Tagir Akhmetshin
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Tagir Akhmetshin has a total of 47 co-authors. Here are the top co-authors by number of shared publications:
- Alexandre Varnek
Laboratory of Chemoinformatics, UMR 7140 CNRS, University of Strasbourg, 4, rue B. Pascal, Strasbourg, France, 67000
12 shared publications - Dragos Horvath
University of Strasbourg
5 shared publications - Timur I. Madzhidov
Chemistry Solutions, Life Science Solutions, Elsevier, Nielsen House, John Smith Drive, Oxford OX4 2WB, United Kingdom
5 shared publications - Dmitry Zankov
Aix-Marseille Universite
4 shared publications - Gilles Marcou
Laboratory of Chemoinformatics UMR 7140 University of Strasbourg Strasbourg France
3 shared publications Timur Madzhidov
Elsevier Nielsen House, John Smith Drive
3 shared publications
Latest Publications
- 21 August 2026
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 66 (17) , 10919-10934
Leveraging the Condensed Graph of Reaction for Clustering Retrosynthetic Pathways
Modern retrosynthetic tools can propose hundreds of alternative pathways for a single target, making it challenging to effectively explore and navigate the resulting route space. We present a CGR-based framework for the analysis and clustering of synthetic routes that integrates both target-centered and all-species-centered perspectives. Entire reaction pathways are encoded as single-molecule graphs (RouteCGR) or reduced representations retaining only target atoms (SB-CGR), enabling automatic identification of strategic bond patterns (SBPs). These representations can be transformed into Morgan fingerprints for a quantitative route similarity assessment. We propose a two-level clustering strategy in which routes are first grouped by shared SBPs, ensuring high interpretability based on key retrosynthetic disconnections, and then further differentiated using RouteCGR similarity to capture variations in starting materials and auxiliary transformations. The method demonstrates near-linear scalability and computational efficiency for large data sets. Application to synthetic routes for apatinib generated by multiple planning tools reveals tool-dependent diversity in strategic disconnections and highlights the benefit of combining tools to expand route space. The framework also supports cross-target analysis, enabling the identification of reusable route families that share common strategic disconnections and building blocks across related molecules. Overall, the SBP-based approach provides an interpretable and scalable solution for automated synthesis route analysis and informed decision-making. The proposed approach is implemented in SynPlanner retrosynthesis planning software and is available as a standalone module at https://github.com/Laboratoire-de-Chemoinformatique/SynPlanner.
- 21 August 2026
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 66 (17) , 10919-10934
Leveraging the Condensed Graph of Reaction for Clustering Retrosynthetic Pathways
Modern retrosynthetic tools can propose hundreds of alternative pathways for a single target, making it challenging to effectively explore and navigate the resulting route space. We present a CGR-based framework for the analysis and clustering of synthetic routes that integrates both target-centered and all-species-centered perspectives. Entire reaction pathways are encoded as single-molecule graphs (RouteCGR) or reduced representations retaining only target atoms (SB-CGR), enabling automatic identification of strategic bond patterns (SBPs). These representations can be transformed into Morgan fingerprints for a quantitative route similarity assessment. We propose a two-level clustering strategy in which routes are first grouped by shared SBPs, ensuring high interpretability based on key retrosynthetic disconnections, and then further differentiated using RouteCGR similarity to capture variations in starting materials and auxiliary transformations. The method demonstrates near-linear scalability and computational efficiency for large data sets. Application to synthetic routes for apatinib generated by multiple planning tools reveals tool-dependent diversity in strategic disconnections and highlights the benefit of combining tools to expand route space. The framework also supports cross-target analysis, enabling the identification of reusable route families that share common strategic disconnections and building blocks across related molecules. Overall, the SBP-based approach provides an interpretable and scalable solution for automated synthesis route analysis and informed decision-making. The proposed approach is implemented in SynPlanner retrosynthesis planning software and is available as a standalone module at https://github.com/Laboratoire-de-Chemoinformatique/SynPlanner.
