Recent Advances in Automated Structure-Based De Novo Drug Design

Abstract

As the number of determined and predicted protein structures and the size of druglike ‘make-on-demand’ libraries soar, the time-consuming nature of structure-based computer-aided drug design calls for innovative computational algorithms. De novo drug design introduces in silico heuristics to accelerate searching in the vast chemical space. This review focuses on recent advances in structure-based de novo drug design, ranging from conventional fragment-based methods, evolutionary algorithms, and Metropolis Monte Carlo methods to deep generative models. Due to the historical limitation of de novo drug design generating readily available drug-like molecules, we highlight the synthetic accessibility efforts in each category and the benchmarking strategies taken to validate the proposed framework.

1. Introduction

Computer-aided drug design (CADD) methods have become more powerful as better hardware and novel methods like machine learning improve the performance of traditional tools.¹ Structure-based (SB) methods such as docking and molecular dynamics play a crucial role in CADD, enhancing our understanding of how small molecules bind to the protein target.² As more and more experimentally determined structures of therapeutic targets become available via X-ray crystallography, nuclear magnetic resonance (NMR) spectroscopy, or cryo-electron microscopy (cryo-EM), SB-CADD methods have sped up numerous drug discovery campaigns. The influence of SB-CADD methods expanded even more as homology modeling bridges the gap between similar protein sequences and determined structures.^3,4 The recent success of AlphaFold in the 14th Critical Assessment of Protein Structure Prediction (CASP) showed the feasibility of highly accurate large-scale structure prediction, leading to the extensive AlphaFold Protein Structure Database filled with more than 200 million structures.^5,6 Despite the improvement in computer hardware, the computational cost for evaluating a protein–ligand complex is still high, limiting the scope of assessment during hit searching. SB-CADD virtual screening campaigns can now screen ultralarge make-on-demand libraries containing millions of molecules, but this covers only a small proportion of the vast drug-like chemical space which is estimated to be up to 10⁶⁰ molecules.⁷ For a search problem of this scale, exhaustive search is infeasible. The situation calls for more efficient ways of exploration.

De novo drug design refers to a subset of methods that aim to design novel molecules with pharmacological properties from scratch.⁸ Compared with SB-CADD virtual screening, de novo design can explore a wider chemical space in a time-efficient manner. Similar to SB-CADD virtual screening, the molecules proposed from de novo design are usually still far from a final drug, but they serve as good starting points for medicinal chemistry to develop. A de novo drug design workflow generally consists of candidate sampling and property evaluation, usually in an iterative fashion. Ligand property evaluation is generally performed through various scoring functions and pharmacological filters. The sampling method, or molecular construction, is usually the main difference between design approaches.⁹

Various sampling methods have evolved significantly over the past years. The first SB-CADD de novo design method, LEGEND (1991), employed an atom-based sampling method, placing atoms and bonds successively in the receptor pocket to explore the chemical space.¹⁰ However, the combinatorial explosion associated with atom-based methods soon drove the field toward fragment-based methods and computing heuristics. Conventional fragment-based sampling methods employ three major strategies: growing, linking, and merging to develop binding fragments into complete drug molecules.^11,12 Evolutionary algorithms are a class of methods extensively used in de novo drug design. Mechanisms inspired by biological evolution are applied in these methods to optimize ligand population generation by generation.^13,14 Monte Carlo Metropolis (MCM) is another sophisticated heuristic for sampling in high-dimensional space. Such methods have been applied in drug discovery to search step-by-step in the chemical space for drug candidates.¹⁵ Recent advancements in machine learning (ML) have brought numerous deep generative models into drug discovery, combining and redefining the tasks in a de novo drug design workflow.¹⁶ This review introduces recent sampling methods ranging from conventional fragment-based methods and evolutionary methods to emerging ML methods (Figure 1).

Figure 1.

Schematic illustration of ligand sampling methods in structure-based de novo design. (a) Fragment-based growing, linking, and merging. (b) Evolutionary algorithms. (c) Monte Carlo Metropolis methods. (d) Deep generative neural networks.

The scoring task in SB-CADD de novo design requires a balance between accuracy and computing time to accomplish iterative searching in a vast library within a reasonable time. Major scoring functions include physics-based force fields, empirical potentials, and knowledge-based scoring functions. Compared with quantum-level calculations, these scoring functions are not only less accurate but also more lightweight in computing resources. Newly developed scoring functions are regularly assessed by their scoring and ranking power, as well as docking and screening power, as in the comparative assessment of scoring functions benchmark.¹⁷ Besides conventional scoring functions, ML scoring functions have emerged over the years, though the scoring performance has been questioned to be dependent on training sets.¹⁸

While predicted binding affinity is the most commonly used evaluation strategy, other metrics including structural similarity to known binders, binding mode, validity, novelty, diversity, and drug-likeness are also frequently employed during or after the design. Another important property for evaluation is synthetic accessibility (SA). SA has been a consistent challenge for the field since its inception. Though a number of methods in the past decade have discovered chemical entities that eventually proceeded to experimental validation, manual alteration to proposed designs prior to actual synthesis is still a frequent occurrence.¹⁹ Many recent approaches have tried to resolve the SA issue, and their efforts will be discussed in this review.

This review will also examine how recent approaches have benchmarked and validated their methods. Although experimental validation is widely agreed to be more convincing than in silico evaluation, few protocols have validated their designs in vitro or in vivo, mostly because of SA concerns. Instead, in silico validation frequently relies on docking studies and molecular dynamics (MD) simulation. Methods differ in the selection of evaluation metrics for benchmarking, as well as their choice of protein targets. These differences will be outlined for each method to compare similar protocols in parallel.

