers (BERT) approach applies a strategy for predicting masked tokens in sequences to guarantee the learning of molecular features automatically [98]. In this sense, self-attention mechanisms are used to learn context-dependent relationships in vast sets of molecular data. At the end of the pre-training process, these models acquire a holistic perspective of the molecule and form molecular representations useful for downstream tasks of property prediction. This adaptation to property prediction can be accomplished with TL to further expand the versatility of these models in addressing diverse datasets [99]. ChemBERTa and SMILES-BERT exemplify this evolutionary advancement in molecular modeling. These models undergo pre-training on millions of compounds to learn intricate chemical rules and token relationships embedded within SMILES notation. Their subsequent fine-tuning for specific property prediction tasks demonstrates the versatility and efficiency of this approach [48,100]. In this sense, Jiang et al. [101] developed a property prediction framework by improving conventional GNN-based approaches through the integration of atom-level, fragment-level, and junction-level views into a unified multi-view heterogeneous graph representation. The proposed transformer-based architecture allows the model to better capture substructure-level chemical knowledge, demonstrating promising performance especially in tasks related to toxicity and solubility prediction [101].

Another research path for molecular property prediction is the use of 3D descriptors due to their ability to consider features that determine complex DTIs where spatial arrangements play a crucial role. Liu et al. [102] introduced a self-supervised learning method that introduces 3D geometric information of molecules to enhance the learning process of 2D molecular graph representation. More recently, Fang et al. [103] proposed an innovative representation learning method, Geometry Enhanced Molecular Representation (GEM), which incorporates molecular geometry into GNN using a specially designed architecture (GeoGNN) and geometry-level self-supervised learning tasks. By leveraging both molecular topology and geometry through modeling the effects of atom, bond, and bond angle, GEM achieved state-of-the-art performance on several molecular property prediction benchmarks, showcasing the advantage of integrating 3D spatial information into the learning process, particularly for properties that depend on specific molecular conformations and spatial arrangements [103].

Another powerful alternative to face the scarcity of labeled data is the contrastive learning strategy. This paradigm uses unlabeled data for self-supervised learning without human annotations. The basic principle is to learn representations automatically by contrasting positive and negative sample pairs. In practice, the objective is to minimize the difference between positive pairs and maximize the difference between negative pairs. Wang et al. [104] demonstrated the effectiveness of this methodology by combining a GNN with self-supervised training on extensive molecular datasets. The idea was to apply three data augmentation methods for molecular graphs, including subgraph removal, atom masking, and bond deletion. Afterward, the goal is to combine the molecular instances in pairs to apply contrastive learning and obtain informative molecular descriptors. Moon et al. [105] extended this line of research by introducing 3DGCL, a lightweight 3D-3D graph contrastive learning framework for molecular property prediction that leverages structural semantics while addressing key limitations of previous approaches. Unlike prior models that relied on heavy datasets, altered molecular graphs, or ignored 3D geometry, 3DGCL constructs positive pairs using conformers thus preserving chemical identity during self-supervised pre-training. The results show that using a reduced number of training samples, the model achieves state-of-the-art or competitive results across six benchmark datasets.

4. Translating AI-driven molecular design into clinical applications

AI is at the cornerstone of the current digital revolution. The integration of AI into drug research and development (R&D) endeavors is on the cusp of a transformative upgrade that can potentially revolutionize the pharmaceutical landscape and unlock unprecedented innovation and efficiency. Nevertheless, while the promise is profound, it is essential to temper expectations. Successfully applying AI requires careful integration as it is not a universal remedy that can instantly convert any ideas into a tangible outcome. Rather, it should be perceived as a valuable tool that, when used judiciously, can effectively help address the root causes of drug failure and streamline the development process [106].

In this chapter, we will explore the signs of success of clinical-stage AI-enabled small molecules and how the regulatory framework is adapting to this new reality.

4.1. Early proof-of-concept of clinical stage AI-enabled small molecules

Integrating AI into drug discovery represents a groundbreaking milestone, offering unparalleled opportunities with potential to revolutionize the field. In the last few years, AI-enabled drug discovery has been a flourishing field with unmatched momentum. A pipeline growth analysis of the years between 2010 and 2021 of 24 “AI-native” drug discovery companies, in which AI is at the core of their discovery strategy, revealed a rapid pipeline growth for a subset of 20 of those companies. Their average annual growth rate was around 36% of asset count, mainly driven by programs at the discovery and preclinical stage, which reflects the early-stage nature of “AI-native” companies [107].

“AI-native” companies appear to have a combined pipeline equivalent to 50% of leading large pharmaceutical companies in-house discovery and preclinical programs, which showcases an impressive landscape [107]. A recent analysis of the clinical pipelines of “AI-native” companies revealed that AI-discovered molecules in phase 1 clinical trials have between 80% and 90% success rate, which is substantially higher than the average reports for the traditional drug discovery pathway. In phase 2, the success rate is around 40%, which is in line with the traditional industry benchmarks. Altogether, these results strongly endorse the clinical potential of AI-discovered molecules, despite the limited sample size of the analyses [108].

The application of AI holds great promise for streamlining every step of the highly complex drug discovery value chain, from identification/design of AI-derived molecules, through target identification, prediction of compound interactions, optimization of clinical trial design, patient recruitment, and trial execution [[109], [110], [111], [112]].