- 8 April 2026
- Published by Agence Bibliographique de l'Enseignement Supérieur
Graph-based neural networks for generation of synthetically accessible molecular structures
Réseaux neuronaux à base de graphes pour la génération de structures moléculaires synthétiquement accessibles: Cette thèse est dédiée au développement de génératifs par réseaux de neurones artificiels de graphes et d'un outil de planification rétrosynthétique amélioré par des méthodes d’apprentissage automatique profond. Les modèles génératifs s'appuient sur le concept d'autoencodeur, très populaire dans les tâches de conception moléculaire de novo et d'analyse QSAR inverse. L'architecture proposée, HyFactor, basée sur le graphe étiqueté par le nombre d'hydrogènes par atome, se révèle performante et utile pour générer des analogues moléculaires. La seconde architecture proposée, VQGAE, donne des résultats comparables à ceux d'HyFactor dans la tâche de reconstruction, mais la représentation latente se montre plus performante pour la recherche par similarité et pour le QSAR. Ces performances sont illustrées par la génération de ligands très actifs pour les récepteurs A2A. La faisabilité synthétique des structures générées a été vérifiée à l'aide d'un nouvel outil de rétrosynthèse, GSLRetro. Cet outil est conçu sur le concept d'auto-apprentissage qui améliore la stabilité des solutions grâce aux précédentes solutions proposées. Le protocole de curation des données de réaction ainsi que de nouvelles techniques de validation croisée pour les modèles QSPR basés sur les réactions sont également présentés. En somme, la combinaison de ces outils constitue une étape importante vers la conception automatisée de molécules.
- 8 April 2026
- Published by Agence Bibliographique de l'Enseignement Supérieur
Graph-based neural networks for generation of synthetically accessible molecular structures
Réseaux neuronaux à base de graphes pour la génération de structures moléculaires synthétiquement accessibles: Cette thèse est dédiée au développement de génératifs par réseaux de neurones artificiels de graphes et d'un outil de planification rétrosynthétique amélioré par des méthodes d’apprentissage automatique profond. Les modèles génératifs s'appuient sur le concept d'autoencodeur, très populaire dans les tâches de conception moléculaire de novo et d'analyse QSAR inverse. L'architecture proposée, HyFactor, basée sur le graphe étiqueté par le nombre d'hydrogènes par atome, se révèle performante et utile pour générer des analogues moléculaires. La seconde architecture proposée, VQGAE, donne des résultats comparables à ceux d'HyFactor dans la tâche de reconstruction, mais la représentation latente se montre plus performante pour la recherche par similarité et pour le QSAR. Ces performances sont illustrées par la génération de ligands très actifs pour les récepteurs A2A. La faisabilité synthétique des structures générées a été vérifiée à l'aide d'un nouvel outil de rétrosynthèse, GSLRetro. Cet outil est conçu sur le concept d'auto-apprentissage qui améliore la stabilité des solutions grâce aux précédentes solutions proposées. Le protocole de curation des données de réaction ainsi que de nouvelles techniques de validation croisée pour les modèles QSPR basés sur les réactions sont également présentés. En somme, la combinaison de ces outils constitue une étape importante vers la conception automatisée de molécules.
- 23 December 2025
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 66 (1) , 744-755
Design of Highly Potent Antibiofilm, Antimicrobial Peptides Using Explainable Artificial Intelligence
Antimicrobial peptides have emerged as a potential alternative to traditional small-molecule antimicrobials. They possess broad-spectrum efficacy and increasingly confront the challenges of bacterial resistance, especially the adaptive resistance of biofilms. However, advanced rational peptide design methods are still required to ensure optimal property profiles of such peptides, while limiting the cost of their synthesis and screening. Here, we present a computational pipeline for the rational de novo design of antimicrobial and antibiofilm peptides based on an explainable artificial intelligence (XAI) framework. The developed framework combines a Wasserstein Autoencoder (WAE) and a nonlinear dimensionality reduction method─generative topographic mapping (GTM). The WAE was used to learn the latent representation of the peptide space, while the GTM guided the generation of novel AMPs through an illustrative depiction of the latent space in the form of 2D maps. The generated peptides were subjected to screening by machine learning models, resulting in the final hit list based on their predicted activity. The efficacy of the peptides generated with the developed pipeline was experimentally verified by synthesis and testing for activity against methicillin-resistant Staphylococcus aureus (MRSA), achieving a 100% hit rate in targeting biofilms. Notably, the most potent antibiofilm peptide developed in this study demonstrated almost one order of magnitude improvement in IC50 value compared with the potent antibiofilm peptide reference "1018", used as a positive control. The developed pipeline is readily extendable for the optimization of additional peptide properties, including cytotoxicity, tendency to aggregate, and proteolytic stability, underscoring its potential utility for rational design of the peptide-based therapeutics.
