FOCUS on NOD2: Advancing IBD Drug Discovery with a User-Informed Machine Learning Framework

Abstract

In this study, we introduce the Framework for Optimized Customizable User-Informed Synthesis (FOCUS), a generative machine learning model tailored for drug discovery. FOCUS integrates domain expertise and uses Proximal Policy Optimization (PPO) to guide Monte Carlo Tree Search (MCTS) to efficiently explore chemical space. It generates SMILES representations of potential drug candidates, optimizing for druggability and binding efficacy to NOD2, PEP, and MCT1 receptors. The model is highly interpretive, allowing for user-feedback and expert-driven adjustments based on detailed cycle reports. Employing tools like SHAP and LIME, FOCUS provides a transparent analysis of decision-making processes, emphasizing features such as docking scores and interaction fingerprints. Comparative studies with Muramyl Dipeptide (MDP) demonstrate improved interaction profiles. FOCUS merges advanced machine learning with expert insight, accelerating the drug discovery pipeline.

Nucleotide oligomerization domain 2 (NOD2) is a canonical pattern-recognition receptor highly expressed in ileal Paneth cells and intestinal stem cells¹ that plays an important role in maintaining host equilibrium.²⁻⁴ It serves as a general sensor for Gram-negative and Gram-positive bacteria by recognizing muramyl dipeptide (MDP), a common constituent of peptidoglycan.^5,6 When activated by a ligand, NOD2 triggers proinflammatory innate immune responses and regulate intestinal microbiota. Mutations in NOD2 correlate with a range of medical conditions like Blau syndrome, asthma, and arthiritis.⁷⁻⁹ More prominently, disruptions in NOD2 signaling are linked to Crohn’s Disease.^10,11 Given the pivotal role of NOD2 in mediating innate immune responses, exploring novel ligands with drug-like quality capable of modulating this crucial interaction is of paramount importance. However, the potential ligands would need to traverse barriers to interact with NOD2 within the cells given its location. Therefore, the novel ligands would need to exhibit optimal bioactivity toward NOD2 receptor and effectively interact with transporters facilitating a more comprehensive approach.

As poor bioavailability remains a leading cause of failure in drug development,¹² solute carrier transporters are becoming attractive targets for therapeutic intervention.¹³ PEPT1 (peptide transporter 1) belonging to this family is present in the mucosa of the small intestine accounts for over 50% of all known clincally relevant drug transporters^14,15 and is known to translocate MDP and other bacterial products into the cytosol of colonic epthelial cells during inflammation.^16,17 Although the structures of transporters are now well elucidated, the underlying mechanisms governing substrate recognition and subsequent conformational alterations continue to be a focal point of ongoing research. MCT1 is a bidirectional transporter of monocarboxylates such as butyrate, lactate, and pyruvate and is widely expressed throughout the intestine. Due to its enhanced transport capability, it is gaining traction in development of novel drug-delivery strategies.¹⁸ These particular residues play a pivotal role in the binding and transportation of peptides, underlining their importance in the MCT1 function.^19,20 Studies have revealed that interaction with lysine residue and hydrobphobic bonds play an important role in transport function of this transporter.²¹⁻²³ Even though, MDP (or any known NOD2 ligand) is not transported by MCT1, we included this transporter in our case study to introduce a level of complexity and to illustrate the feasibility of designing molecules that can theoretically align with transporters possessing distinct substrate preferences. In this context, our focus with PEPT1 and MCT1 is guided by the computational imperative to design ligands that not only theoretically interact with NOD2, an intracellular receptor, but also have the potential to be transported into cells. This dual consideration in computational models is pivotal as it aims to suggest ligands that are predicted to be biologically active at the site of NOD2 and possess attributes that could enable them to overcome cellular barriers, a crucial aspect in the theoretical framework of successful drug development. However, searching for alternative therapeutic molecules with desirable properties is not a trivial task.

The conventional approach of drug discovery is both time-consuming and costly, with approximately 80–90% of candidates experiencing failure at the clinical evaluation stage.²⁴ The lead optimization (LO) phase is particularly resource-intensive and the most expensive phase of this process.²⁵ This multiparameter optimization problem involves pinpointing compounds that balance multiple and often conflicting objectives like drug-like properties and adequate bioactivity.²⁶ The advent of deep learning generative models for the de novo design of molecular structures presents a promising approach to accelerate drug discovery, spurring a wide spectrum of research as covered in these reviews.²⁷⁻³⁰ Despite the proven efficiency and potential of deep learning models, the lack of transparency in these “black-box” models is a major limitation. The complexity of these models often makes them inscrutable to users, obscuring their decision-making algorithms. This obscurity limits scientists’ abilities to use these models for forward-looking suggestions, typically relegating their use to evaluating already created compounds and narrowing their role in decision-making. This opacity deepens the divide between chemists and AI platforms. Models should be interpretable to not only enhance the trust in the generated solutions but also to incorporate feedback from domain experts. Additionally constrained design of these generative models curtails their potential, particularly as they struggle with tinkering design—a prevalent method for molecular exploration—due to the myriad of constraints affecting potential modifications.