AI-developed drugs can bring several unique clinical advantages compared with traditional small molecules discovered through conventional pipelines. These advantages stem from how AI models integrate vast amounts of data (omics, chemical libraries, clinical data, and real-world evidence) to optimize design not only for activity but also for developability, safety, and patient usability. Some of the most relevant advantages include optimized dosage and pharmacokinetics [113], improved administration routes [114], improved formulation and drug delivery [115], and early toxicity prediction [116].

There have been reports of several modes of discovery for AI-derived molecules, namely: AI-generated small molecules [117]; molecules with AI-discovered drug targets [118]; AI-repurposed molecules [119]; AI-discovered biologics [120]; and AI-discovered vaccines [121].

Several “AI-native” drug discovery companies have been able to progress molecules into clinical trials, raising the expectations of investors and pharmaceutical partners for R&D programs. An analysis of the field between 2015 and 2023 revealed that “AI-native” companies and their large pharmaceutical companies partners have been able to progress 75 molecules into clinical testing, of which 67 were in ongoing clinical trials as of December 2023. Most of these molecules are currently in phase 1, although some have been able to progress to phase 2 and one is in the very early stages of phase 3 (trials started in late 2024). Such molecules have been identified through a wide range of modes-of-discovery [108]. Paying closer attention to the AI-generated small molecules, Table 3 systematizes noteworthy examples that have advanced into clinical trials, informing the reader on how successfully these AI-derived assets are performing and the early signs of their potential.

Table 3.

Artificial intelligence (AI)-discovered small molecules in clinical trials.

Drug name	Company	Platform name	Phase 1	Phase 2	Therapeutic area	Target
A2A-252	A2A Pharmaceuticals	SCULPT™	NCT06136884: recruiting	–	Leukemias and solid tumors	TACC3 PPI inhibitor
BMF-219	A2A Pharmaceuticals/Biomea Fusion	SCULPT™	NCT05153330: terminateda	–	Leukemias and solid tumors	Menin inhibitor
			NCT05631574: terminateda
BMF-219	A2A Pharmaceuticals/Biomea Fusion	SCULPT™	NCT05731544: completed (phase 1/2)	NCT06152042: terminatedb	Restore functional β cell mass to halt or reverse diabetes progression	Menin inhibitor
AC-699	Accutar Biotechnology	Accutar Virtual Screen	NCT05654532: recruiting	–	Breast cancer	ER degrader
AC-682	Accutar Biotechnology	Accutar Virtual Screen	NCT05080842: terminatedc	–	Breast cancer	ER degrader
			NCT05489679: terminatedc
AC-0176	Accutar Biotechnology	Accutar Virtual Screen	NCT05673109: terminatedd	–	Prostate cancer	AR degrader
			NCT05241613: terminatedd
AC-0676	Accutar Biotechnology	Accutar Virtual Screen	NCT05780034: recruiting	–	Hematological cancers	BTK degrader (WT and C481S)
C4X3256 (INDV-2000)	C4X Discovery/Indivior	Conformetrix	NCT04413552: completed	NCT06384157: active, not recruiting	Opioid use disorder	OX1R antagonist
			NCT04976855: completed
			NCT05694533: completed
EVOEXS-21546	Exscientia/Evotec	Precision Design	NCT04727138 (1a): completed	–	Solid tumors carrying high adenosine signatures	A2A receptor antagonist
		Centaur Chemist™	NCT05920408 (1b/2): terminatede
DSP-1181	Exscientia/Sumitomo Dainippon	Precision Design	NCT not found	–	Obsessive-compulsive disorder	5-HT1A receptor agonist
		Centaur Chemist™
DSP-0038	Exscientia/Sumitomo Dainippon	Precision Design	NCT not found	–	Alzheimer's disease psychosis	Dual-targeted 5-HT1A receptor agonist and 5-HT2A receptor antagonist
		Centaur Chemist™
EXS-617	Exscientia/GT Apeiron	Precision Design	NCT05985655 (1/2): recruiting	–	Advanced solid tumors	CDK7 inhibitor
		Centaur Chemist™
EXS-4318	Exscientia/BMS	Precision Design	NCT05760937: completed	–	Inflammatory and autoimmune conditions	PKC-θ inhibitor
		Centaur Chemist™
INS018-055	InSilico Medicine	Chemistry42®	NCT05154240: completed	NCT05938920: completed	Idiopathic pulmonary fibrosis	TNIK inhibitor
				NCT05975983: recruiting
ISM-0003091	InSilico Medicine/Exelixis	Chemistry42®	NCT05932862: recruiting	–	BRCA mutant tumors	USP1 inhibitor
MDR-001	MindRank	Molecule Pro™	NCT not found	NCT06606483: not yet recruiting	Treatment of obesity and type 2 diabetes mellitus	GLP-1 Receptor agonist
NDI-034858 (TAK-279) (zasocitinib)f	Nimbus Therapeutics/Takeda	Not found	–	NCT05153148 (2b): completed	Active psoriatic arthritis/moderate-to-severe plaque psoriasis	TYK2 inhibitor
				NCT04999839 (2b): completed
NDI-101150	Nimbus Therapeutics	Not found	NCT05128487 (1/2): active, not recruiting	–	Cancer immunotherapy for solid tumors	ATP-competitive HPK1 inhibitor
PHI-101	Pharos iBio	Chemiverse Chemical	NCT04842370: unknown statusg	–	Relapsed or refractory AML	FLT3 inhibitor
PHI-101	Pharos iBio	Chemiverse Chemical	NCT04678102: unknown statusg	–	Platinum-resistance/refractory ovarian, fallopian tubal, and primary peritoneal cancer	CHK2 inhibitor
RLY-1971 (GDC-1971)	Relay Therapeutics/Genentech	Dynamo Platform™	NCT04252339: completed	–	Solid tumors	SHP2 inhibitor
			NCT05954871: active, not recruiting
			NCT05487235: completed
RLY-2608	Relay Therapeutics	Dynamo Platform™	NCT05216432: recruiting	–	Breast cancer and solid tumors	PI3KαPAN inhibitor
RLY-5836	Relay Therapeutics	Dynamo Platform™	NCT05759949: completed	–	Breast cancer and solid tumors	PI3KαPAN inhibitor
RLY-4008	Relay Therapeutics	Dynamo Platform™	NCT04526106 (1/2): active, not recruiting	–	Solid tumors	FGFR2 inhibitor
SNX-281	Stingthera/Merck	Not found	NCT04609579: terminatedh	–	Hematologic malignancies and solid tumors	STING activator
SGR-2921	Schrödinger	Not found	NCT05961839: recruiting	–	Hematologic malignancies and solid tumors	CDC7 inhibitor
SGR-1505	Schrödinger	Not found	NCT05544019: recruiting	–	Hematologic malignancies, solid tumors, and autoimmune diseases	MALT1 inhibitor
TOS-358	Totus Medicines	Not found	NCT05683418: recruiting	–	Solid tumors	PI3Kα inhibitor