- 23 December 2025
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 66 (1) , 744-755
Design of Highly Potent Antibiofilm, Antimicrobial Peptides Using Explainable Artificial Intelligence
Antimicrobial peptides have emerged as a potential alternative to traditional small-molecule antimicrobials. They possess broad-spectrum efficacy and increasingly confront the challenges of bacterial resistance, especially the adaptive resistance of biofilms. However, advanced rational peptide design methods are still required to ensure optimal property profiles of such peptides, while limiting the cost of their synthesis and screening. Here, we present a computational pipeline for the rational de novo design of antimicrobial and antibiofilm peptides based on an explainable artificial intelligence (XAI) framework. The developed framework combines a Wasserstein Autoencoder (WAE) and a nonlinear dimensionality reduction method─generative topographic mapping (GTM). The WAE was used to learn the latent representation of the peptide space, while the GTM guided the generation of novel AMPs through an illustrative depiction of the latent space in the form of 2D maps. The generated peptides were subjected to screening by machine learning models, resulting in the final hit list based on their predicted activity. The efficacy of the peptides generated with the developed pipeline was experimentally verified by synthesis and testing for activity against methicillin-resistant Staphylococcus aureus (MRSA), achieving a 100% hit rate in targeting biofilms. Notably, the most potent antibiofilm peptide developed in this study demonstrated almost one order of magnitude improvement in IC50 value compared with the potent antibiofilm peptide reference "1018", used as a positive control. The developed pipeline is readily extendable for the optimization of additional peptide properties, including cytotoxicity, tendency to aggregate, and proteolytic stability, underscoring its potential utility for rational design of the peptide-based therapeutics.
- Open Access
- 1 September 2025
- Published by American Chemical Society (ACS) in ChemRxiv
Leveraging Condensed Graphs of Multi-step Reaction Routes for Clustering Retrosynthetic Pathways
Retrosynthetic planning often produces a multitude of potential synthetic routes, creating challenges for chemists in selecting the most viable options. Existing clustering approaches used in tools like ASKCOS and AiZynthFinder rely on abstract or computationally intensive metrics that often lack chemical interpretability. In this work, we propose a novel, chemically intuitive clustering methodology implemented in the SynPlanner retrosynthetic tool, based on the concept of Strategic Bond Patterns (SBPs). By extending the Condensed Graph of Reaction (CGR) formalism to multi-step retrosynthetic pathways, we introduce the RouteCGR and its simplified variant, SB-CGR, which encodes the key disconnections leading to a target molecule. This two-level clustering strategy - first by SBP, then by generalized pseudo-reactions - groups routes by their underlying synthetic logic while tolerating variations in the reactions’ order and application of non-strategic steps (protection/deprotection). Comparison against clustering algorithms from ASKCOS and AiZynthFinder demonstrates that the SBP-based method produces fewer, more interpretable clusters and aligns more closely with chemists’ reasoning and intuition. SynPlanner’s SBP approach uncovers a broader diversity of unique synthetic strategies and demonstrates faster, more meaningful clustering. This framework facilitates efficient exploration of synthetic options, enhances interpretability, and offers a practical solution for navigating the combinatorial complexity of retrosynthetic planning.
- 1 September 2025
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessLeveraging Condensed Graphs of Multi-step Reaction Routes for Clustering Retrosynthetic Pathways
Retrosynthetic planning often produces a multitude of potential synthetic routes, creating challenges for chemists in selecting the most viable options. Existing clustering approaches used in tools like ASKCOS and AiZynthFinder rely on abstract or computationally intensive metrics that often lack chemical interpretability. In this work, we propose a novel, chemically intuitive clustering methodology implemented in the SynPlanner retrosynthetic tool, based on the concept of Strategic Bond Patterns (SBPs). By extending the Condensed Graph of Reaction (CGR) formalism to multi-step retrosynthetic pathways, we introduce the RouteCGR and its simplified variant, SB-CGR, which encodes the key disconnections leading to a target molecule. This two-level clustering strategy - first by SBP, then by generalized pseudo-reactions - groups routes by their underlying synthetic logic while tolerating variations in the reactions’ order and application of non-strategic steps (protection/deprotection). Comparison against clustering algorithms from ASKCOS and AiZynthFinder demonstrates that the SBP-based method produces fewer, more interpretable clusters and aligns more closely with chemists’ reasoning and intuition. SynPlanner’s SBP approach uncovers a broader diversity of unique synthetic strategies and demonstrates faster, more meaningful clustering. This framework facilitates efficient exploration of synthetic options, enhances interpretability, and offers a practical solution for navigating the combinatorial complexity of retrosynthetic planning.