2. Growing, Linking, and Merging

Fragment-based drug discovery usually begins with a screen of diverse fragments, often by computational virtual screening, but sometimes with in vitro methods. Compared with larger and more complex molecules that are less likely to bind, fragments bind less tightly but more reliably.^11,12 These fragments are selected to form a library for further expansion into larger optimized molecules. Common fragment expansion strategies are growing, linking, and merging. Growing is the most commonly used strategy. Growing starts with a single core in the pocket, and subsequent additions of fragments then aim to extend the ligand into the rest of the pocket with improved affinity. Linking starts with two fragments occupying different nonoverlapping portions of the pocket, and the goal is to find a linker with suitable flexibility to maintain the original fragments’ binding modes. Merging combines the two fragments in different but overlapping parts of the pocket, with the common structure forming the core. Table 1 summarizes the program packages in the past eight years that adopted one or more of the fragment-based strategies.

Table 1. Recent Fragment-Based Ligand Design Packages.

Method	Ligand Construction	Synthetic Accessibility	Validation (method; target)
LigBuilder V3(20)	Growing/Linking/Genetic Algorithm	Retrosynthesis analysis	(Prototype) in vitro; COX2/LTA4H
			(Full protocol) in silico MM/GBSA method; HIV-1 protease/reverse transcriptase
NAOMInext(21)	Growing	Reaction-rule based	In silico docking/alignment; aurora A kinase, carbonic anhydrase II, acetylcholine-binding protein, protease factor VIIa
PINGUI(22)	Merging	Reaction-rule based	In vitro/In silico docking; β2AR
de novoDOCK(23)	Growing	Torsion environment from synthesizable database	In silico docking; HIVgp41
AutoT&T 2(24)	Merging	Real molecule reference library	In silico; angiotesin converting enzyme, VEGFR2, β-lactamase
OpenGrowth(25)	Growing	Fragment connection probability based on drug library	In silico MD; HIV-1 protease
			In vitro/In vivo; PDE3A-SLFN12 complex26
Frag4Lead(27)	Growing	Commercially available fragment database	In vitro/In silico docking; aspartyl protease endothiapepsin
LeadOp+R(28)	Growing	Reaction-rule based	In silico MD; Tie-2 kinase, human 5-lipoxygenase
AutoCouple(29)	Growing	Reaction-rule based	In vitro/In silico MD; CBP bromodomain

2.1. Ligand Construction

Fragment-based ligand construction usually begins with the selection and placement of anchors or growing centers. Most methods take a docked or cocrystallized ligand directly as the starting point^{21,22,24,25,28} or as the reference to lookup analog anchors from a given fragment library.^27,29 LigBuilder V3 has a de novo design mode called Chemical Space Exploring Algorithm, which performs iterative growing and fragment extraction operations on a pool of seed structures derived from a single sp3 carbon, hence avoiding preassigned seed structures and allowing broader exploration in the chemical space.²⁰De novo DOCK generates building block libraries including anchors from ZINC, breaking the molecules at each rotatable bond into rigid fragments, which are then oriented to the binding site via a graph matching algorithm.²³ In some methods, the user also needs to specify the sites of optimization on an anchor.²⁸ For most reaction-rule based methods, this step is unnecessary, since the reaction patterns automatically define the sites.

Building block sampling is a 2-fold problem: chemical space and conformational space sampling. Some fragment-based approaches therefore require docking of the fragment library to filter out undesired structures before ligand construction.^22,24 Fragment sampling for reaction-rule based methods is straightforward since candidates are restricted to reagents compatible with the reaction. Methods that are not reaction-rule based derive fragment connection probabilities from real molecules and use these to guide ligand construction.^23,30 AutoT&T2 limits the search space by searching for matched bonds between the reference library and the input lead molecule and carrying out a systematic crossover for all matches.²⁴ Frag4Lead collects hit analogs from commercial databases, with the common substructure aligned to the input hit.²⁷ Conformational sampling is typically through docking and in-site optimization of the product before the next iteration. NAOMInext generates conformations on the fly with a dynamic strategy that switches between breath-first-search and depth-first-search.²¹

Besides ligand flexibility, protein conformation is another factor to consider during design, though most methods only sample side-chain flexibility. OpenGrowth simultaneously grows ligands in several conformations of the protein, together with a rotamer search and geometry optimization on the chosen fragment, taking both protein and ligand flexibility into account.²⁵ LigBuilder V3 allows multitarget design, targeting multiple binding sites or multiple conformations of a protein. Specifically, the multitarget growing mode synchronously grows identical fragments at the same growing site to maintain 2D structure consistency, while 3D conformations are independently optimized in corresponding targets. The ensemble linking mode grows each fragment independently and flexibly before attempting to link among suitable ones.²⁰

2.2. Synthetic Accessibility

SA concerns in fragment-based methods usually arise at two stages: fragment source and fragment connection. The first is addressed through using a commercially available fragment library²⁷ or, at the very least, having a real molecule/drug library as reference.²³⁻²⁵ The second is addressed in some approaches by ensuring connection validity with reaction rules and reaction-based fragment libraries.^21,22,28,29 Fragment-based methods today are mostly iterative. The absence of SA supervision during fragment connection can lead to synthetically inaccessible molecules, especially at later iterations. Reaction-rule based approaches are therefore a general trend, particularly given the increasing number of reaction databases in recent years.³¹ Moving ahead, the concept of make-on-demand libraries presents promising prospects for the advancement of fragment-based methods. Enamine has generated a combinatorial library containing an impressive 36 billion readily accessible molecules through the enumeration of products from their existing compounds and reactions.³² This strategy significantly mitigates SA concerns at both the fragment source and fragment connection stages. These vast and dynamic libraries are anticipated to be instrumental in future fragment-based methods.