^a Biomea Fusion, Inc. is no longer pursuing oncology indications for BMF-219. No safety concerns or efficacy observations led to this study closure.

^b Sponsor made a business decision to terminate the study based on prioritization of portfolio.

^c Voluntarily terminated the study since the sponsor's development strategy was adjusted.

^d Due to subject benefit-risk ratio changes, sponsor decides to voluntarily terminate study.

^e Emerging data demonstrates the challenge for EXS21546 to reach a suitable therapeutic index.

^f Has recently reached phase 3 clinical trials (Trial Nos.: NCT06671496 and NCT06671483).

^g Study has passed its completion date and status has not been verified in more than two years.

^h Terminated by sponsor decision, not due to safety concerns.

–: no data. NCT: national clinical trial; TACC3: transforming acidic coiled-coil-containing protein 3; PPI: protein-protein interaction; ER: estrogen receptor; AR: androgen receptor; BTK: Bruton's tyrosine kinase; WT: wild type; OX1R: orexin type 1 receptor; A2A: adenosine A2A receptor; 5-HT_1A: serotonin 1A receptor; CDK7: cyclin-dependent kinase 7; PKC-θ: protein kinase C theta; TNIK: TNF receptor associated factor 2 (TRAF2) and non-catalytic region of tyrosine kinase (NCK) interacting kinase; USP1: ubiquitin-specific protease 1; GLP-1: glucagon-like peptide-1; TYK2: tyrosine kinase 2; ATP: adenosine triphosphate; HPK1: hematopoietic progenitor kinase 1; AML: acute myeloid leukemia; FLT3: Feline McDonough sarcoma (FMS)-like tyrosine kinase 3; CHK2: checkpoint kinase 2; SHP2: Src homology 2 domain-containing protein tyrosine phosphatase 2; PI3KαPAN: phosphoinositide 3-kinase alpha pan inhibitor; FGFR2: fibroblast growth factor receptor 2; STING: stimulator of interferon genes; CDC7: cell division cycle 7-related protein kinase; MALT1: mucosa-associated lymphoid tissue lymphoma translocation protein 1.

INS018_055, for example, with an AI-driven generative design, is optimized for potency, safety, and oral bioavailability. Its clinical data shows promising results compared with existing drugs such as nintedanib. INS018_055 aims for once-daily dosing with improved tolerability, while nintedanib requires twice-daily intake and is limited by gastrointestinal adverse events. Both are orally administered, but the AI designed drug has optimized absorption and reduced food effect, which improves patient usability. In terms of safety, INS018_055 has reduced off-target kinase inhibition, which leads to fewer systemic side effects. Additionally, INS018_055 shows a trend toward improved forced vital capacity and favorable tolerability. Overall, this direct comparison shows that the existing drug nintedanib is effective, but it is limited by the side effect burden and dosing complexity. On the contrary, INS018_055 has the potential for better tolerability and simpler dosing, which leads to improved patient adherence in long-term treatment [122].

Another example from Table 3 is BMF-219, a covalent menin inhibitor with early signals of durable disease responses in menin-dependent leukemias and potential disease-modifying effects in diabetes through beta-cell restoration. For patients with menin-dependent cancers, BMF-219 offers a precision, oral option that may produce deep and durable remissions with less systemic toxicity and more outpatient-friendly dosing than standard cytotoxic regimens. However, this evidence should be taken carefully, since the evidence data is still from early phase 1 clinical trials [123]. In diabetes, if later-phase trials confirm, BMF-219 holds potential to shift care from chronic symptomatic control to a short course that restores beta-cell function, simplifying management and reducing injection burden. That is a high-impact potential advantage but remains investigational and contingent on confirming safety and durable efficacy [124].