Therefore, to generate novel ligands for NOD2 that can also be transported into the cells as NOD2 targets, we have developed a framework called FOCUS. In this study, we introduce the Framework for Optimized Customizable User-Informed Synthesis (FOCUS). The name “FOCUS” was carefully chosen to encapsulate the essence of our approach: it is a Framework designed to optimize the synthesis process, making it highly Customizable to adapt to specific research needs. It is User-Informed, meaning it actively incorporates feedback and insights from domain experts throughout the drug discovery process. The Synthesis aspect highlights our model’s ability to generate novel compounds effectively. By utilizing Proximal Policy Optimization (PPO) to guide Monte Carlo Tree Search (MCTS), the FOCUS method offers a strategic and efficient exploration of the chemical space, emphasizing its designed purpose to “focus” efforts on the most promising areas of research. It integrates interpretability, detailed data analysis at each iterative stage, and has he capability for user-driven tinkering, addressing challenges that are inherent in navigating the expansive chemical space. One key additional aspect of FOCUS is interaction fingerprints. Since computational molecular docking, often a part of generative models, can be highly inaccurate necessitating the visual inspection of docking poses,³¹ there is a need for mechanisms that extend beyond merely predicting binding affinity. Interaction fingerprints (IFPs) capture the three-dimensional interactions within molecular complexes through vector or binary representations, capturing more binding-relevant information compared to atomic fingerprints.³² In the field of drug discovery, precise characterization of these receptor/ligand interactions enables medicinal chemists to evaluate the potential of a ligand for further investigation and pharmaceutical development. Research groups have undertaken different approaches to decompose interactions from relying on pharmacophoric features³² to incorporating atomic information.³³ A notable challenge is the absence of a clear consensus on the empirical thresholds for distance and angles used to describe bonds.³⁴ Nonetheless, several deep learning models have emerged that utilize IFPs to improve binding prediction and retain binding mode information.^35,36 FOCUS uses BINANA2³⁷ to decompose the interactions between our generated molecules and the three receptors as it has been successfully employed in other studies to produce more accurate results.³⁸⁻⁴⁰

In the context of drug discovery, especially concerning molecules that modulate biological targets like NOD2, it is imperative to address multiobjective optimization challenges that encompass not only biological activity but also drug-like and ADMET properties. FOCUS tackles this by providing a transparent, user-interactive platform that leverages both computational predictions and domain knowledge from medicinal chemists. The detailed data analysis provided at each stage of the generative process permits users to guide the model through the vast chemical space effectively, focusing on regions that may provide molecules that balance the often conflicting objectives of high biological activity, favorable drug-like properties, and satisfactory ADMET profiles. Moreover, considering the challenge of simultaneously optimizing binding affinity toward multiple receptors, FOCUS ensures a systematic and informed exploration, providing a robust methodology that is capable of handling the intricate problem of generating molecules that can adeptly navigate the multifaceted landscape of multitarget drug discovery.

The FOCUS method depicted in Figure 1 follows a structured workflow. It starts with a template SMILES, upon which an LSTM network applies Monte Carlo techniques to generate analogues. This process involves replacing partial sequences in the template SMILES, with optimization geared toward druggability properties as defined by a reward function. This partial insertion MCTS method was derived from a published work.⁴¹ After generating a set of SMILES, the top six with the highest reward scores are selected for further analysis. These are assessed for their docking scores, and their interaction fingerprints with NOD2 and the two transporter receptors are meticulously documented. This data is then input into the Proximal Policy Optimization (PPO) algorithm. The PPO algorithm is designed to refine the docking score optimization against the three receptors and to enhance the accuracy of interaction fingerprint predictions. Its output influences the exploration constant in the Monte Carlo Tree Search (MCTS) and guides the adjustment of reward weights within MCTS. Thus, PPO is instrumental in directing the MCTS toward generating SMILES that are likely to possess desirable drug-like properties and demonstrate effective binding to the three receptors of interest.

Figure 1.

MCTS navigates the chemical space guided by multiobjective optimization for druggability-related properties. PPO determines the weights and exploration constants, iteratively refining the molecular generation process. The model operates in a loop, providing comprehensive reports, including interpretability results and various analysis plots after each episode, allowing for expert feedback and user-tinkering.

A brief overview of the techniques and their integration is provided here, with detailed technical specifics and configurations available in the Supporting Information.

LSTM Training

Our approach begins with training an LSTM model, well-suited for sequential data such as molecular structures, to capture the essential patterns from the ZINC database.⁴² This model lays the groundwork for the subsequent generative phase, leveraging the inherent capabilities of LSTM networks for managing sequential data with long-term dependencies.⁴³⁻⁴⁵

Generating New Sequences Using MCTS

The trained LSTM model informs the Monte Carlo Tree Search (MCTS) algorithm, which iteratively generates and evaluates new molecular sequences. MCTS generates analogues by substituting portions of a template molecule, effectively exploring the vast chemical space by balancing exploratory moves with exploitation. The process is driven by a composite reward function, designed to optimize several molecular properties critical for drug effectiveness. These properties include the Quantitative Estimate of Drug-likeness (QED), the partition coefficient logP, and molecular weight, which are calculated using RDKit functions, synthetic Accessibility (SA) calculated using a separate method,⁴⁶ and caco-2 permeability scores are predicted using a model from another study.⁴⁷ The technical specifics of MCTS, including update rules and configuration, along with a detailed explanation of the reward function’s components, are provided in the Supporting Information. The validity of generated SMILES was ascertained using various checks from the RDKit chemoinformatics library, considering only the representations that passed these validation steps in the reward calculation for the MCTS algorithm. Figure S1 illustrates example of top 6 analogues generated by a single MCTS process.