2.3. Validation

Methods reviewed here were validated at different levels (Figure 2). At the very least, the methods were tested to see if the known binders or analogs could be reconstructed from a set of precursors or seed structures. The successfully recovered molecules from these methods also adopted poses within 2.0 Å root-mean-square deviation (RMSD) to native structures.^21,23 However, such validation cannot test a method’s ability to generate novel molecules or to extract good ligands from a larger chemical space. Therefore, a majority of methods estimate the binding affinity of the compounds designed with generic databases, through docking or MD simulation. The reported methods mostly have top molecules estimated to be as potent or more potent than known binders or FDA-approved drugs.^20,25,28

Figure 2.

Levels of confidence for drug design validation.

A few methods verified their designed ligands in vitro. A prototype of LigBuilder V3 has been experimentally validated by designing a COX2/LTA₄H dual-functional inhibitor and yielded a ligand binding to COX2 and LTA₄H with an IC₅₀ of 7.1 and 7.0 μM, respectively.³³ PINGUI found four ligands targeting the β₂-adrenergic receptor (β₂AR) with improved binding affinity over the core fragments used as growing centers.²² Frag4Lead proposed ten binders targeting aspartyl protease endothiapepsin that have been confirmed by X-ray crystallography, with the best-proposed molecule improving affinity over its reference by 266-fold and two compounds having affinities below 10 μM.²⁷ Five molecules targeting CBP bromodomain and proposed by AutoCouple have been confirmed as nanomolar binders with the best one having a 10-fold increase in K_d compared with the known binder. The authors of AutoCouple then completed a series of optimizations to the proposed molecules to identify a hybrid molecule with good binding affinity and selectivity confirmed in vivo.²⁹ OpenGrowth was applied to generate anagrelide analogs as potential molecular glues between phosphodiesterase 3A (PDE3A) and Schlafen 12 protein (SLFN12). The in vitro apoptosis induction activities of 14 synthesized analogs increased significantly with the best one having an IC₅₀ of 0.3 nM. Further in vivo testing of the best compound also showed better tumor growth inhibition than the known drug anagrelide.²⁶

For reaction-rule based methods, it is also important to confirm the validity or viability of the proposed synthetic routes. While molecules tested in vitro automatically prove their SA, methods lacking experimental validation have sought verification from the literature. For example, the synthetic routes suggested by LeadOp+R match past experimental studies in the literature, and nine of their proposed molecules for Tie-2 kinase were found to have been synthesized before.²⁸ However, such validation cannot prove the SA of novel molecules that have not been synthesized.

3. Evolutionary Algorithms

Evolutionary algorithms (EAs) are powerful approaches for solving search and optimization problems that involve multiple, conflicting objectives. They mimic the concept of Darwinian evolution in that the fittest molecules are selected generation by generation. The genetic algorithm (GA) is the most commonly used type of EA.³⁴ Other types include genetic programming, evolutionary strategy, and evolutionary programming.

An EA starts with an initial parent population (often randomly chosen). In the case of drug design, the initial population is usually a set of chemical compounds. Random, biological evolution-inspired operations such as reproduction, mutation, selection, and crossover are applied to individuals in the parent population to produce the “children”. Mutation introduces new information into the population, and crossover combines information from existing individuals to generate new populations. GA also employs the replication (or elitism) operator to carry the fittest molecules unchanged into the next generation. All “children” structures in the new population not only will be evaluated with a fitness function, usually in some form of the binding affinity, but can also involve properties such as drug-likeness, toxicity, and similarity to known actives.

Various selection strategies involving the “Roulette wheel”, “Tournament”, and “Ranking” are employed in each round to select a diverse set of fit molecules to function as the parent population in the next round. “Roulette” assigns an area weighted by fitness to each proposal on a metaphorical roulette. By giving each proposal a chance, roulette introduces randomness into each generation, and the exploration is less likely to be trapped in a local minimum. “Tournament” randomly samples a subgroup from the proposals and picks the fittest ones. “Ranking” directly chooses the best-scoring proposals but has the risk of selecting low diversity compounds at later generations with high convergence.³⁵ While each strategy has its pros and cons, many EA methods incorporate more than one strategy to balance randomness, fitness, and diversity.

The iteration of offspring generation, evaluation, and selection continues until a user-set termination criterion is met, at which point the molecules will have converged to a set of locally optimized “fittest” compounds, substantially better than the initial pool. With independent runs of the EA starting from different sets of initial populations, the vast chemical space can be efficiently explored. EAs have been widely used in de novo drug design over the past two decades. Here, we summarize some recent EA methods to complement existing reviews^9,13,14 on this topic (Table 2).

Table 2. Recent De Novo Drug Design Methods Using Evolutionary Algorithms.