The fast-paced advances seen in the use of AI technology in R&D represent a groundbreaking milestone that holds particular promise for the development of novel and effective treatments for several diseases, a process that is typically fraught with failure. Unsurprisingly, the majority of AI-generated small molecules are in the field of oncology and are being investigated as the next potential blockbusters. The emphasis on anticancer treatments is likely due to the high unmet medical needs, the abundance of biologically well-validated therapeutic targets and the commercial attractiveness to capture investment [107,125].

Considering how pivotal the molecules listed in Table 3 may become as future drugs for patients with a wide range of diseases, we delved into the search for the available intellectual property information. Table 4 is intended to provide more details on such molecules by elucidating their putative chemical scaffolds. The disclosure of the chemical scaffolds aims to shed light on the novelty of these compounds and therefore provide readers with information on where there is freedom to operate and develop new chemical structures without litigation risks.

Table 4.

Elucidation of putative chemical scaffolds for the artificial intelligence (AI)-discovered small-molecules in clinical trials.

Drug name	Patent number	Chemical scaffolds
A2A-252	WO 2022/221194 A1
BMF-219	WO 2020/142557 A1
AC-699 & AC-682	WO 2020/106933 A1 and WO 2021/118629 A1
AC-0176	WO 2020/214952 A1 and WO 2021/061642 A1
AC-0676	WO 2022/169798 A2
C4X3256 (INDV-2000)	WO 2016/034882 A1
EVOEXS-21546	WO 2019/233994 A1
DSP-1181	WO 2018/168738 A1
DSP-0038	US 10,745,401 B2
EXS-617	WO 2022/134641 A1
EXS-4318	WO 2022/234299 A1
INS018-055	WO 2022/179528 A1
ISM-0003091	WO 2024/086790 A1
MDR-001	WO 2023/029380 A1
NDI-034858 (TAK-279) (Zasocitinib)	WO 2020/112937 A1 and WO 2020/081508 A1
NDI-101150	WO 2021/050964 A1
PHI-101	WO 2019/231220 A1 and WO 2019/231221 A1
RLY-1971 (GDC-1971)	WO 2019/183367 A1
RLY-2608 and RLY-5836	WO 2021/222556 A1
RLY-4008	WO 2020/231990 A1
SNX-281	WO 2020/232378 A1 and WO 2020/232375 A1
SGR-2921	WO 2021/113492 A1 and WO 2022/055963 A1
SGR-1505	WO 2024/020534 A2
TOS-358	WO 2022/198024 A1

Given the length and cost associated with drug discovery, drug repurposing has emerged as a valuable approach to accelerate the pace of the R&D value chain, significantly reducing both the time and financial investment required. The AI-driven repurposing of existing approved drugs for previously unidentified therapeutic indications, through the use of big data and proprietary AI algorithms, unlocks further potential applications for these drugs, thus enabling the faster development of in-licensed drugs [119,126]. Table 5 systematizes examples of AI-repurposed molecules that have successfully progressed through various stages of clinical trials, covering a wide range of therapeutic indications.

Table 5.

Artificial intelligence (AI)-repurposed molecules in clinical trials.