After MCTS generates the valid SMILES, they are arranged in descending order starting with the SMILE that has the highest inner reward at the top. The top 6 molecules based on the inner reward are selected to run binding analysis on i.e. docking score with NOD2, PEPT1 and MCT1 are calculated using Autodock smina and their interaction fingerprint is generated using BINANA. Within the reward function, FOCUS carefully balances the use of docking scores and BINANA interaction fingerprints to select molecules. Docking scores initially filter out molecules with less promising binding profiles. However, to overcome the limitations of docking score predictions, the model also utilizes BINANA interaction fingerprints. These fingerprints provide a detailed analysis of molecular interactions, with a special emphasis on hydrophobic interactions, which are crucial for binding to NOD2.⁴⁸ Therefore, the reward function is designed to give priority to molecules that not only meet the criteria based on docking scores but also show a high degree of hydrophobic interactions. The SMILE with the highest outer reward is used as the new template molecule for the next iteration.

Chemical Space Navigation Using PPO

Our methodology includes exploring chemical space using a combination of Monte Carlo Tree Search (MCTS) and Proximal Policy Optimization (PPO). PPO is a reinforcement learning algorithm which has emerged as a reliable and efficient method to train deep reinforcement learning models.^49,50 It helps adjust the balance between exploration and exploitation, which is crucial when navigating through complex molecular structures.

In our setup, PPO guides how MCTS makes decisions by dictating the weights of the MCTS reward function and determining the exploration constant. MCTS builds a tree to explore different molecular options, selecting paths based on a scoring system. PPO refines this scoring system by updating the rules MCTS uses to choose and evaluate possible molecular structures. This means that PPO adjusts how MCTS explores new areas and how it uses information from previous successful outcomes.

By using PPO to enhance the decision-making in MCTS, we ensure that our search is not just wide-ranging but also focused on finding molecules with the best properties. PPO updates these decisions continuously as it receives new data from MCTS results, making the whole process more effective.

Model Validation and Decision-Making Interpretation

As detailed in the Supporting Information, the generated molecules are rigorously validated using the RDKit cheminformatics library to ensure their structural and chemical validity. The framework is designed with a strong focus on interpretability and user control. The interpretability of our deep learning model’s decision-making process is enhanced through the use of SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) analyses. These tools provide critical insights into how various features influence the model’s predictions, helping to understand and improve the molecular generation process.^51,52 SHAP values offer a deep dive into the contribution of each feature to the model’s predictions, enabling users to understand how different chemical properties or structural elements of the SMILES influence the model’s output. LIME, on the other hand, provides localized interpretations, illustrating how changes in the input features affect the predictions on an individual prediction basis.

Both the MDP molecule and its synthesized analogs have shown to be highly pyrogenic, leading to serious side effects upon use, and thus, are ruled out for direct clinical application.⁵³ This prohibits the application of MDP and emphasizes the need to design ligands that replicate the bioactivity of MDP but with an improved safety profile. An important inclusion in our method is the Caco-2 score⁵⁴ which reflects the Caco-2 permeability, the gold standard to predict drug intestinal absorption.^55,56 MDP suffers from a poor drug quality profile with undesirable QED, Caco-2, and LogP scores. Therefore, a critical task is to ensure that the generated molecules possess properties falling within an favorable range.

We applied FOCUS to MDP (template molecule), showcasing a systematic improvement in molecular properties as the model iterates through episodes. Figure 2 demonstrates a notable shift from the first to the last episode. Initially, the generated molecules exhibited lower range of QED with highest being 0.2. They also had negative values of LogP, which is not favorable as the aim is to have more hydrophobic molecules. A highly negative distribution of Caco-2 score and high molecular weight were also unfavorable for permeability. However, as the episodes progressed, a transition toward more favorable property ranges is observed. By the final episode, the distribution reflects higher QED (Figure 2A), improved logP (Figure 2B), better Caco-2 scores (Figure 2C), and lower MW (Figure 2D), indicating the model’s enhanced capability in generating molecules with desirable attributes.

Figure 2.

Progression of molecular properties over the course of the generative process, comparing the initial (first episode) and final (last episode) generated molecules. Two distinct KDE curves depict the distributions of property values at these two stages, with the orange line indicating the boundary marking the desirable ranges.

Unlike other properties, the distribution plot of SA score (Figure 2F) remains within an acceptable range throughout all episodes, affirming the robustness of the model in maintaining good synthetic accessibility from the outset. The template molecule (MDP) possesses a favorable SA score, which is consistently mirrored in the generated molecules across episodes. The Tanimoto similarity plot (Figure 2E) reveals an increase in molecular diversity over the episodes. Initially, the generated sequences were closely aligned with the template MDP, reflecting higher Tanimoto similarity values. As episodes advanced, a broader spread in Tanimoto similarity values is observed, indicating the generation of more diverse molecular structures. This trend is consistent with the MCTS’s partial substitution mechanism, which initially generates molecules similar to the template but diversifies the generated structures as iterations progress. Figure S2 demonstrates the progression of ADMET properties across different optimization episodes.