Method	Ligand Construction	Synthetic Accessibility	Validation (method; target)
Dock_GA(36)	GA	Torsion environment from synthesizable database	In silico docking; protein ligand complexes from SB2012 testset, SAR-CoV-2 Mpro
SECSE(37)	Rule-based/GA	Retro-synthesis module	In silico docking; phosphoglycerate dehydrogenase
EMGA(38)	Evolutionary Strategy/Transformer ANN	SA scores	In silico MD; SAR-CoV-2 Mpro
Steinmann etal.(39)	Graph-based GA	SA scores and filers	In silico docking; chorismate mutase, β2AR, DDR1, β-cyclodextrin
			In vitro; SAR-CoV-2 Mpro
AutoGrow4(35)	GA	Reaction-based mutation	In silico docking; poly(ADP-ribose) polymerase 1
MoleGear(40)	Graph-based EA	None	In silico docking/alignment; HIV-1 protease

3.1. Ligand Construction

A distinguishing aspect between methods is molecular representation. The structure of chemical entities resembles graphs in computer science, naturally leading to a molecular graph representation, where atoms and bonds correspond to nodes and edges. In contrast, the simplified molecular input line entry specification (SMILES) is a linear string representation that is derived from the molecular graphs.⁴¹ The difference in molecular representation is reflected in EA molecular construction operations: Graph-based methods alter a graph representation of the molecule,^39,40 while SMILES-based methods modify a string representation.^35,37,38 An advantage of using SMILES representation is rapid reaction-rule based modification through SMILES arbitrary target specification (SMARTS), as exemplified by Systemic Evolutionary Chemical Space Explorer (SECSE)³⁷ and AutoGrow4.³⁵

The initial population is often randomly drawn from drug-like libraries such as ZINC.^35,39 Elend et al. trained a neural language model on a subset of the ZINC database to generate initial SMILES strings character by character.³⁸ Dock_GA, SECSE, and MoleGear are fragment-based methods. Dock_GA generates a fragment library with the same infrastructure as de novo DOCK.^23,36 SECSE proposed a fragment generation algorithm that can enumerate up to twelve heavy atom fragments and build up a collection containing 121 million fragments as the starting point of the workflow.³⁷ MoleGear uses a fragment library generated from 1990 compounds selected from the National Cancer Institute diversity set.^40,42

Structural operators vary slightly between the methods. Mutation is the most common operator that performs structural transformations like ring-open, ring-closing, atom insertion, or deletion, etc. In the Evolutionary Molecular Generation Algorithm (EMGA) proposed by Elend et al., the neural language model also serves as a mutation operator, and it randomly deletes, adds, and replaces atoms in a SMILES.³⁸ AutoGrow4 has a reaction-rule based mutation operator using 36 click-chemistry reactions from AutoClickChem,⁴³ 58 reactions published by Hartenfeller et al.,⁴⁴ and any user-defined sets.³⁵ In comparison, SECSE has both a classical mutation operator and a reaction operator. Elitism, as a feature of GA, is also common among the methods. SECSE introduced a graph-based deep learning module trained with docked samples of the population and can subsequently assess the quality of the rest of the population, speeding up the elite selection.³⁷ The fragment-based methods also have a growing operator, similar to conventional fragment-based growing. Crossover is not as frequently employed, possibly due to its higher computational cost than other operators. Dock_GA has a 3D crossover operator that constructs molecules in the binding site environment.³⁶ Besides the typical operators, SECSE also has a bioisostere operator that allows the interconversion of classical or nonclassical bioisosteric replacements.³⁷

Fitness assessment for SB-CADD EA methods need to be fast since a large pool of molecules is proposed with every new population. Besides docking, similarity and diversity scores are also common fitness metrics. AutoGrow4 includes a diversity score that measures a molecule’s uniqueness relative to the others in the generation as an optional secondary metric.³⁵ SECSE and MoleGear use similarity to reference known binders as the fitness metric in a mode parallel to the docking evaluation, therefore enabling ligand-based drug design as an option.^37,40 SA metrics are also sometimes included in addition to the primary docking fitness score. SECSE includes a retrosynthesis module in its fitness evaluation,³⁷ while Elend et al. and Steinmann et al. have an SA score component in their fitness functions.^38,39 Other properties like drug-likeness are usually set as filters before docking,^35,37 but there are exceptions such as Elend et al. which includes the drug-likeness and toxicity as weighted score terms in its fitness function.³⁸

3.2. Synthetic Accessibility

Because of the nature of EA, it is harder to keep track of SA compared with other fragment-based growing or merging methods. The mutation and the crossover operators introduce complexity into the formation of new ligands, making reaction-rule based solutions hard to apply. Nevertheless, existing methods have incorporated reaction rules into the mutation operator³⁵ or as a separate operator.³⁷ Iterations of operations on a population also require the SA consideration to be fast since numerous molecules are assessed in the process. As a result, several approaches make use of SA scores as the solution.^38,39 SA scores are metrics that measure the molecular complexity and are able to rank or filter large collections of molecules in a mere time.^45,46 Retrosynthesis analysis is a more resource-intensive way to evaluate SA and is often utilized in postgeneration inspection.³⁷ Although recent EA methods consider SA in one or more of the above directions, it is not uncommon to obtain ligands with poor SA scores or lengthy synthetic routes, as brought up in several methods. Top molecules proposed by SECSE, for example, were predicted to be synthesizable within 15 steps.³⁷ Such concerns directly hinder the in vitro validation of the proposed molecules and limit the methods’ applicability.