- 31 December 2024
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 65 (1) , 15-21
SynPlanner: An End-to-End Tool for Synthesis Planning
SynPlanner is an end-to-end tool for designing customized retrosynthetic planners from reaction data. It includes a reaction data curation pipeline (reaction atom-to-atom mapping, reaction standardization, and filtration), reaction rule extraction, retrosynthetic model training, and retrosynthetic planning. The tool is designed to be as flexible as possible, supporting the customization of each step of the pipeline to address different needs in the development of customized retrosynthetic planning solutions. The retrosynthetic planning in SynPlanner is performed by Monte Carlo Tree Search (MCTS) guided by graph neural networks for node expansion (retrosynthetic rule predictions) and evaluation (precursor synthesizability prediction). The solution can be accessed by a simple graphical user interface and a command line interface and is accompanied by a collection of tutorials. SynPlanner is available on GitHub at https://github.com/Laboratoire-de-Chemoinformatique/SynPlanner.
- 31 December 2024
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 65 (1) , 15-21
SynPlanner: An End-to-End Tool for Synthesis Planning
SynPlanner is an end-to-end tool for designing customized retrosynthetic planners from reaction data. It includes a reaction data curation pipeline (reaction atom-to-atom mapping, reaction standardization, and filtration), reaction rule extraction, retrosynthetic model training, and retrosynthetic planning. The tool is designed to be as flexible as possible, supporting the customization of each step of the pipeline to address different needs in the development of customized retrosynthetic planning solutions. The retrosynthetic planning in SynPlanner is performed by Monte Carlo Tree Search (MCTS) guided by graph neural networks for node expansion (retrosynthetic rule predictions) and evaluation (precursor synthesizability prediction). The solution can be accessed by a simple graphical user interface and a command line interface and is accompanied by a collection of tutorials. SynPlanner is available on GitHub at https://github.com/Laboratoire-de-Chemoinformatique/SynPlanner.
- Open Access
- 5 December 2024
- Published by Wiley in Molecular Informatics
- Vol. 44 (1) , e202400265
From High Dimensions to Human Insight: Exploring Dimensionality Reduction for Chemical Space Visualization
Dimensionality reduction is an important exploratory data analysis method that allows high‐dimensional data to be represented in a human‐interpretable lower‐dimensional space. It is extensively applied in the analysis of chemical libraries, where chemical structure data ‐ represented as high‐dimensional feature vectors‐are transformed into 2D or 3D chemical space maps. In this paper, commonly used dimensionality reduction techniques ‐ Principal Component Analysis (PCA), t‐Distributed Stochastic Neighbor Embedding (t‐SNE), Uniform Manifold Approximation and Projection (UMAP), and Generative Topographic Mapping (GTM) ‐ are evaluated in terms of neighborhood preservation and visualization capability of sets of small molecules from the ChEMBL database.
- 5 December 2024
- Published by Wiley in Molecular Informatics
- Vol. 44 (1) , e202400265
Open AccessFrom High Dimensions to Human Insight: Exploring Dimensionality Reduction for Chemical Space Visualization
Dimensionality reduction is an important exploratory data analysis method that allows high‐dimensional data to be represented in a human‐interpretable lower‐dimensional space. It is extensively applied in the analysis of chemical libraries, where chemical structure data ‐ represented as high‐dimensional feature vectors‐are transformed into 2D or 3D chemical space maps. In this paper, commonly used dimensionality reduction techniques ‐ Principal Component Analysis (PCA), t‐Distributed Stochastic Neighbor Embedding (t‐SNE), Uniform Manifold Approximation and Projection (UMAP), and Generative Topographic Mapping (GTM) ‐ are evaluated in terms of neighborhood preservation and visualization capability of sets of small molecules from the ChEMBL database.