Having established that FOCUS was instrumental in generating molecules with a more appealing drug profile, we advanced to scrutinize the impact of the exploration constant on our outcomes, particularly focusing on binding-related properties which are not optimized by the MCTS. We categorized the properties into three domains: drug quality, binding score, and interaction fingerprints. Figure 3 illustrates a comparison of outcomes between scenarios of local exploitation 0.3 ≤ c and global exploration 0.7 ≥ c. Notably, global exploration was associated with more favorable ranges in the drug quality profile as opposed to local exploitation. Although the mean docking score for NOD2 remained similar in both scenarios, global exploration exhibited a narrower distribution range. On the other hand, local exploitation was favorable for a better docking score with PEPT1, while for MCT, the distribution was akin in both scenarios except for several outliers in local exploitation, indicating a docking score in the positive, which suggests energetically unfavorable binding. In terms of interaction signature, global exploration yielded a higher average count of hydrophobic bonds, with NOD2 averaging at 29.5 as opposed to 26.4 in local exploitation. It has been established that hydrophobic interactions are directly proportional to NOD2 activation.⁵⁷ Therefore, generating molecules with high hydrophobic interactions is desirable. Conversely, local exploitation showed better results for hydrogen bonds with NOD2. Even in interactions with PEPT1 and MCT, where the average count of bonds remained consistent for both scenarios, local exploitation featured more data points with higher bond counts. These findings highlight the necessity of both local exploitation and global exploration in the quest for suitable molecules, underscoring the challenging nature of navigating the chemical space. The heatmap in Figure S3 illustrates the Spearman correlation coefficients between various molecular properties and interaction scores. Notably, certain correlations can provide significant insights into the chemical properties that influence molecular interactions. By leveraging these correlations, molecular properties can be strategically prioritized in the MCTS reward function. Specifically, properties that are demonstrated to have strong positive correlations with key performance indicators, such as binding affinity, are assigned higher rewards. Conversely, properties that exhibit strong negative correlations with desired outcomes are penalized within the reward structure.

Figure 3.

Distributions of (A) drug-quality profiles, (B) binding scores, and (C) interaction signatures, differentiating between molecules generated through local exploitation and global exploration. Additionally, reference lines indicate the corresponding values for Muramyl Dipeptide (MDP), serving as a comparative benchmark to evaluate the relative performance. The green arrow on the y-axis represents the target value range, indicating the direction toward favorable molecular property profiles.

Following each episode, a comprehensive data report is generated, providing a detailed view on various molecular aspects. Figure 4A, depicted as a radar chart, offers a comparative analysis of the number of molecules meeting the set criteria versus those that do not, across different categories. It is observed that for certain properties such as QED, Caco-2, SA, logP, and hydrophobic bonds with NOD2 and MCT, a significant majority of molecules satisfy the criteria. Conversely, for Molecular Weight (MW), π Stacking, salt bridges, cation-π, and π–π bonds, a lesser proportion of molecules meet the established benchmarks. The subsequent subplot, Figure 4B, delineates the safety and reasonability profile of the generated molecules. It is illustrated that a predominant share of the molecules adheres to RDKit’s valency and charge criteria and are devoid of PAINS (Pan-Assay Interference Structures).

Figure 4.

(A) Chart conveying the proportion of generated molecules adhering to, or deviating from, established criteria, across various molecular properties and descriptors. (B) A validity, reasonability, and PAINS assessment plot illustrating the proportion of generated molecules adhering to chemical validity, reasonability, and absence of PAINS. (C) A comparative analysis of recurrent substructures found in molecules categorized as ‘good’ and ‘bad’ drugs based on predefined pharmacological profiles. (D) Identified pharmacophoric features across molecules categorized into four groups based on their bioactivity and safety profiles.

Next, we segregated the molecules into two distinct profiles based on their drug-likeness quality: a good drug-like profile and a bad drug-like profile. The categorization was predicated on QED values (good = QED > 0.5, bad = QED < 0.5) and logP values (good = logP in positive, bad = logP in negative). This bifurcation was orchestrated to delve deeper into the substructural distinctions between molecules with superior drug-like qualities and those lacking them.

The composition of functional groups within molecules significantly impacts their drug-like properties. In Figure 4C, we observed higher aromatic rings and amine groups but lower amide and carbonyl groups. Aromatic rings are frequently utilized in drug design due to their propensity to engage in varied interactions with target receptors, attributed to their distinctive chemical features like electronegativity, volume, and lipophilicity, which influence biological interactions.⁵⁸ On the other hand, amide groups are known for their hydrogen bonding capabilities and can partake in less polar interactions, such as stacking interactions with aromatic heterocycles, which are essential in facilitating ligand–receptor interactions at binding sites.⁵⁹ The carbonyl groups in amides, due to their electron-withdrawing nature, render amides less basic compared to amines, which might influence the drug’s interaction with its targets. Amine groups play a fundamental role in drug discovery, especially through the formation of amide bonds via coupling with carboxylic acids, a commonly used reaction in drug discovery.⁶⁰ Additionally, prodrugs of amines have been explored to address challenges like low aqueous solubility, poor membrane permeability, and chemical instability, thereby potentially enhancing the pharmacokinetic properties of amine-containing drugs.⁶¹ For instance, the elevated presence of aromatic rings could bolster interaction with targets, while the increase in amine groups might facilitate beneficial chemical reactions for drug discovery. Conversely, the reduction in amide and carbonyl groups might be an endeavor to modulate the basicity and electron-withdrawing characteristics of the generated molecules, which could, in turn, affect their interaction profiles and overall drug-like properties. The modulation in the proportions of these functional groups underscores a potential shift toward molecules with better stability, binding affinity, and pharmacokinetic profiles, aligning with the drug-like properties the model aims to optimize for.