3.3. Validation

Molecular docking was extensively used in the reviewed EA methods to report both the predicted binding pose and the predicted binding energy of the proposed molecules compared to a reference compound. The majority of the methods was able to propose molecules with similar predicted pose to the reference and better predicted binding energies. Steinmann et al. also compared the method’s performance to conventional high-throughput SBVS. All molecules in the ZINC subset where the initial population was sampled from were docked, and the top scores are compared to that of the generated molecules. For the case of DDR1 and β₂AR, the reported methods found 1.9 times as many molecules with a good docking score (<−9.0) relative to known binders (−6.8/–6.9) by docking only 1.6 times as many molecules compared to SBVS.³⁹ Besides docking, some methods also used MD simulation on a filtered list of top molecules to predict the binding energies, which is computationally more demanding.³⁸

In vitro validation of the proposed molecules is generally lacking in these papers, most likely due to SA concerns. An early version of the method by Steinmann et al. was applied to SARS-CoV-2 M^pro in the COVID Moonshot project in 2020.⁴⁷ One out of 10 submitted molecules proceeded to experimental validation but was later shown to have low inhibition.³⁹

4. Monte Carlo Metropolis

Monte Carlo methods are a class of computational algorithms that solve problems through iterative random sampling. It has been extensively used in many optimization problems and sampling from probability distributions. The Metropolis criterion, which decides if the new state of each iteration is accepted or rejected, is often combined with the sampling. Monte Carlo Metropolis (MCM) methods are far from uncommon in CADD, with their utilization in fields ranging from molecular docking to small molecule drug discovery. For example, RosettaLigand has been a very successful example of applying MCM to flexible ligand docking.^48,49 The employment of MC methods in de novo drug design can be traced back to 1991 in LEGEND,¹⁰ and the MCM concept has since been incorporated into the drug design workflows repeatedly during the 1990s.⁵⁰ Yet at the time, most of these workflows adopted an atom-based ligand construction model, which suffered from exploding combinatorial search space concerns.^9,15 For two decades, as fragment-based methods become more common and complement the shortcomings of early atom-based approaches, MCM methods have been silent in de novo drug design until recently.

In 2017, Oglic et al. designed a Metropolis-Hastings Markov Chain MC (MCMC) method which is then updated in the following year to perform active search in the chemical space.^15,51 With a moderately active parent compound, candidates represented using vertex-labeled graphs are generated by substitutions at specific sites. The proposal generator includes various filters such as Lipinski’s rule and prohibition of specific synthetically inaccessible substitutions. The evaluation oracle performs rigid docking of a large number of ligand conformations. Binary feedback is returned based on a docking score threshold, and this feedback is incorporated in the Metropolis-Hastings criteria. The method was tested against integrin receptors that are important in idiopathic pulmonary fibrosis. A known inhibitor was used as the parent compound and, with constraints, defined the design space to around 185,000 compounds. The method was able to recover 19 out of 26 known actives with predicted binding affinity more favored than the parent compound. Although the result is encouraging, the SA considerations during molecular generation are weakly implemented, and the authors plan to incorporate actual reaction information into the algorithm in the future.

Xie et al. published an LB approach in 2021 that employs Metropolis-Hastings MCMC and graph neural network (GNN) to perform multiobjective drug discovery.⁵² Though being a novel attempt to combine conventional algorithms with machine learning tools, this approach named Markov Molecular Sampling is however beyond the discussion of the current review which focuses primarily on SB-CADD de novo design methods.

5. Machine Learning

Deep learning (DL) is a subclass of ML that incorporates multilayers of artificial neural networks (ANNs) to represent data in a rather complex latent space. Like other de novo design methods, DL methods need to solve the problems of molecular generation, property prediction, and molecular optimization, which are also the key differences between different DL methods.⁵³ The success of deep generative models in other fields including natural language processing and computer vision has inspired the utilization of these sophisticated models in de novo drug design.

Molecular representation is an important aspect of DL de novo design methods, as it decides how a molecule is interpreted by the generative model. SMILES and graphs account for most of the two-dimensional (2D) representations used in the deep generative models in de novo drug design. Yet incorporating protein–ligand interactions with 2D representations is challenging, and years of efforts have mostly been LB, learning information primarily from known actives. Recently, several SB-CADD DL methods have been published to take advantage of the protein structures and build models trained on generic or target-specific databases to learn the intrinsic rules of protein–ligand interactions. Although some of these SB-CADD methods still employ the 2D representations with conformation generation processes to sample in the three-dimensional (3D) space, others adopt a 3D generative model where the configuration and conformation of a molecule are sampled simultaneously inside a protein pocket. These 3D generative models often need a 3D featurization for both the ligand and the protein. Such featurization includes cubic grid-based, Euclidean distance matrix (EDM)-based, and Cartesian coordinate-based, which have been reviewed in detail before.¹⁶

Over the past decade, there has been an increasing interest in DL methods, which can be seen by the soaring number of papers. There has been plenty of discussion on DL de novo drug design in recent reviews, both LB and SB.^9,16,53 For this review, we focus specifically on the SB-CADD DL methods (Table 3).

Table 3. Recent Structure-Based Deep Learning De Novo Drug Design Approaches.