- Open Access
- 16 August 2024
- Published by American Chemical Society (ACS) in ChemRxiv
From High Dimensions to Human Comprehension: Exploring Dimensionality Reduction for Chemical Space Visualization
Dimensionality reduction is an important exploratory data analysis method that allows high-dimensional data to be represented in a human-interpretable lower-dimensional space. It is extensively applied in the analysis of chemical libraries, where chemical structure data — represented as high-dimensional feature vectors—are transformed into 2D or 3D chemical space maps. In this paper, commonly used dimensionality reduction techniques — Principal Component Analysis (PCA), t-Distributed Stochastic Neighbor Embedding (t-SNE), Uniform Manifold Approximation and Projection (UMAP), and Generative Topographic Mapping (GTM) — are evaluated for exploration of subsets of small molecule organic compounds from ChEMBL database. The performance of these methods is examined in terms of neighborhood preservation and visualization capabilities, and the strengths and limitations are discussed.
- 16 August 2024
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessFrom High Dimensions to Human Comprehension: Exploring Dimensionality Reduction for Chemical Space Visualization
Dimensionality reduction is an important exploratory data analysis method that allows high-dimensional data to be represented in a human-interpretable lower-dimensional space. It is extensively applied in the analysis of chemical libraries, where chemical structure data — represented as high-dimensional feature vectors—are transformed into 2D or 3D chemical space maps. In this paper, commonly used dimensionality reduction techniques — Principal Component Analysis (PCA), t-Distributed Stochastic Neighbor Embedding (t-SNE), Uniform Manifold Approximation and Projection (UMAP), and Generative Topographic Mapping (GTM) — are evaluated for exploration of subsets of small molecule organic compounds from ChEMBL database. The performance of these methods is examined in terms of neighborhood preservation and visualization capabilities, and the strengths and limitations are discussed.
- Open Access
- 6 August 2024
- Published by American Chemical Society (ACS) in ChemRxiv
SynPlanner: an end-to-end tool for synthesis planning
SynPlanner is an open-source tool for retrosynthetic planning, designed to increase flexibility in training and developing customized retrosynthetic planning solutions from raw chemical data. It integrates Monte Carlo Tree Search (MCTS) with graph neural networks to evaluate applicable reaction rules (policy network) and the synthesizability of intermediate products (value network). SynPlanner can be used directly with pre-trained policy/value networks or fine-tuned on custom data through an automated end-to-end training pipeline. Additionally, SynPlanner enables the training of custom value functions on discovered synthesis routes to improve predictive performance. The tool includes original modules for atom-to-atom mapping, reaction curation, standardization, and extraction of reaction rules, ensuring the reproducibility of the training pipeline from initial data to trained retro-synthetic models. SynPlanner is available on GitHub at https://github.com/Laboratoire-de-Chemoinformatique/SynPlanner
- 6 August 2024
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessSynPlanner: an end-to-end tool for synthesis planning
SynPlanner is an open-source tool for retrosynthetic planning, designed to increase flexibility in training and developing customized retrosynthetic planning solutions from raw chemical data. It integrates Monte Carlo Tree Search (MCTS) with graph neural networks to evaluate applicable reaction rules (policy network) and the synthesizability of intermediate products (value network). SynPlanner can be used directly with pre-trained policy/value networks or fine-tuned on custom data through an automated end-to-end training pipeline. Additionally, SynPlanner enables the training of custom value functions on discovered synthesis routes to improve predictive performance. The tool includes original modules for atom-to-atom mapping, reaction curation, standardization, and extraction of reaction rules, ensuring the reproducibility of the training pipeline from initial data to trained retro-synthetic models. SynPlanner is available on GitHub at https://github.com/Laboratoire-de-Chemoinformatique/SynPlanner
- Open Access
- 10 July 2024
- Published by American Chemical Society (ACS) in ChemRxiv
CoLiNN: A Tool for Fast Chemical Space Visualization of DNA-Encoded Libraries Without Enumeration
Visualization of the combinatorial library chemical space provides a comprehensive overview of available compound classes, their diversity, and physicochemical property distribution - key factors in drug discovery. Typically, this visualization requires time- and resource-consuming compound enumeration, standardization, descriptor calculation, and dimensionality reduction. In this study, we present the Combinatorial Library Neural Network (CoLiNN) designed to predict the projection of compounds on a 2D chemical space map using only their building blocks and reaction information, thus eliminating the need for compound enumeration. Trained on 2.5K virtual DNA-Encoded Libraries (DELs), CoLiNN demonstrated high predictive performance, accurately predicting the compound position on Generative Topographic Maps (GTMs). GTMs predicted by CoLiNN were found very similar to the maps built for enumerated structures. In the library comparison task, we compared the GTMs of DELs and the ChEMBL database. The similarity-based DELs / ChEMBL rankings obtained with “true” and CoLiNN predicted GTMs were consistent. Therefore, CoLiNN has the potential to become the go-to tool for combinatorial compound library design – it can explore the library design space more efficiently by skipping the compound enumeration.