In Figure 4D, we employed a meticulous pharmacophore analysis to discern the underlying molecular features correlating with these attributes. We divided the molecules into categories of good drug-likeness (0.5 ≥ QED and positive values of LogP) and bad drug-likeness (QED < 0.5 and negative values of LogP), as well as strong binding profiles as described in the table and poor binding profiles ((docking score with NOD2 < −5.8, docking score with PEP < −7.0, docking score with MCT1 < −4.5). We observed distinct molecular trends among these categories, particularly concerning the quantity of hydrogen bond acceptors, the presence of a positive ionizable core, and the distribution of lumped hydrophobes. We noticed the bad drug-like molecules exhibited a higher quantity of HBAs which aligns with the established understanding that a higher number of HBAs can contribute to increased polarity and decreased lipophilicity. A decrease in lumped hydrophobes in molecules with a poor binding profile could signify a reduction in hydrophobic interactions, which are fundamental for stable binding with target proteins, especially within hydrophobic pockets or regions. The lack of sufficient hydrophobic interactions could lead to a weaker binding affinity, explaining the poor binding profile observed in these molecules. Positive ionizable features indicate regions of the molecule capable of bearing a positive charge, essential for electrostatic interactions with negatively charged groups on targets.⁶² An increase in the positive ionizable core in molecules associated with a poor binding profile could be indicative of an enhanced positive charge. This might not be favorable for interactions with specific targets, especially if these targets have regions of positive charge or neutral charge distribution. The electrostatic repulsion or lack of favorable electrostatic interactions could adversely affect the binding affinity of these molecules to the targets, thereby resulting in a poor binding profile.

This analysis of substructures and pharmacophores serves dual purposes. First, it validates the trends observed against established domain knowledge, ensuring the logic and validity of the generation process remain intact. Second, it augments our understanding, as seen in this instance, by offering insights into the binding pockets of target receptors. Such knowledge is invaluable as it empowers users to fine-tune parameters in subsequent episodes, thus refining the molecule generation process to better align with desired objectives. Ten molecules generated by FOCUS and ten MDP analogues were randomly selected for analysis. Their interactions with the NOD2 receptor were examined and compared to highlight the similarities in Figure S4.

In the field of molecular generation and optimization, interpretability and the ability for user-driven adjustments are crucial. Our model is designed to provide clear insights into the molecular generation process, while also allowing users to modify its parameters. Through enhanced interpretability and user-centric tinkering capabilities, our model becomes a flexible tool, capable of adapting to various molecular design objectives.

SHAP (SHapley Additive exPlanations) is a model-agnostic method that uses cooperative game theory to fairly allocate contributions of feature values to the prediction for a particular instance, offering both local interpretability and the ability to compute global feature importance.⁵¹ The SHAP summary plot in Figure 5A provides a quantifiable measure of feature importance in our generative model, which predicts the weights of the ’Reward’ input and the exploration constant through a combination of MCTS and PPO.

Figure 5.

(A) SHAP summary plot illustrating the mean influence of various features, such as docking scores, interaction fingerprints, and internal rewards, on the model’s predictions, categorizing them into five distinct output-based groups the exploration constant, and weights associated with QED, penalized logP, MW, and Caco-2 score. (B) LIME subplots exemplifying local interpretability for two distinct data points (8 and 367), each presenting the contribution of features toward model prediction. The x-axis indicates the impact on the model’s prediction, while the y-axis enumerates specific features and conditions.

Figure 5A shows the SHAP summary plot which provides quantified metrics of feature impact on multiple outputs of the generative model, including the exploration constant and four associated weights related to specific molecular properties. The ’Reward’ feature demonstrates a notable influence across all model outputs, with a SHAP value peaking at 0.07, while docking score with PEPT1 exhibits the least influence, potentially due to its variable nature or possible redundancy. Classes 0 through 4 represent distinct model outputs: the exploration constant, and weights associated with QED, penalized logP, MW, and Caco-2 score, respectively. Each class displays similar magnitudes in their stacked SHAP values, indicating a consistent influence of features across outputs. Interestingly, the docking score with MCT1 has greater importance than that with NOD2 and PEP, despite known binding affinities of the template molecule. This might indicate the model’s capability to explore new chemical spaces and generalize beyond training data, potentially identifying new molecular interactions, particularly given the significant impact of hydrophobic bonds, a key element in molecular binding and docking scores. This analysis of feature contributions and interactions provides detailed insights into the predictive mechanisms, enhancing understanding and interpretability of the multiobjective optimization in the SMILES generation process.

While the SHAP values offer a global perspective, highlighting the overall importance of each feature across all predictions, it is imperative to delve into localized, instance-specific interpretations to uncover the nuanced influences of features on individual predictions. Leveraging LIME (Local Interpretable Model-agnostic Explanations), we explore how altering specific features within localized regions affects the predictive outcome, offering a detailed, microscopic view that complements the macroscopic insights derived from SHAP.⁵²

In our LIME analysis, we aim to provide a detailed, microscopic view of the model’s decision-making process on an individual prediction basis. This approach complements the macroscopic insights derived from the SHAP analysis. To achieve this, we randomly selected data points from our data set, which comprises [total number of data points], for a more unbiased and representative analysis. Specifically, data points 8 and 367 were chosen to demonstrate the variability and adaptability of the model in different scenarios, showcasing the instance-specific influence of various features.