Drug name	Company	Phase 1	Phase 2	Phase 3	First indication	Repurposed indication
Bavisant (BEN-200)	BenevolentAI	–	NCT03194217: completed	–	Attention-deficit hyperactivity disorder	Treatment of excessive daytime sleepiness in subjects with Parkinson's disease
Azelaprag (BGE-105)	BioAge Labs/Amgen	NCT06141889: completed	–	–	Heart failure	Increase overall weight loss and improve quality of weight loss vs. incretin drugs alone. Prevention of adverse muscle atrophy-related outcomes for older patients in the ICU
Talabostat (BXCL-701)	BioXcel Therapeutics	NCT03910660 (phase 1/2): active, not recruiting	–	–	Solid and liquid tumor malignancies	Small cell neuroendocrine prostate cancer
Talabostat (BXCL-701)	BioXcel Therapeutics	NCT04123574: withdrawna	NCT05558982: recruiting	–	Solid and liquid tumor malignancies	Metastatic pancreatic ductal adenocarcinoma
Talabostat (BXCL-701)	BioXcel Therapeutics	NCT05703542: recruiting	–	–	Solid and liquid tumor malignancies	Acute myeloid leukemia
Talabostat (BXCL-701)	BioXcel Therapeutics	–	NCT04171219: terminatedb	–	Solid and liquid tumor malignancies	Advanced solid cancers
IGALMI™/Dexmedetomidine (BXCL501)	BioXcel Therapeutics	NCT04010305 (1b): completed	NCT03708315: completed	NCT04276883: completed	Approved as IGALMI™ for the acute treatment of agitation associated with schizophrenia or bipolar I or II disorder in adults	Acute treatment of agitation associated with bipolar disorders/schizophrenia (at home)
		NCT05025605: recruiting		NCT04268303: completed
				NCT05658510: active, not recruiting
IGALMI™/Dexmedetomidine (BXCL501)	BioXcel Therapeutics	NCT04251910 (phase 1/2): completed	NCT05276830: terminated3	NCT05271552: completed	Approved as IGALMI™ for the acute treatment of agitation associated with schizophrenia or bipolar I or Il disorder in adults	Acute treatment of agitation associated with Alzheimer's dementia
				NCT05665088: terminatedc
IGALMI™/Dexmedetomidine (BXCL501)	BioXcel Therapeutics	NCT04470050 (phase 1/2): completed	–	–	Approved as IGALMI™ for the acute treatment of agitation associated with schizophrenia or bipolar I or Il disorder in adults	Opioid use disorder
		NCT05712707 (phase 1/2): active, not recruiting
IGALMI™/Dexmedetomidine (BXCL501)	BioXcel Therapeutics	NCT04827056: completed	–	–	Approved as IGALMI™ for the acute treatment of agitation associated with schizophrenia or bipolar I or Il disorder in adults	Post-traumatic stress disorder
Sulindac (HLX-0201)	Healx	–	NCT04823052: withdrawnd	–	Nonsteroidal anti-inflammatory drug	Fragile X syndrome
Fasoracetam (NB-001)	Nobias Therapeutics	–	NCT05290493: completed	–	Attention deficit disorder with hyperactivity	Neuropsychiatric symptoms of 22q11DS
REC-2282	Recursion Pharmaceuticals	–	NCT05130866 (phase 2/3): recruiting	–	pan-HDAC inhibitor	Progressive NF2-mutated meningiomas
Tolcapone (SOM0226/CRX-1008)	SOM Biotech/Corino Therapeutics	NCT02191826 (phase 1/2): completed	–	–	Parkinson's disease	Transthyretin amyloidosis in patients with familial amyloid polyneuropathy
Bevantolol (SOM-3355)	SOM Biotech	–	NCT03575676 (phase 2a): completed	–	Angina pectoris and hypertension	Chorea movements associated with Huntington's disease
			NCT05475483 (phase 2b): completed
SAR-407899/OPL-0401	Valo Health	–	NCT05393284: active, not recruiting	–	Microvascular angina (syndrome X)	Non-proliferative diabetic retinopathy and proliferative diabetic retinopathy

^a No longer relevant to field.

^b There is no partial response/complete response (PR/CR) observed in the first stage, so study stopped without proceeding to the stage 2 of efficacy stage.

^c Study was terminated for business reasons, not due to safety or efficacy concerns.

^d Healx has experienced delays to the site activation of the study and this has had an adverse impact on the recruitment timeline which is delaying the progress of other projects in their Fragile X Syndrome (FXS) program.

–: no data. NCT: national clinical trial; ICU: intensive care unit; 22q11DS: 22q11.2 deletion syndrome; HDAC: histone deacetylase; NF2: neurofibromatosis type 2.

As this review suggests, the application of AI to drug discovery is a widely expanding field with potential to play a likely transformative role in the state-of-the-art of current R&D methodologies. The spotlight that AI is assuming in the R&D value chain can be witnessed by the rapid rise in the number of high-value partnerships between well-established large pharmaceutical companies and “AI-native” companies. The latter have invested time and money into developing the software and tools that pharmaceutical companies do not have, while the former have a deep expertise in the industry and regulatory fields and are ready to share the risk [127,128]. The depth and breadth of knowledge resulting from these partnerships is designed to deliver high-quality drug candidates in a more expeditious way and with the highest likelihood of successful progression into clinical development. This holds great promise as a starting point for therapeutic breakthroughs, enabling the creation of industry-leading small molecule programs that address critical unmet medical needs for patients [127,128].

Although the use of AI in drug discovery remains in its infancy, recent investments and reported clinical trial successes provide initial proof of concept of its growing influence and potential to address the complexity of drug discovery. As the pool of AI-driven drug candidates entering clinical trials is expected to grow in the coming years, it will become clear how pivotal AI can be in streamlining the process of identifying and developing effective drug-like molecules and significantly reducing the financial losses and failures associated with the R&D drug discovery process.

4.2. Evolving regulatory frameworks for AI in drug development

As AI continues to evolve, it is poised to revolutionize the drug discovery landscape and accelerate the development of new medicines to improve patient care. However, despite the promise of AI to reshape the field of drug discovery, its adoption and integration into R&D pipelines still faces many challenges, most notably stringent regulatory considerations.

Applying existing regulatory guidelines to AI-driven drug discovery processes creates ambiguity and poses compliance challenges for drug developers. For instance, regulatory frameworks are still adapting to assess the specific technical risks and benefits of AI-based models. This adds further layers of complexity to the already lengthy and costly drug approval process, posing a challenge that hampers AI-developed molecules from reaching the market [129].

FDA acknowledges the increasing use of AI throughout the drug discovery lifecycle and its potential to expedite the development of safe and effective drugs across a range of therapeutic areas. In fact, the FDA has witnessed a remarkable increase in the number of drug and biological application submissions with AI components over the past few years [130,131]. Therefore, in alignment with its mission to protect, promote, and advance public health, the FDA released a discussion paper in May 2023. The paper aims to stimulate a debate among stakeholders regarding the use of AI and contribute to the development of an AI regulatory framework in R&D. The discussion paper provides an overview of considerations and good practices for the general application of AI and raises key questions soliciting feedback from stakeholders [130]. As a result of this discussion, the FDA issued a draft guidance in January 2025 that provides industry and other interested parties with recommendations on the use of AI to support regulatory decisions about the safety, effectiveness, or quality of a drug or biological product [[130], [131]].