Method	Ligand Representation	Molecular Generation	Validation (method; target)
DiffSBDD(54)	3D coordinates	Diffusion	In silico docking; 100 proteins from CrossDocked2020 and 130 complexes from Binding MOAD
RELATION(55)	3D property grids	VAE/AAE	In silico docking; protein kinase B alpha, CDK2
Ragoza et al.(56)	3D property grids	VAE	In silico docking; 10 random proteins from CrossDocked2020 Mutation study, Shikimate kinase
DeepLigBuilder(57)	3D coordinates	MPNN	In silico docking; SARS-CoV-2 Mpro
SBMolGen(58)	SMILES	RNN	In silico FMO and MD; CDK2, EGFR, AA2AR, ADRB2
MolAICal(59)	SMILES/graphs	Sequence-based/GNN	In silico MD; glucagon receptor, SARS-CoV-2 Mpro
Xu et al.(60)	SMILES	RNN	In silico docking; mitogen-activated protein kinase 14
DEVELOP(61)	Graphs	GNN, CNN	In silico docking; menin-MLL
LiGANN(62)	3D property grids	BicycleGAN	In silico docking; delta opioid 7TM receptor, CHK1, TNNI3K, and IRAK-4 kinase
Armstrong et al.(63)	Graphs	GCN, VAE	In silico docking; protein–ligand complexes from scPDB
Grechishnikova etal.(64)	SMILES	Transformer	In silico docking; Insulin-like growth factor 1 receptor, VEGFR2
cMolGPT(65)	SMILES	Transformer	In silico QSAR prediction; EGFR, HTR1A, S1PR1
Luo et al.(66)	3D coordinates	GNN	In silico Docking; 100 proteins from CrossDocked2020

5.1. Ligand Construction

The methods summarized here employ a variety of ANNs to generate ligands. The underlying difference is ligand representation (Figure 3a). SMILES-based methods generate strings character by character within a chemical context. Recurrent neural network (RNN), a sequence-based model commonly used in text generation tasks, serves this purpose well and for many years has been employed as the generative model in SMILES-based methods.^58,60 More recently, transformer-based approaches have achieved success in many sequence processing tasks. The self-attention mechanism of transformer allows long-range dependencies, while also being faster than recurrent networks.⁶⁷ Several groups have incorporated the transformer model into structure-based drug design.^64,65 Graph-based methods represent atoms and bonds in nodes and edges, intuitively leading to the utilization of graph neural networks (GNNs) that process graph data.^59,61,63 3D-based methods are mostly cube grid-based, where the molecules are translated into property grids such as elements, aromaticity, hydrogen bond donors and acceptors, formal charge, etc. In these cases, a variational autoencoder (VAE)^55,56 or a generative adversarial network (GAN)⁶² can perform the feature extraction and generate latent vectors that correspond to 3D druglike molecules. Then a decoder, usually a long short-term memory network, is required to transform these vectors into readable formats such as SMILES.^55,62 For the same decoding problem, Ragoza et al. implemented an atom fitting algorithm that combines iterative atom detection with gradient descent to deduce a 3D molecular structure from a density grid.⁵⁶ Besides grid-based featurization, other 3D-based methods adopt Cartesian coordinate-based representation, which generates rotationally and translationally invariant 3D embeddings that lead to full and unambiguous 3D structures.^54,57,66 DiffSBDD implemented a diffusion model together with molecule inpainting to generate structures within the molecular context.⁵⁴ DeepLigBuilder introduced a novel graph generative model, consisting of a state encoder with Message Passing Neural Network (MPNN) architecture and a policy network, to iteratively generate valid 3D druglike structures.⁵⁷ Luo et al. used an autoregressive algorithm to sample atoms sequentially from a changing probability density, leading to unambiguous and multimodal ligand outputs.⁶⁶

Figure 3.

(a) Common molecular representations and deep learning models utilized in structure-based deep learning de novo design approaches. (b) Protein information can be included in ligand generation in two ways: conditioned generation and docking-based backpropagation.

A significant difference between the methods is how the protein information is incorporated into ligand generation, as per the definition of SB. The methods reviewed here lie within one of the two categories: generation conditioned on target receptor and docking score backpropagation (Figure 3b). In the first type, the receptor is transformed into constraints or additional variables and integrated into the ligand generation model. How the receptor is represented or interpreted differs between methods, but a majority adopt a similar representation as the ligand representations mentioned above, and several methods encode the protein into constraint vectors via a convolution neural network (CNN).^56,60−62 DiffSBDD used two strategies for protein conditioning: One considers the protein as a fixed 3D context during denoising of the diffusion model; the other learns the joint distribution of protein–ligand complexes.⁵⁴ RELATION extracted pharmacophore features from crystal structures and the constraints were based on the root-mean-square of the matched feature pair distance and the number of matched pairs.⁵⁵ Xu et al. transformed the binding site with a coarse-grained strategy using the sorted eigenvalues of the Coulomb matrix descriptor.⁶⁰ Armstrong et al. represented the binding site in a graph-theoretic manner and trained their graph convolutional network (GCN) with EDM-based representation.⁶³ The second type of method utilizes a common concept in DL called backpropagation to bias the ligand generation through a docking-based reward function. A Monte Carlo tree search (MCTS) can be combined into the molecular generation to optimize any intermediate state while guided by docking score, as done by DeepLigBuilder⁵⁷ and SBMolGen.⁵⁸ Apart from the two categories, MolAICal’s generative model involves no protein information and outputs fragments which then undergo classical SB-GA to generate complete ligands.⁵⁹ The transformer proposed by Grechishnikova et al. considers the drug design problem as a translation from the amino acid sequence to SMILES and therefore needs only the protein sequence as the input.⁶⁴

DL methods generate molecules through sampling in the latent space learned by the model. Unlike the real chemical space which is discrete, the latent space contains infinite points. DL generative models therefore can output an enormous number of molecules after training. These outputs need to be filtered and evaluated in a fast manner. Validity is a common metric to check in DL methods due to the continuous nature of the latent space, usually with cheminformatic toolkits like RDKit. Today most models can reach 80–90% validity by training on large chemical databases. Other filters and metrics include novelty, uniqueness, drug-likeness, and SA, as in conventional de novo design methods. Affinity prediction after molecular docking of the filtered molecules can be done with DL-based scoring functions trained with generic protein complex databases, as seen in several methods.^56,59,63 To facilitate sampling and improve the quality of final outputs, RELATION employed Bayesian optimization with docking scores or quantitative structure–activity relationship (QSAR) scores and greatly improved the performance.⁵⁵