The random selection of these data points ensures that the analysis is not biased toward any particular type of prediction or outcome. Data point 8, for example, illustrates the significant impact of the docking score with MCT1 (Feature 3), especially when this score exceeds 0, indicating no or unfavorable binding. This negatively influences the model’s predictive outcome. In contrast, data point 367 highlights different influences and thresholds of other features, such as “Reward” (Feature 0) and interaction with a glutamate residue in PEPT1 (Feature 12).

The variability observed between these randomly selected instances sheds light on the model’s sensitivity to different molecular scenarios, adjusting its predictions by weighting features differently based on the unique characteristics of each input.

Activity landscapes provide a critical view of the relationship between molecular similarity and bioactivity, revealing instances of activity cliffs where small structural changes lead to significant activity differences.^63,64 These cliffs present a challenge for structure–activity relationship (SAR) modeling but also offer opportunities for model refinement through user intervention.⁶⁵

In the presented activity landscape plot (Figure 6) including receptors MCT1, PEP, and NOD2, the relationship between Tanimoto similarity and activity difference uncovers a spectrum of insightful observations and anomalies, notably in the form of activity cliffs. The general trend, where a higher Tanimoto similarity corresponds to a diminished activity difference, adheres to conventional expectations. However, intriguing exceptions are observed, such as instances in NOD2 and PEPT1 where molecules, despite a Tanimoto similarity of 0.6, exhibit an activity difference of 2, indicating a significant alteration in activity despite relatively minor structural differences. Particularly stark is the landscape for MCT1, which unveils a varied terrain of activity differences extending from 5 to 10, even at high Tanimoto similarities of up to 0.85, presenting a challenging scenario for predictive modeling and highlighting a pronounced activity cliff. Through a meticulous analysis and understanding of activity cliffs, users can judiciously modify models, integrating penalties or biases against specific molecular modifications or sequences that induce cliffs, aiming to enhance predictive performance of the model.

Figure 6.

Juxtaposition of Tanimoto similarity against activity difference, revealing general trends and exceptional cases (activity cliffs) across the three receptors. Notably, MCT1 exhibits a diverse landscape, with high Tanimoto similarities corresponding to activity differences extending up to 10. For NOD2 and PEPT1, notable cliffs are observed at a Tanimoto similarity of 0.6, indicating a doubled activity difference. An example of the structure behind the activty cliff for each receptor is displayed.

Deciphering the interaction signatures and binding residues of generated molecules provides a pivotal platform to rationalize and compare the binding affinities and specificities of novel candidates with known ligands, such as Muramyl Dipeptide (MDP). Furthermore, a meticulous exploration of these molecular interactions and the residued involved, especially within the binding pocket, not only enables a the mechanistic understanding of ligand–receptor engagements but also facilitates identification of suitable candidates for further investigation and eventual testing.

In Figure 7A, a frequency plot illustrates the ten most prevalent residues in NOD2, PEPT1, and MCT1, demonstrating a variety of interactions such as hydrophobic contacts, π–π and π-stacking, hydrogen bonding, and salt bridges with the generated molecules. A study demonstrated that strong binding alone is not enough to activate NOD2.⁶⁶ The exact mechanism of interactions leading to successful NOD2 signaling is not very well-understood and remains an active area of research.⁶⁷ Hydrophobic contacts are associated with increased NOD2 activation.⁵⁷ TRP887 emerges as the most common residue engaging in hydrophobic contacts, π–π and π-stacking interactions, contributing to a stabilized ligand binding. Agonists and inhibitors exhibit very similar sets of residues, but the nature of interactions might differ. Hydrogen bonds with SER913 and TRP911 were observed in both MDP and a NOD2 agonist.⁶⁸ However, MDP forms an additional H bond with ARG857 (interacts with both peptide and carbohydrate part), whereas the inhibitor forms both a hydrogen and a cation-π bond with the same residue.⁶⁸ The molecules generated by FOCUS were a mixture of both agonists and inhibitors, as it was observed that there were molecules which formed cation-π bonds with ARG857. Other notable residues include PHE831, TRP911, and ARG857, with ARG857 being significantly involved in hydrogen bonding, enhancing molecular recognition and interaction with ligands. Additionally, LYS969 forms salt bridges with glutamate residues at positions 943 and 939, hinting at potential electrostatic interactions crucial for structural integrity and function. Notably, most of the hydrogen bonding occurs with residues ARG857, ARG803, and SER913. Figure S5 shows the frequency plot of MDP analogues with NOD2 and PEPT1 receptors.

Figure 7.

(A) Frequency plot delineating the ten most common residues engaged in interactions with generated molecules, segmented by interaction type (hydrogen bonds, hydrophobic contacts, etc.) for NOD2, PEP, and MCT1 receptors. (B) Detailed visualization of binding interactions of randomized four candidate molecules compared to MDP, highlighting crucial interactions in NOD2. Each subplot presents a distinct binding signature. The highlighted portions in the 2D structure correspond to the structures common with MDP. (C) The drug-related properties of the candidate molecules were compared with MDP marking an improvement.

In the depicted figure, the top three residues identified are TRP294, GLU595, and VAL626, playing potentially significant roles in the transporter function. Specifically, hydrogen bonds are observed with residues GLU595, ASN171, TRP622, and TYR167, facilitating crucial interactions for ligand recognition and binding. Notably, GLU595 engages predominantly in salt bridges, a type of ionic interaction vital for maintaining structural integrity and function of the transporter. This aligns with the observation that specific interactions, such as the hydrogen bond with asparagine (N329) on TM8 and engagement with glutamate (E595) on TM10, have been observed in drugs known to be transported by PEPT1.^69,70 Observations indicate that PHE362, TRP369, PHE365, and ALA339 are the most common residues interacting with MCT1. These interactions underline the importance of specific amino acid residues like Arginine and Lysine for ionic interactions, Aspartate and Glutamate for hydrogen bonding and ionic interactions, and aromatic residues like Tryptophan for hydrophobic interactions, in orchestrating the precise functionality of transporter proteins.⁷¹

Although over a hundred potential candidates were generated, Figure 7B selectively illustrates the binding interactions of four random candidates in comparison with MDP, particularly focusing on NOD2, a receptor of key interest in our study. All 4 candidates exhibited a multifaceted binding profile. Arginine, with its positively charged guanidinium group, often forms stable hydrogen bonds that can contribute significantly to ligand binding affinity and specificity in receptor–ligand interactions. Candidates 1, 2, and 3 all form pi-pi bonds with PHE831 due to their aromatic nature. These three candidates also form cation-pi interactions, contributing to the energetics with tryptophan residues. TRP887, an amino acid with an indole side chain, engages in two notable t-stacking interactions in Candidate 4, likely involving its aromatic ring and a corresponding aromatic ring or conjugated system within the ligand. These t-stacking interactions with tryptophan are pivotal, potentially stabilizing the ligand and ensuring its optimal orientation within the binding pocket, thereby influencing both the affinity and specificity of ligand binding. In Candidate 3, GLU939 forms a salt bridge, with its deprotonated carboxylate group typically interacting with a positively charged moiety on the ligand. This salt bridge is crucial, offering stabilization of charged entities within the predominantly hydrophobic environment of the binding pocket, and significantly contributing to binding affinity. Moreover, it plays a strategic role in determining the ligand’s orientation and positioning within the pocket, enhancing the precision and stability of the ligand–receptor interaction. Moreover, a comparative analysis of the generated molecules and MDP, illustrated through a bubble plot, revealed enhancements in three pivotal properties: QED, logP, and Caco-2 scores, indicating that the generated molecules not only exhibit favorable binding signatures but also possess improved druggable properties.

Addressing the challenge of identifying NOD2 modulators, where known ligands are predominantly variations of Muramyl Dipeptide (MDP), is critical in expanding the scope of potential binding interactions and molecular scaffolds in NOD2-targeted drug discovery. This limitation confines the exploration of the vast chemical space and may hinder the discovery of novel agonists with improved selectivity, efficacy, or reduced adverse effects compared to MDP-based compounds. Diversification beyond MDP derivatives is essential for uncovering new mechanisms of NOD2 activation, which could lead to the development of more effective therapeutic strategies for diseases associated with NOD2 dysfunction, such as Crohn’s disease. Recent research efforts exemplify the active pursuit of structurally diverse NOD2 agonists.⁴⁸ Such studies not only broaden the chemical diversity of NOD2 modulators but also contribute valuable insights into the receptor’s pharmacology, paving the way for the development of innovative therapeutic options.

Generative modeling represents a significant approach in the discovery of new bioactive molecules, providing a mechanism to explore new possibilities in computational drug discovery. However, exploring the complex chemical space effectively requires both advanced machine learning models and the expertise of medicinal chemists. Medicinal chemists provide crucial insights into molecular interactions, synthetic feasibility, and structure–activity relationship (SAR) trends, all of which are challenging to fully encode into computational models.^72,73

The FOCUS framework was developed to prioritize the interpretability of model decisions and facilitate user interaction. Unlike most models that rely solely on data-driven predictions, FOCUS enables the investigation and adjustment of results after each iteration, emphasizing a human-driven approach over a purely model-driven one. Furthermore, we based our molecule design process on the crystal structure of rabbit NOD2 in its ADP-bound state, serving as an essential reference. This strategy allowed us to pinpoint important ligand-residue interactions through BINANA analysis, improving the specificity of our predictions. Recognizing the limitations inherent to docking scores, especially their potential inaccuracies in predicting binding affinities, FOCUS complements this approach by integrating BINANA interaction fingerprints. This inclusion significantly enhances our assessment framework, enabling a detailed evaluation of crucial molecular interactions. Particular emphasis is placed on hydrophobic interactions, due to their strong correlation with improved NOD2 receptor engagement. This dual methodology combines initial docking score filtering with detailed BINANA interaction analysis, forming the core of our strategy for evaluating molecular candidates comprehensively. This approach highlights the innovation of our model and provides a thorough analysis. This is crucial, given the imprecision of docking scores as standalone indicators of binding affinity and the challenge posed by the limited diversity among known NOD2 ligands, primarily variations of MDP. By leveraging both the crystal structure insights and the computational analysis provided by BINANA, we mitigate the risk of over-reliance on docking scores, thereby enhancing the robustness and accuracy of our predictions.