This draft guidance fulfills the FDA's commitment of using AI in a way that ensures the agency's robust scientific and regulatory standards are met. It provides stakeholders with a risk-based, seven-step credibility assessment framework focused on the question of interest, context of use, model risk, credible planning, execution, documentation, assessment, and life-cycle maintenance. While it does not explicitly itemize criteria like transparency, reproducibility, or data provenance, these critical factors are embedded in the process, especially through the documentation steps and planning phases required to establish the credibility plan [132].

Echoing the FDA's efforts, the EMA is also paying attention to the rapidly evolving application of AI in the drug discovery lifecycle. Therefore, in July 2023, it has published a draft reflection paper, setting out the current thinking on the use of AI to support the safe and effective development, regulation, and use of human and veterinary medicines. This reflection paper is lifecycle-oriented and calls for sponsor-provided evidence on how AI is used at each stage. The aim is to open a dialogue with stakeholders to define the way forward and ensure that the full potential of these innovations can be realized for the benefit of patients' and animal health [133].

The China National Medical Products Administration (NMPA) has been very active in regulating AI-related medical devices, releasing six pivotal guidelines in 2023 [134]. These guidelines are significant in their scope and detail. They clarify the technical review requirements essential to this sector and are a step towards establishing a robust standard system. This system aims to enhance the development of China's growing AI industry in the medical field and amplify its international influence. The NMPA is also evolving its policies and technical review practices for drugs. However, the most concrete documents thus far have focused on AI tools used as devices or diagnostics [134].

In 2024, the UK's Medicines and Healthcare products Regulatory Agency (MHRA) outlined its strategic approach to AI, which is based on five key principles: safety, security, and robustness; appropriate transparency and explainability; fairness; accountability and governance; and contestability and redress [135]. In its role as a regulator of AI products, MHRA's principles are largely focused on medical devices with limited discussion of the medicine's framework. However, as a public service organization providing time-critical decisions, MHRA does discuss the use of AI in the medicine's lifecycle, but primarily in the context of improving the quality of applications. Overall, MHRA is using international recognition routes (relying on decisions from trusted regulators) and emphasizes balancing innovation with patient safety [135].

In February 2025, Health Canada released a pre-market guidance for ML-enabled medical devices. This guidance introduces the concept of a predetermined change control plan and provides Health Canada with a mechanism to address cases in which regulatory pre-authorization of planned changes to ML systems is required to mitigate known risks [136]. Regarding the use of AI in drug development, manufacturing, trial design, or pharmacovigilance, Health Canada expects sponsors to apply existing drug-regulatory requirements (evidence of safety/efficacy, data integrity, traceability, and post-market vigilance) and increasingly leans on cross-agency principles (FDA/MHRA) and Pan-Canadian AI for Health (AI4H) Guiding Principles [137].

Overall, the discussion of these subjects among regulators and stakeholders has created momentum toward regulatory readiness for AI-derived therapies. For this to happen, collaboration between pharmaceutical companies, “AI-native” startups, academic institutions, and regulatory agencies will be essential for the ethical implementation of AI in the drug discovery R&D value chain, fostering an environment where AI-derived drugs may become the standard [138].

In summary, although no AI-developed drugs have yet received approval to reach the market, the promising results reported for these drugs in the clinical trial landscape are encouraging. Such results strongly suggest that harnessing AI in R&D has great potential to help drug development companies improve the speed of approval and market access of their drugs, serving the ultimate goal of optimizing the drug development lifecycle. This could be the key to bringing critical drugs to market faster and more efficiently, ultimately revolutionizing the drug development process and patient care [138].

5. Challenges to harness the full potential of AI in drug discovery

Despite the significant progress and optimistic outlook surrounding AI methodologies in pharmaceutical R&D, the quest to unlock the full potential of AI in drug discovery is still in its early stages, with several challenges hindering its seamless integration into drug discovery processes. Bridging the gap between academic advances and real-world therapeutic outcomes demands targeted efforts to overcome domain-specific challenges.

Data quality, curation, and accessibility are the forefront challenges that warrant attention. The availability of reliable and comprehensive datasets is the cornerstone of effective AI-based drug development, as these are essential components for model training and validation [139]. While large chemical and bioactivity databases exist, those capturing protein-ligand binding affinities, ADMET profiles, and disease-specific phenotypes often suffer from inconsistency, sparsity, and measurement noise [140]. For instance, variations in binding affinity values across different assays (e.g., half-maximal inhibitory concentration (IC₅₀) vs. equilibrium dissociation constant (K_D)) complicate the training of predictive models and require careful harmonization strategies, such as converting different binding measurements to a common scale or training separate models for each measurement type [141]. Moreover, biased chemical space coverage by overrepresentation of drug-like molecules or easily synthesizable scaffolds can lead to overfitting and poor generalizability to underexplored chemical regions or novel targets. Moreover, widely used resources such as ChEMBL and PubChem, while extensive, contain redundant, missing, or misannotated entries. In practice, these challenges are often addressed through data cleaning processes such as deduplication, annotation correction, harmonization of molecular identifiers (e.g., InChIKeys), and application of outlier detection methods to identify potentially erroneous measurements prior to model training. Cross-database integration also benefits from standardized assay ontologies and mapping tools, which ensure comparability across datasets collected under different experimental conditions [142].