5.2. Synthetic Accessibility

Despite the popularity, DL methods still face SA challenges. In fact, few DL methods consider SA in their protocol design and validation. It is somewhat understandable since layers of neural networks form a “black box” that makes it hard to incorporate SA measures. Some methods apply SA scores as a filter or metric to optimize during molecular generation.^58,63 Most methods set the training data set to real molecules or drugs, expecting that SA can be a feature inherently learned. Another potential solution is multiobjective optimization. For most current deep generative models, predicted binding affinity is the primary and only objective to optimize during molecular generation. A multiobjective model would be able to optimize molecules based on several metrics other than predicted binding affinity, for example, drug-likeness and SA scores. Such multiobjective models have been realized by Armstrong et al. and also mentioned in the future work of several other methods.^55,58,63

5.3. Validation

DL methods are black boxes. It is therefore important to test how each part of the model contributes to the overall performance during benchmarking. In SB-CADD de novo design, protein structural information is incorporated in many ways as discussed above. How much did the protein participate in molecular generation and did it bias the process as anticipated? This was investigated by several methods, especially those with a model conditioned on receptors. These methods built uncontrolled models that do not have access to protein information and compared the performance to the controlled models, usually by looking at recovery rate or similarity to known binders.^60,61 Ragoza et al. answered these questions by evaluating the effect of mutation at the pocket on the outputs. For this purpose, shikimate kinase was mutated at interacting and noninteracting residues. The generative model responded to the pocket variants and generated molecules with corresponding changes.⁵⁶ Methods that used docking-based MCTS molecular optimization compared the performance with and without MCTS.^57,58 Some methods also looked at the effect of including known binders, either during training or as part of the input, where the generation process is much like lead optimization.^57,60 Ragoza et al. designed a bias-toward-reference factor in their method that switches the mode between de novo design and lead optimization. In their benchmark, the latent space was interpolated by varying the bias factor during sampling, producing a series of ligands from novel ones to analogs.⁵⁶

Further evaluation of DL methods proceeds analogously to conventional methods. In some cases where a 3D conformer can be directly generated, the quality of the generated pose can be compared with a docked pose. Ragoza et al., for example, compared the poses before and after minimization, and less than 20% of the generated molecules moved more than 2 Å RMSD, indicating that the majority of the molecules has a stable conformation in the pocket even before minimization.⁵⁶ Pharmacophore recapture is also a common evaluation, especially for the cubic grid-based methods where properties are specifically encoded into molecular generation.^55,57,61 Affinity prediction is typically done by molecular docking or MD simulation. The baseline performance is usually random decoys from the training database, and known binders are the next level of comparison. Most methods are able to generate molecules predicted to be as potent as known binders, and some are able to propose even better molecules. More than 15% of the generated molecules by Ragoza et al. have better predicted affinity than the reference molecules.⁵⁶ SBMolGen was able to generate molecules with better predicted binding affinity than known actives for cyclin-dependent kinases 2 (CDK2), epidermal growth factor receptor erbB1 (EGFR), adenosine A2a receptor (AA2AR), and beta-2 adrenergic receptor (ADRB2).⁵⁸ One-third of the molecules generated by DEVELOP and targeting menin and mixed lineage leukemia (MLL) fusion proteins have a predicted binding affinity greater or equal to the ground truth molecules.⁶¹ The method proposed by Luo et al. generated on average more than 29% of molecules with higher predicted affinity than reference ligands over 100 proteins from the CrossDock2020⁶⁸ data set. Their method was also successful in linker prediction, recovering 48% of the test molecules.⁶⁶

6. Conclusion

Since the rise of computer-based de novo design in the 1990s, methods in the field have evolved rapidly. De novo design has revolutionized drug discovery by developing in silico heuristics to speed up searching in the vast chemical space.⁵⁰ The advantages of de novo design become more obvious when fragment-based libraries further speed up search. The relevance of SB-CADD de novo design is increasing as more and more crystal structures and homology models are available. The direct inclusion of target information in the search process makes the proposed molecules and their predicted interactions more exact and specific. As docking methods and computer power continue to improve, protein flexibility is not as obstructive as it once was, and new methods in the field all are able to address this. By providing novel scaffolds and constructive structural ideas, de novo design has aided medicinal chemists in developing patentable leads with desired properties.¹⁹ A great example of application is COVID-19, where de novo drug design methods made a rapid response to the newly discovered disease, yielding numerous novel drug molecules and reducing the time of development for treatment. This topic has been extensively reviewed elsewhere,⁶⁹ and some of the methods that targeted COVID-19 are also included in this review.^{36,38,39,57,59}

Despite the advances in the field, there are still challenges yet to be solved. Scoring remains a limiting factor for SB-CADD methods, with SB-CADD de novo design relying heavily on the performance of scoring functions. Conventional scoring methods behave poorly in screening, giving a low hit rate and many false positives.^17,70 ML scoring functions are limited by the scope of their training sets, making scoring novel targets with few known binders unreliable.¹⁸ Quantitatively assessing the protein–ligand interactions in an accurate and fast manner is critical for the success of these methods. Iterative searching, scoring, and optimization in a vast library require a balance between accuracy and computing time. It can thus be foreseen that future advances in scoring functions will effectively improve the performance of SB-CADD de novo design methods.