- 10 July 2024
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessCoLiNN: A Tool for Fast Chemical Space Visualization of DNA-Encoded Libraries Without Enumeration
Visualization of the combinatorial library chemical space provides a comprehensive overview of available compound classes, their diversity, and physicochemical property distribution - key factors in drug discovery. Typically, this visualization requires time- and resource-consuming compound enumeration, standardization, descriptor calculation, and dimensionality reduction. In this study, we present the Combinatorial Library Neural Network (CoLiNN) designed to predict the projection of compounds on a 2D chemical space map using only their building blocks and reaction information, thus eliminating the need for compound enumeration. Trained on 2.5K virtual DNA-Encoded Libraries (DELs), CoLiNN demonstrated high predictive performance, accurately predicting the compound position on Generative Topographic Maps (GTMs). GTMs predicted by CoLiNN were found very similar to the maps built for enumerated structures. In the library comparison task, we compared the GTMs of DELs and the ChEMBL database. The similarity-based DELs / ChEMBL rankings obtained with “true” and CoLiNN predicted GTMs were consistent. Therefore, CoLiNN has the potential to become the go-to tool for combinatorial compound library design – it can explore the library design space more efficiently by skipping the compound enumeration.
- 25 July 2022
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 62 (15) , 3524-3534
HyFactor: A Novel Open-Source, Graph-Based Architecture for Chemical Structure Generation
Graph-based architectures are becoming increasingly popular as a tool for structure generation. Here, we introduce novel open-source architecture HyFactor in which, similar to the InChI linear notation, the number of hydrogens attached to the heavy atoms was considered instead of the bond types. HyFactor was benchmarked on the ZINC 250K, MOSES, and ChEMBL data sets against conventional graph-based architecture ReFactor, representing our implementation of the reported DEFactor architecture in the literature. On average, HyFactor models contain some 20% less fitting parameters than those of ReFactor. The two architectures display similar validity, uniqueness, and reconstruction rates. Compared to the training set compounds, HyFactor generates more similar structures than ReFactor. This could be explained by the fact that the latter generates many open-chain analogues of cyclic structures in the training set. It has been demonstrated that the reconstruction error of heavy molecules can be significantly reduced using the data augmentation technique. The codes of HyFactor and ReFactor as well as all models obtained in this study are publicly available from our GitHub repository: https://github.com/Laboratoire-de-Chemoinformatique/HyFactor.
- 25 July 2022
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 62 (15) , 3524-3534
HyFactor: A Novel Open-Source, Graph-Based Architecture for Chemical Structure Generation
Graph-based architectures are becoming increasingly popular as a tool for structure generation. Here, we introduce novel open-source architecture HyFactor in which, similar to the InChI linear notation, the number of hydrogens attached to the heavy atoms was considered instead of the bond types. HyFactor was benchmarked on the ZINC 250K, MOSES, and ChEMBL data sets against conventional graph-based architecture ReFactor, representing our implementation of the reported DEFactor architecture in the literature. On average, HyFactor models contain some 20% less fitting parameters than those of ReFactor. The two architectures display similar validity, uniqueness, and reconstruction rates. Compared to the training set compounds, HyFactor generates more similar structures than ReFactor. This could be explained by the fact that the latter generates many open-chain analogues of cyclic structures in the training set. It has been demonstrated that the reconstruction error of heavy molecules can be significantly reduced using the data augmentation technique. The codes of HyFactor and ReFactor as well as all models obtained in this study are publicly available from our GitHub repository: https://github.com/Laboratoire-de-Chemoinformatique/HyFactor.