In addition, biased chemical space coverage, caused by the overrepresentation of drug-like molecules or easily synthesizable scaffolds, can lead to overfitting and poor generalizability to novel targets. Recent studies have highlighted the value of semi-supervised learning approaches that leverage large amounts of unlabeled data to mitigate sparsity and improve generalization [143], particularly when combined with active learning strategies that iteratively prioritize experimental validation. Finally, privacy concerns and intellectual property restrictions in the pharmaceutical industry continue to hinder the sharing of novel data. Federated learning and privacy-preserving data sharing frameworks are emerging as potential solutions to balance data accessibility with confidentiality [144].

Together, these strategies, ranging from database-specific cleaning and normalization protocols to semi-supervised and federated learning frameworks, represent practical directions for improving the quality and usability of datasets in AI-driven drug discovery.

A critical bottleneck that hinders the real-world adoption of DL methods in drug discovery is their inherent lack of explainability, making it challenging to obtain an appropriate interpretation of the model decision outputs. Given how pivotal it is for the pharmaceutical industry to have reliable and trustworthy results, the ability to understand the rationale behind AI-proposed clinical candidates is essential [140]. While strategies such as attention mechanisms, saliency maps, and post-hoc attribution methods have emerged to improve explainability, the field still lacks a consensus approach for robust interpretability that would streamline the regulatory approval. Domain-aware explainability, such as identifying which substructures or interactions contribute to predicted binding affinity, would significantly enhance adoption.

Recently, concrete examples have emerged explicitly focused on ensuring the interpretability of drug predictions and mechanisms of action [145]. For instance, attention-based transformer models have been applied as a tool to evaluate and generate candidate drugs. This information can be used to estimate the importance of each molecular region for the biological activity of the molecules [146]. Similarly, GNN-based property predictors can be coupled with node importance scores to locate pharmacophore-like substructures within molecules, providing chemists with interpretable links between model predictions and FGs [147]. Alienation maps have also been used to project contributions from binding affinity or protein-binding complexes, allowing domain experts to verify that the model emphasizes chemically meaningful interactions [148].

Regulatory agencies such as the FDA and EMA have explicitly called for interpretable and transparent AI models in drug development, emphasizing the importance of accountability in decision support systems for clinical candidates. Current guidelines emphasize the need for traceable predictions, human oversight, and auditable decision-making processes. Without interpretation, AI-assisted drug candidates encounter substantial obstacles to regulatory approval, regardless of their predictive performance.

Another central issue is the modeling of protein-ligand interactions, especially for targets with limited 3D structural information. While structure-based models, such as GNNs and docking-AI hybrids, have made strides, they still struggle with flexible binding pockets, water-mediated interactions, and metalloproteins. One promising direction has been the integration of multi-modal data, where molecular graphs or SMILES sequences are fused with protein representations derived from language models, evolutionary profiles, or 3D structural data. One integration strategy is cross-attention mechanisms, where ligand and protein embeddings interact via attention mechanisms, allowing the model to highlight complementary interaction regions. Another alternative is the graph-sequence embedding fusion, where drug molecules are represented as graphs and proteins as sequences, with joint embeddings learned through concatenation, bilinear pooling, or attention-based fusion. One example is GraphDTA, which uses graphs of drugs and protein sequences encoded via convolutional or recurrent layers, integrating them through concatenated embedding for drug affinity prediction [149]. MolTrans applies transformer-based encoders to both SMILES drugs and protein sequences, with interaction-aware attention layers that explicitly capture substructure-motif relationships, demonstrating improved interpretability and performance [150]. Other frameworks incorporate structural embeddings from AlphaFold2 or 3D surface descriptors to enhance predictions when experimental protein structures are missing.

Each approach has its advantages and disadvantages. Cross-attention models often yield greater interpretability, highlighting which ligand fragments interact with which protein motifs, but they require large datasets and are computationally demanding. Simpler embedding-fusion models are easier to train and widely applicable, but may underperform on complex binding scenarios. The 3D-aware fusion promises higher accuracy in flexible or novel binding sites, but it is limited by the quality and availability of structural data.

Moreover, ensuring that AI tools remain accessible and actionable for medicinal chemists, biologists, and regulatory decision makers is an ongoing challenge. The complexity of these models often results in opaque, inflexible structures that resist integration into existing drug development pipelines [148]. In addition, developing and implementing robust AI-based outputs requires expensive, highly parallelized computing power that threatens the limited budgets of smaller companies and academic laboratories who want to make a meaningful contribution to the drug R&D process. Such financial barriers further hamper the contribution of these players and prevent the creation of a diverse and impactful pipeline of novel compounds. Addressing these practical barriers to access and infrastructure is therefore as crucial as advancing algorithmic innovation. Both are necessary to ensure that AI can have a broad and equitable impact on drug discovery.