SA is another common issue. Fragment-based methods try to overcome the problem by using druglike fragments and reaction-rule based molecular generation, but the SA diminishes during the process of iterative optimization. Other methods employ some sort of SA scores, sometimes as metrics to optimize and most times as filters. These scores measure the molecular complexity, usually calculated from a fixed set of molecules, and hardly agree with each other or the medicinal chemists. DL methods have even more concerns about SA, despite their popularity. The synthesizability of the training data set restricts the SA of the output molecules from the generative models, and SA is usually not an objective during training.⁷¹ This also calls for multiobjective optimization in the next stage of DL methods, where binding affinity is not the only metric to optimize, and more properties bias the molecular generation toward more druglike and synthetically accessible molecules.

From the reviewed methods, we also observed a lack of standardized benchmark workflow. A de novo design method should validate its proposed molecules experimentally, but such validation is rarely performed in these methods, partly because of the SA concerns mentioned above. Even in silico, there is no unified benchmark strategy in the field. Common evaluation strategies include structural and binding mode similarity to known binders, predicted binding affinities compared to known binders or random decoys, novelty and diversity (and validity for DL methods), and various drug-likeness metrics. The reviewed methods adopt one or more of these strategies, validating their proposed framework at different levels, making it difficult to compare the performance of different methods. With the rapid advance in the field and the emergence of numerous novel approaches, future methods should have more comprehensive benchmarks to convince the scientific public. Alternatively, public benchmarking exercises like Critical Assessment of Computational Hit-finding Experiments (CACHE) provide the community with opportunities to test the computational methods experimentally and under a standardized setting.⁷² Results from these exercises will be collected and released to the public, serving as valuable resources to guide further advancement. As the community grows with the abundance of available structures, we hope more SB-CADD de novo design methods can be adopted to facilitate future drug development.

Table 4 contains data and software availability.

Table 4. Data and Software Availability.

Method	License	Source
LigBuilder V3(20)	Free for all	http://repharma.pku.edu.cn/ligbuilder3/
NAOMInext(21)	Free for academic	http://uhh.de/naomi
PINGUI(22)	Web application	www.kolblab.org/scubidoo/pingui
de novoDOCK(23)	Free for academic	https://dock.compbio.ucsf.edu/
AutoT&T 2(24)	300 USD for academic;	http://www.sioc-ccbg.ac.cn/software/att2/
	3000 USD for industrial
OpenGrowth(25)	Free for all	http://opengrowth.sourceforge.net/
Frag4Lead(27)	Not public	N/A
LeadOp+R(28)	Not public	N/A
AutoCouple(29)	Unknown	Scripts available at https://github.com/Caflisch-Group/AutoCouple_Python-based
Dock_GA(36)	Free for academic	https://dock.compbio.ucsf.edu/
SECSE(37)	Open source	https://github.com/KeenThera/SECSE
EMGA(38)	Not public	N/A
Steinmann et al.(39)	Open source	https://github.com/cstein/GB-GA/tree/feature-glide_docking
AutoGrow4(35)	Open source	https://durrantlab.pitt.edu/autogrow4/
MoleGear(40)	Not public	N/A
Oglic et al.(15)	Not public	N/A
DiffSBDD(54)	Open source	https://github.com/arneschneuing/DiffSBDD
RELATION(55)	Unknown	https://github.com/micahwang/RELATION
Ragoza et al.(56)	Open source	https://github.com/mattragoza/liGAN
DeepLigBuilder(57)	Not public	N/A
SBMolGen(58)	Open source	https://github.com/clinfo/SBMolGen
MolAICal(59)	Free for academic	https://molaical.github.io/
Xu et al.(60)	Not public	N/A
DEVELOP(61)	Open source	https://github.com/oxpig/DEVELOP
LiGANN(62)	Web application	https://playmolecule.com/LiGANN/
Armstrong et al.(63)	Not public	N/A
Grechishnikova et al.(64)	Unknown	https://github.com/dariagrechishnikova/molecule_structure_generation
cMolGPT(65)	Unknown	https://github.com/VV123/cMolGPT
Luo et al.(66)	Open source	https://github.com/luost26/3D-Generative-SBDD

Glossary

Abbreviations

AA2AR

adenosine A2a receptor

AAE

adversarial autoencoder

ADRB2

beta-2 adrenergic receptor

ANNs

artificial neural networks

SB-CADD

structure-based computer-aided drug design

CDK2

cyclin-dependent kinases 2

CNNs

convolutional neural networks

DL

deep learning

EAs

evolutionary algorithms

EDM

Euclidean distance matrix

EGFR

epidermal growth factor receptor erbB1

FMO

fragment molecular orbital

GAs

genetic algorithms

GANs

generative adversarial networks

GCN

graph convolution network

GNN

graph neural network

LB

ligand-based

MCM

Monte Carlo Metropolis

MCMC

Markov Chain Monte Carlo

MCTS

Monte Carlo tree search

MD

molecular dynamics

ML

machine learning

MM/GBSA

molecular mechanics with generalized Born and surface area solvation

RMSD

root-mean-square deviation

RNNs

recurrent neural networks

SA

synthetic accessibility

SBVS

structure-based virtual screening

SMILES

simplified molecular input line entry specification

VAE

variational autoencoders

VEGFR2

vascular endothelial growth factor receptor 2

β₂AR

β₂-adrenergic receptor

Data Availability Statement

The data and software availability information can be found in the table at the end of the text.

Author Contributions

The manuscript was written through the contributions of all authors. All authors have given approval to the final version of the manuscript.

The authors declare no competing financial interest.