A key strength of FOCUS lies in its ability to handle the complexity of multiobjective optimization, with our study involving three receptors and respective interaction fingerprints, alongside various druggability properties including an ADMET property. In the context of drug discovery, particularly for molecules modulating biological targets like NOD2, it is imperative to navigate the multiobjective optimization challenges that encompass not merely biological activity but also drug-like and ADMET properties. It is pivotal to consider not only the principal receptor of interest but also transporter receptors and other biological entities involved in the upstream and downstream processes. This necessitates a comprehensive approach toward the drug discovery pathway, ensuring every step—from molecule generation to its entire biological journey—is tracked and optimized. This demonstrated the capability of FOCUS to navigate complex, multiobjective problems. While FOCUS effectively designed candidates for potential bioactivity and yielded expected results in substructure and pharmacophore analysis, the use of SMILES notations imposes limitations due to the potential introduction of inaccuracies in molecular representations.^74,75 Alternatives like DeepSMILES⁷⁶ or SELFIES,⁷⁷ and 3D molecular information,⁷⁸ may augment future predictive and generative capacities. FOCUS effectively designed candidates with potential NOD2, PEPT1 and MCT1 activity, and the results in substructure and pharmacophore analysis were in line with expectations while also providing new insights regarding binding residues and interactions. While MCTS is proficient at exploring large and complex search spaces, it can sometimes struggle with ensuring that its explorations are strategically directed toward profitable regions of the space.⁷⁹ By employing PPO to modulate the exploration constant and dictate the weights of the reward function, the search is directed toward more promising areas of the chemical space without being overly exploitative. This alleviates the issue of the search becoming stagnated in local optima or dispersing too widely to be effective. Additionally, PPO’s adaptive characteristics pave the way for a search strategy that is not only dynamic but also finely attuned to the multifaceted optimization landscape that is a staple in molecular generation tasks. It ensures the search maintains a healthy balance between leveraging known, productive regions of the space and diving into novel, unexplored territories.

Given the vastness and high-dimensionality of the chemical space, FOCUS adopts a strategy that emphasizes optimizing the search process rather than attempting to understand the entire landscape. This perspective is especially relevant in the current context where the representation of molecules is subject to ongoing debate, and methodologies are perpetually evolving. In summary, while generative modeling in drug discovery continues to advance, FOCUS demonstrates the importance of combining computational capabilities with human expertise. Navigating the complex space of chemical compounds requires robust models and the ability to explore, adjust, and refine strategies, thus improving the exploration process rather than solely focusing on the end goal.

In this research, the FOCUS framework has generated a portfolio of molecules with potential efficacy against NOD2. These molecules that meet the criteria specifically Synthetic Accessibility Score suggest that they are viable candidates for synthesis and further investigation. The crucial next step is the empirical validation of these selected molecules. This validation process can be effectively conducted through a high-throughput (HT) NOD2 assay, which is an efficient method for screening and identifying compounds with potent NOD2 binding capabilities. The advantage of a high-throughput approach lies in its capacity to test a broad array of compounds within a relatively short time frame, thereby expediting the discovery and validation process. Following the empirical testing, detailed biochemical assays to measure the direct interaction between the molecules and NOD2 can provide more specific insights into binding affinities and kinetics. For molecules showing promising interactions, cell-based assays can further assess their ability to modulate NOD2 activity within a biological context, examining effects on downstream signaling pathways and cellular responses. Moreover, toxicity screening is essential to identify any potential cytotoxic effects early in the discovery process.

Animal models representing NOD2-mediated conditions could also be employed to evaluate the in vivo efficacy and safety profiles of leading candidates. The insights gained will be invaluable for refining the FOCUS framework. This iterative process, where computational predictions are honed based on laboratory results, is essential for the progressive enhancement of drug discovery methodologies. By effectively narrowing down the pool of potential candidates, FOCUS significantly reduce the number of compounds requiring empirical testing. This not only saves valuable time and resources but also accelerates the identification of promising candidates. However, it is important to acknowledge that several iterative cycles of computational analysis and empirical validation are essential to precisely hone in on the most effective molecules. The aim of these computational tools like FOCUS is not to replace laboratory experiments but to complement them. By providing a more focused starting point for empirical validation, these tools can significantly expedite the experimental phase, enhancing the overall efficiency and efficacy of the drug discovery process.^80,81 Iterative cycles of computational analysis and empirical validation are essential to precisely zero in on the most effective molecules, demonstrating the complementary roles of computational and experimental methodologies in advancing pharmaceutical research.

Data Availability Statement

The source code can be found on GitHub: https://github.com/LMSE/FOCUS_v1.

Supporting Information Available

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsmedchemlett.4c00148.

• Detailed computational methodologies employed in the study, including the setup and operational details of the Monte Carlo Tree Search (MCTS) algorithm, complete with its mathematical formulations and reward function specifications; explanations of the Proximal Policy Optimization (PPO) algorithm’s integration with MCTS, specifying the algorithmic settings and hyperparameters used; validation procedures using the RDKit cheminformatics library for the generated SMILES strings, including a series of validation checks such as molecular structure conversion, 3D coordinate generation, and molecular sanitization steps; Figures S1–S5 illustrating the MCTS process, the relationships between molecular properties via heatmaps, and comparisons of generated molecular structures; and tables outlining the conditions and rewards for calculating the outer reward of the MCTS algorithm to provide a comprehensive view of the computational framework (PDF)

The authors declare no competing financial interest.

Special Issue

Published as part of ACS Medicinal Chemistry Lettersvirtual special issue “Exploring the Use of AI/ML Technologies in Medicinal Chemistry and Drug Discovery”.