Another important challenge that warrants attention is the limited generalizability of computational models across different application domains and their insufficient robust experimental validation. It is common to find cutting-edge work that performs well within a narrowly defined training domain, but does not translate effectively to real-world problems. A more robust integration of wet-lab feedback, including in vitro binding assays, cytotoxicity testing, and early-stage ADMET profiling, into the model retraining cycle could serve as a form of active learning. One actionable strategy would be to develop closed-loop pipelines, where generated molecules are synthesized and tested, and experimental results directly inform the next generation of molecules via RL or Bayesian optimization frameworks. Moreover, the ethical and legal issues related to the intellectual property of AI-generated compounds must be considered. Questions of inventorship, data ownership, and liability for incorrect predictions arise when models are trained with poor quality publicly available data.

Finally, yet importantly, the regulatory landscape presents additional challenges. Although regulatory agencies such as the EMA and the FDA have begun to take steps to regulate the use of AI in drug development, a final guidance on this framework has not yet been issued. The official issuance of a risk-based regulatory guideline is essential to ensure the trustworthy use of AI computational models and to help maintain rigorous standards for product quality, safety, and efficacy. Furthermore, as mentioned in Section 4.2, the lack of clear guidelines is not just a procedural issue, but also hinders investment and validation. Companies and academic groups are uncertain about how AI-derived results will be evaluated in regulatory submissions. An official framework would therefore not only streamline compliance but also encourage broader adoption and resource allocation toward AI-enabled drug discovery [132,133,138].

6. Conclusions and future perspective

The transition from traditional drug discovery to AI-driven drug discovery presents several challenges, but remains an essential step for pharmaceutical companies seeking to lead in innovation and efficiency. Although AI may not be the sole game-changer in drug R&D, it can surely assist in addressing the root causes of drug failure and expediting the protracted process of bringing new drugs to the market. Nonetheless, the advances that have been made so far are likely to represent only a small fraction of the potential that AI can bring to the drug discovery landscape.

From a computational perspective, it is essential to invest in the development of scalable and generalizable methodologies, capable of handling intricate, chemically diverse, and multimodal datasets, including molecular graphs, protein sequences, gene expression, and phenotypic profiles. Future AI systems must be able to transfer knowledge across targets, operate within heterogeneous biological contexts, and address the intrinsically multi-objective nature of drug design, for example, balancing potency against toxicity or optimizing efficacy while maintaining selectivity and synthesizability. Recent developments in multi-modal learning promise to bridge these gaps by integrating data from multiple biological and chemical sources to yield more context-aware predictions.

Another promising direction is the generation of synergistic drug combinations, particularly relevant in oncology and infectious diseases. AI models are beginning to shift from single-compound design to the co-design of drug pairs or even triplets that maximize therapeutic synergy while minimizing adverse effects. This paradigm opens new opportunities in personalized medicine and polypharmacology.

Equally important is the integration of synthesizability-aware generation, ensuring that computational designs are not only effective in silico but also viable in real-world synthesis. This involves incorporating retrosynthetic analysis, differentiable synthesis scores, and dynamic feedback from synthesis planning tools directly into model training and generation cycles.

Current retrosynthesis tools are fundamentally constrained by biased reaction databases and template-based methods that are sensitive to small perturbations in input data and fail to generalize beyond well-documented chemistries, and lack generalizability beyond narrow contexts. These constraints propagate systematic biases into AI training, undermining predictive reliability, and limiting impact on unexplored chemical space. Overcoming these issues will require broader, better-curated datasets and the development of flexible, template-free methodologies capable of capturing true synthetic diversity.

The establishment of closed-loop pipelines, combining AI-driven molecule generation with experimental validation and iterative model refinement using real-world wet-lab data, is key to making AI truly actionable in drug development. These systems will enable adaptive learning from experimental failures, helping to narrow the gap between virtual design and clinical application.

Realizing the full potential of AI in drug discovery will require a collaborative effort among researchers from academia and industry, healthcare professionals, and policymakers to enable access to high-quality data, increase model transparency and explainability, and establish robust frameworks that meet the rigorous regulatory standards for drug quality, safety, and efficacy. Successful initiatives such as Machine Learning Ledger Orchestration for Drug Discovery (MELLODDY), which enabled multiple pharmaceutical companies to collaboratively train models on proprietary chemical libraries without sharing raw data, and Open Targets, which integrates genetics and genomics data to systematically identify and prioritize drug targets, illustrate how cross-sector partnerships can accelerate progress.

In the future, integrating AI with traditional drug discovery processes will usher in a transformative era in pharmaceutical research, propelled by emerging trends and breakthrough advances. As a result, the judicious use of AI will be critical to maintain a competitive edge in a rapidly changing industry.

CRediT authorship contribution statement

Rita I. Oliveira: Writing – review & editing, Writing – original draft, Visualization, Investigation, Conceptualization. Tiago O. Pereira: Writing – review & editing, Writing – original draft, Visualization, Investigation, Conceptualization. Maryam Abbasi: Writing – review & editing. Jorge A.R. Salvador: Writing – review & editing, Supervision. Joel P. Arrais: Writing – review & editing, Supervision.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

The following is the Supplementary data to this article: