Arijit Roy
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Arijit Roy has a total of 59 co-authors. Here are the top co-authors by number of shared publications:
Sowmya Ramaswamy Krishnan
11 shared publications- Sowmya Ramaswamy Krishnan
Department of Biotechnology, Bhupat and Jyoti Mehta School of Biosciences, Indian Institute of Technology Madras , Chennai 600036
10 shared publications - Gopalakrishnan Bulusu
Hyderabad (Life Sciences Division) TCS Innovation Labs Tata Consultancy Services Limited Hyderabad India
9 shared publications - Navneet Bung
TATA Consultancy Services (India)
9 shared publications Rajgopal Srinivasan
TCS Research (Life Sciences division), Tata Consultancy Services, Hyderabad, India
8 shared publicationsNavneet Bung
TCS Research (Life Sciences Division), Tata Consultancy Services Limited, Hyderabad, India
7 shared publications
Latest Publications
- Open Access
- 17 November 2026
- Published by Springer Nature in Journal of Cheminformatics
SKiP: a Structure-oriented Kinetics Prediction suite for enzyme-substrate complexes
- 17 November 2026
- Published by Springer Nature in Journal of Cheminformatics
Open AccessSKiP: a Structure-oriented Kinetics Prediction suite for enzyme-substrate complexes
- Open Access
TRACEDD: A Tool-grounded Reasoning and Agentic Coordination for Explainable Drug Design
Drug discovery depends on coordinated decisions across target validation, structure analysis, molecular design, developability assessment and synthetic feasibility, but current computational methods often operate as disconnected tools. Here, we introduce TRACEDD (Tool-grounded Reasoning and Agentic Coordination for Explainable Drug Design), a framework that makes three primary contributions: (1) It establishes a ‘tool-first’ multi agentic architecture where LLMs orchestrate validated computational tools rather than replace them, ensuring scientific rigor. (2) It implements a multi-agent system that mirrors expert discovery teams, enabling transparent and traceable decision-making through a Reason-Act-Observe loop. (3) It demonstrates an end-to-end workflow, from target validation to synthesis planning, that adaptively handles real-world data variability, such as the absence of experimental structures. The framework decomposes discovery into specialized agents for target validation, druggability assessment, molecular generation, lead optimization, ADMET evaluation, literature evidence integration and retrosynthesis, all operating through a Reason–Act–Observe workflow. Using JAK2 as a representative case, we show that the system can retrieve experimental protein structures, invoke AlphaFold when structures are unavailable, identify druggable pockets and perform de novo molecular generation. Known JAK2 inhibitors are used to define design hypotheses and guide reinforcement learning-based molecular generation, with docking scores/predicted pIC50 and other physicochemical/ADMET properties serving as reward and prioritization signals. The framework demonstrates a tool-first, reasoning-driven approach in which each major decision is linked to explicit tool invocation, intermediate evidence. By combining agentic orchestration with domain-specific computational tools, the system supports transparent, adaptable and human-verifiable molecular design workflows, providing a foundation for more reliable AI-assisted drug discovery.
Open AccessTRACEDD: A Tool-grounded Reasoning and Agentic Coordination for Explainable Drug Design
Drug discovery depends on coordinated decisions across target validation, structure analysis, molecular design, developability assessment and synthetic feasibility, but current computational methods often operate as disconnected tools. Here, we introduce TRACEDD (Tool-grounded Reasoning and Agentic Coordination for Explainable Drug Design), a framework that makes three primary contributions: (1) It establishes a ‘tool-first’ multi agentic architecture where LLMs orchestrate validated computational tools rather than replace them, ensuring scientific rigor. (2) It implements a multi-agent system that mirrors expert discovery teams, enabling transparent and traceable decision-making through a Reason-Act-Observe loop. (3) It demonstrates an end-to-end workflow, from target validation to synthesis planning, that adaptively handles real-world data variability, such as the absence of experimental structures. The framework decomposes discovery into specialized agents for target validation, druggability assessment, molecular generation, lead optimization, ADMET evaluation, literature evidence integration and retrosynthesis, all operating through a Reason–Act–Observe workflow. Using JAK2 as a representative case, we show that the system can retrieve experimental protein structures, invoke AlphaFold when structures are unavailable, identify druggable pockets and perform de novo molecular generation. Known JAK2 inhibitors are used to define design hypotheses and guide reinforcement learning-based molecular generation, with docking scores/predicted pIC50 and other physicochemical/ADMET properties serving as reward and prioritization signals. The framework demonstrates a tool-first, reasoning-driven approach in which each major decision is linked to explicit tool invocation, intermediate evidence. By combining agentic orchestration with domain-specific computational tools, the system supports transparent, adaptable and human-verifiable molecular design workflows, providing a foundation for more reliable AI-assisted drug discovery.
- Open Access
Mechanistically informed adaptive dosing for cancer immunotherapy using AI‑guided decision making
Optimizing dose and schedule remains a central challenge in oncology drug development, particularly for immunotherapies where fixed dosing regimens often fail to account for patient specific heterogeneity in tumor-immune dynamics. Here, we present a hybrid quantitative systems pharmacology-reinforcement learning-Monte Carlo Tree Search (QSP-RL-MCTS) framework for personalized immunotherapy dosing that formulates dose selection as a sequential decision‑making problem. The approach integrates a mechanistic QSP model of prostate cancer immunotherapy, transcriptomics informed virtual patient populations and data driven AI system comprising reinforcement learning and Monte Carlo tree search. Reinforcement learning is used to learn adaptive generalized dosing policies that optimize treatment outcomes across the population, while Monte Carlo Tree Search provides forward-looking evaluation of RL predicted dosing trajectories to refine patient‑specific decisions. On benchmarking against fixed dosing regimens of ipilimumab, the remission rate of the proposed model (95.2%) was comparable to the highest fixed dosing regimen of 10 mg/kg per dose while the median total dose (72 mg/kg) of the proposed model designed regimen was comparable to the lowest fixed dosing regimen of 3 mg/kg per dose. The model is generalizable across different dosing protocols and can be extended to predict optimal dose under different therapeutic scenarios. Analysis of the learned dosing trajectories enables stratification of patients into distinct response groups and identifies drug activity rate as the dominant determinant of long‑term treatment outcome. These results demonstrate how mechanistically guided artificial intelligence can transform population‑level dose optimization into patient‑specific, biologically interpretable treatment strategies for precision immuno‑oncology.
Open AccessMechanistically informed adaptive dosing for cancer immunotherapy using AI‑guided decision making
Optimizing dose and schedule remains a central challenge in oncology drug development, particularly for immunotherapies where fixed dosing regimens often fail to account for patient specific heterogeneity in tumor-immune dynamics. Here, we present a hybrid quantitative systems pharmacology-reinforcement learning-Monte Carlo Tree Search (QSP-RL-MCTS) framework for personalized immunotherapy dosing that formulates dose selection as a sequential decision‑making problem. The approach integrates a mechanistic QSP model of prostate cancer immunotherapy, transcriptomics informed virtual patient populations and data driven AI system comprising reinforcement learning and Monte Carlo tree search. Reinforcement learning is used to learn adaptive generalized dosing policies that optimize treatment outcomes across the population, while Monte Carlo Tree Search provides forward-looking evaluation of RL predicted dosing trajectories to refine patient‑specific decisions. On benchmarking against fixed dosing regimens of ipilimumab, the remission rate of the proposed model (95.2%) was comparable to the highest fixed dosing regimen of 10 mg/kg per dose while the median total dose (72 mg/kg) of the proposed model designed regimen was comparable to the lowest fixed dosing regimen of 3 mg/kg per dose. The model is generalizable across different dosing protocols and can be extended to predict optimal dose under different therapeutic scenarios. Analysis of the learned dosing trajectories enables stratification of patients into distinct response groups and identifies drug activity rate as the dominant determinant of long‑term treatment outcome. These results demonstrate how mechanistically guided artificial intelligence can transform population‑level dose optimization into patient‑specific, biologically interpretable treatment strategies for precision immuno‑oncology.
- Open Access
EnzFinder: a sustainable alternative to chemical synthesis
Enzymes have emerged as an important alternative to traditional catalysts in chemical industries over the past few decades owing to their sustainable nature. The application of enzymes in chemical synthesis relies on their ability to catalyse promiscuous reactions. Promiscuous activity of enzymes is an abundant phenomenon in nature; approximately 37% of Escherichia coli K12 enzymes show promiscuous activity. This highlights the vast expanse of chemical reactions that can be made biochemically feasible through selection of the correct candidate enzymes. Here, we present EnzFinder, a promiscuous enzyme prediction tool that filters candidate enzymes based on similarity in chemical transformation patterns and subsequently ranks them using substrate–product similarity, enabling enzyme prioritization up to the fourth level of EC classification without requiring sequence information. On a blind benchmarking set of 2,309 biochemical reactions, the method achieves substantially higher prediction accuracy than existing rule-based and deep-learning approaches, with improvements exceeding 20% at the sub-subclass level and significantly higher coverage at the fourth level. Application to industrially relevant reactions demonstrates EnzFinder’s ability to identify alternative enzymes with higher substrate similarity and improved kinetic potential. Furthermore, integration of EnzFinder with in silico retrosynthesis tools enables effective prioritization of enzymatic steps within hybrid chemical–biological pathways. Together, these results establish EnzFinder as a practical and interpretable tool for accelerating enzyme discovery and promoting greener, enzyme-driven synthesis routes.
Open AccessEnzFinder: a sustainable alternative to chemical synthesis
Enzymes have emerged as an important alternative to traditional catalysts in chemical industries over the past few decades owing to their sustainable nature. The application of enzymes in chemical synthesis relies on their ability to catalyse promiscuous reactions. Promiscuous activity of enzymes is an abundant phenomenon in nature; approximately 37% of Escherichia coli K12 enzymes show promiscuous activity. This highlights the vast expanse of chemical reactions that can be made biochemically feasible through selection of the correct candidate enzymes. Here, we present EnzFinder, a promiscuous enzyme prediction tool that filters candidate enzymes based on similarity in chemical transformation patterns and subsequently ranks them using substrate–product similarity, enabling enzyme prioritization up to the fourth level of EC classification without requiring sequence information. On a blind benchmarking set of 2,309 biochemical reactions, the method achieves substantially higher prediction accuracy than existing rule-based and deep-learning approaches, with improvements exceeding 20% at the sub-subclass level and significantly higher coverage at the fourth level. Application to industrially relevant reactions demonstrates EnzFinder’s ability to identify alternative enzymes with higher substrate similarity and improved kinetic potential. Furthermore, integration of EnzFinder with in silico retrosynthesis tools enables effective prioritization of enzymatic steps within hybrid chemical–biological pathways. Together, these results establish EnzFinder as a practical and interpretable tool for accelerating enzyme discovery and promoting greener, enzyme-driven synthesis routes.
- Open Access
- 1 January 2026
- Published by Royal Society of Chemistry (RSC) in Digital Discovery
- Vol. 5 (8) , 3328-3339
EnzFinder: a sustainable alternative to chemical synthesis
Enzymes have emerged as an important alternative to traditional catalysts in chemical industries over the past few decades owing to their sustainable nature. The application of enzymes in chemical synthesis...
- 1 January 2026
- Published by Royal Society of Chemistry (RSC) in Digital Discovery
- Vol. 5 (8) , 3328-3339
Open AccessEnzFinder: a sustainable alternative to chemical synthesis
Enzymes have emerged as an important alternative to traditional catalysts in chemical industries over the past few decades owing to their sustainable nature. The application of enzymes in chemical synthesis...
- 29 September 2025
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
MolDecor: Leveraging Transformers to Decorate Bioactive Molecules for Property Optimization
Lead optimization is a critical stage in drug discovery, where promising molecules (lead molecules) are further optimized. It involves the refinement of the chemical structure of the lead molecule to improve its pharmacological properties and drug-like characteristics for development into potential therapies. In this study, we developed a pipeline that includes (a) the creation of a property-specific fragment (decorator) library, (b) learning fragment-scaffold relationship using a BERT-based transformer model, and (c) decorating a given scaffold using fragments from the generated fragment library for improving the properties of the lead molecule. This transformer-based model, MolDecor (Molecule Decorator), was trained on drug-like molecules to learn the optimal decorators for property optimization at single or multiple attachment points on the main scaffold of the lead molecule. The model was fine-tuned on specific property data sets like solubility and affinity using transfer learning to optimize these properties. In this study, an automated method was developed to generate a property-specific decorator library. By learning the relationship between scaffolds and decorators, the model avoids bias toward the most commonly used decorators. This also ensures the easy synthesizability of the generated molecules. The model was tested on the anticancer drug (Thalidomide), an antimalarial molecule (Compound 2), and the estrogen receptor modulator (Cyclofenil) to enhance solubility. Additionally, the model was applied to optimize the affinities of molecules targeting Janus kinase 1.
- 29 September 2025
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
MolDecor: Leveraging Transformers to Decorate Bioactive Molecules for Property Optimization
Lead optimization is a critical stage in drug discovery, where promising molecules (lead molecules) are further optimized. It involves the refinement of the chemical structure of the lead molecule to improve its pharmacological properties and drug-like characteristics for development into potential therapies. In this study, we developed a pipeline that includes (a) the creation of a property-specific fragment (decorator) library, (b) learning fragment-scaffold relationship using a BERT-based transformer model, and (c) decorating a given scaffold using fragments from the generated fragment library for improving the properties of the lead molecule. This transformer-based model, MolDecor (Molecule Decorator), was trained on drug-like molecules to learn the optimal decorators for property optimization at single or multiple attachment points on the main scaffold of the lead molecule. The model was fine-tuned on specific property data sets like solubility and affinity using transfer learning to optimize these properties. In this study, an automated method was developed to generate a property-specific decorator library. By learning the relationship between scaffolds and decorators, the model avoids bias toward the most commonly used decorators. This also ensures the easy synthesizability of the generated molecules. The model was tested on the anticancer drug (Thalidomide), an antimalarial molecule (Compound 2), and the estrogen receptor modulator (Cyclofenil) to enhance solubility. Additionally, the model was applied to optimize the affinities of molecules targeting Janus kinase 1.
- Open Access
- 26 August 2025
- Published by Springer Nature in Scientific Data
- Vol. 12 (1) , 1-12
A structure-oriented kinetics dataset of enzyme-substrate interactions
Enzymes are essential biological catalysts that drive nearly all biochemical reactions. Understanding their efficiency and specificity involves studying enzyme kinetics, particularly the parameters kcat and Km. However, there is limited data linking these kinetic parameters with the three-dimensional (3D) structures of enzyme-substrate complexes. Since enzyme function is determined by its structure, such mapping enhances insight into structural basis of enzymatic function and supports applications in enzyme design, synthetic biology and metabolic engineering. To address this critical gap, this work presents SKiD (Structure-oriented Kinetics Dataset), a comprehensive, structured dataset integrating kcat and Km values with the corresponding 3D structural data. This is accomplished by integrating data from existing bioinformatics resources using automated programs to process the data and enhancing it with computational predictions. The erroneous data encountered during data integration is manually resolved. Metadata such as literature and assay conditions (e.g., pH and temperature) are preserved. The 3D coordinates of the modelled enzyme-substrate complexes are provided along with their UniProtKB identifier.
- 26 August 2025
- Published by Springer Nature in Scientific Data
- Vol. 12 (1) , 1-12
Open AccessA structure-oriented kinetics dataset of enzyme-substrate interactions
Enzymes are essential biological catalysts that drive nearly all biochemical reactions. Understanding their efficiency and specificity involves studying enzyme kinetics, particularly the parameters kcat and Km. However, there is limited data linking these kinetic parameters with the three-dimensional (3D) structures of enzyme-substrate complexes. Since enzyme function is determined by its structure, such mapping enhances insight into structural basis of enzymatic function and supports applications in enzyme design, synthetic biology and metabolic engineering. To address this critical gap, this work presents SKiD (Structure-oriented Kinetics Dataset), a comprehensive, structured dataset integrating kcat and Km values with the corresponding 3D structural data. This is accomplished by integrating data from existing bioinformatics resources using automated programs to process the data and enhancing it with computational predictions. The erroneous data encountered during data integration is manually resolved. Metadata such as literature and assay conditions (e.g., pH and temperature) are preserved. The 3D coordinates of the modelled enzyme-substrate complexes are provided along with their UniProtKB identifier.
- Open Access
- 2 July 2025
- Published by American Chemical Society (ACS) in ChemRxiv
ACPOpt: A deep learning approach to design structure-aware property optimized anti-cancer peptides
Peptide-based therapeutics have gained significant attention for their potential in targeting specific proteins with high specificity and minimal toxicity. In cancer therapy, anti-cancer peptides (ACPs) have emerged as promising candidates due to their ability to selectively interact with the target protein. However, the design of peptides with optimal biological activity and stability remains challenging. Recent advancements in generative artificial intelligence (AI) provide opportunities to address these challenges by leveraging existing peptide datasets to identify potent sequences with optimized therapeutic properties. This study presents a novel pipeline, ACPOpt (Optimized Anti-cancer peptides), for the generation of target-specific ACPs. The pipeline utilizes deep learning algorithms to design peptides optimized for binding to the Pim1 kinase, a key protein involved in cancer progression and treatment resistance, aiming to enhance the specificity and efficacy of anti-cancer treatments. In the current work, soluble, non-hemolytic peptides with high affinity were generated against the substrate protein binding site of the Pim1 kinase. By integrating advanced peptide generation models with property optimization, this approach offers potential pathway for the development of target-specific peptide-based therapies.
- 2 July 2025
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessACPOpt: A deep learning approach to design structure-aware property optimized anti-cancer peptides
Peptide-based therapeutics have gained significant attention for their potential in targeting specific proteins with high specificity and minimal toxicity. In cancer therapy, anti-cancer peptides (ACPs) have emerged as promising candidates due to their ability to selectively interact with the target protein. However, the design of peptides with optimal biological activity and stability remains challenging. Recent advancements in generative artificial intelligence (AI) provide opportunities to address these challenges by leveraging existing peptide datasets to identify potent sequences with optimized therapeutic properties. This study presents a novel pipeline, ACPOpt (Optimized Anti-cancer peptides), for the generation of target-specific ACPs. The pipeline utilizes deep learning algorithms to design peptides optimized for binding to the Pim1 kinase, a key protein involved in cancer progression and treatment resistance, aiming to enhance the specificity and efficacy of anti-cancer treatments. In the current work, soluble, non-hemolytic peptides with high affinity were generated against the substrate protein binding site of the Pim1 kinase. By integrating advanced peptide generation models with property optimization, this approach offers potential pathway for the development of target-specific peptide-based therapies.
- Open Access
SKiD: A Structure-Oriented Kinetics Database of Enzyme-Substrate Interactions
Enzymes are essential biological catalysts that drive nearly all biochemical reactions. Understanding their efficiency and specificity involves studying enzyme kinetics, particularly the parameters k cat and K m . However, there is limited data linking these kinetic parameters with the three-dimensional (3D) structures of enzyme-substrate complexes. Since enzyme function is determined by its structure, such mapping enhances insight into structural basis of enzymatic function and supports applications in enzyme design, synthetic biology and metabolic engineering. To address this critical gap, this work presents SKiD (Structure-oriented Kinetics Database), a comprehensive, structured database integrating k cat and K m values with the corresponding 3D structural data. This is accomplished by integrating data from existing bioinformatics resources using automated programs to process the data and enhancing it with computational predictions. The erroneous data encountered during data integration is manually resolved. Metadata such as literature and assay conditions (e.g., pH and temperature) are preserved. The 3D coordinates of the modelled enzyme-substrate complexes are provided along with their UniProtKB identifier. The database is freely accessible from: https://doi.org/10.5281/zenodo.15355031 .
Open AccessSKiD: A Structure-Oriented Kinetics Database of Enzyme-Substrate Interactions
Enzymes are essential biological catalysts that drive nearly all biochemical reactions. Understanding their efficiency and specificity involves studying enzyme kinetics, particularly the parameters k cat and K m . However, there is limited data linking these kinetic parameters with the three-dimensional (3D) structures of enzyme-substrate complexes. Since enzyme function is determined by its structure, such mapping enhances insight into structural basis of enzymatic function and supports applications in enzyme design, synthetic biology and metabolic engineering. To address this critical gap, this work presents SKiD (Structure-oriented Kinetics Database), a comprehensive, structured database integrating k cat and K m values with the corresponding 3D structural data. This is accomplished by integrating data from existing bioinformatics resources using automated programs to process the data and enhancing it with computational predictions. The erroneous data encountered during data integration is manually resolved. Metadata such as literature and assay conditions (e.g., pH and temperature) are preserved. The 3D coordinates of the modelled enzyme-substrate complexes are provided along with their UniProtKB identifier. The database is freely accessible from: https://doi.org/10.5281/zenodo.15355031 .
- Open Access
- 20 March 2025
- Published by Oxford University Press (OUP) in Nucleic Acids Research
- Vol. 53 (6)
DRLiPS: a novel method for prediction of druggable RNA-small molecule binding pockets using machine learning
Ribonucleic Acid (RNA) is the central conduit for information transfer in the cell. Identifying potential RNA targets in disease conditions is a challenging task, given the vast repertoire of functional non-coding RNAs in a human cell. A potential druggable target must satisfy several criteria, including disease association, cellular accessibility, binding pockets for drug-like molecules, and minimal cross-reactivity. While several methods exist for prediction of druggable proteins, they cannot be repurposed for RNAs due to fundamental differences in their binding modality. Taking all these constraints into account, a new structure-based model, Druggable RNA-Ligand binding Pocket Selector (DRLiPS), is developed here to predict binding site-level druggability of any given RNA target. A novel strategy for sampling negative binding sites in RNA structures using three parallel approaches is demonstrated here to improve model specificity: backbone motif search, exhaustive pocket prediction, and blind docking. An external blind test dataset has also been curated to showcase the model's generalizability to both experimental and modelled apo state RNA structures. DRLiPS has achieved an F1-score of 0.70, precision of 0.61, specificity of 0.89, and recall of 0.73 on this external test dataset, outperforming two existing methods, DrugPred_RNA and RNACavityMiner. Further analysis indicates that the features selected for model-building generalize well to both apo and holo states with a backbone RMSD tolerance of 3 Å. It can also predict the effect of binding site single point mutations on druggability, which can aid in optimizing synthetic RNA aptamers for small molecule recognition. The DRLiPS model is freely accessible at https://web.iitm.ac.in/bioinfo2/DRLiPS/.
- 20 March 2025
- Published by Oxford University Press (OUP) in Nucleic Acids Research
- Vol. 53 (6)
Open AccessDRLiPS: a novel method for prediction of druggable RNA-small molecule binding pockets using machine learning
Ribonucleic Acid (RNA) is the central conduit for information transfer in the cell. Identifying potential RNA targets in disease conditions is a challenging task, given the vast repertoire of functional non-coding RNAs in a human cell. A potential druggable target must satisfy several criteria, including disease association, cellular accessibility, binding pockets for drug-like molecules, and minimal cross-reactivity. While several methods exist for prediction of druggable proteins, they cannot be repurposed for RNAs due to fundamental differences in their binding modality. Taking all these constraints into account, a new structure-based model, Druggable RNA-Ligand binding Pocket Selector (DRLiPS), is developed here to predict binding site-level druggability of any given RNA target. A novel strategy for sampling negative binding sites in RNA structures using three parallel approaches is demonstrated here to improve model specificity: backbone motif search, exhaustive pocket prediction, and blind docking. An external blind test dataset has also been curated to showcase the model's generalizability to both experimental and modelled apo state RNA structures. DRLiPS has achieved an F1-score of 0.70, precision of 0.61, specificity of 0.89, and recall of 0.73 on this external test dataset, outperforming two existing methods, DrugPred_RNA and RNACavityMiner. Further analysis indicates that the features selected for model-building generalize well to both apo and holo states with a backbone RMSD tolerance of 3 Å. It can also predict the effect of binding site single point mutations on druggability, which can aid in optimizing synthetic RNA aptamers for small molecule recognition. The DRLiPS model is freely accessible at https://web.iitm.ac.in/bioinfo2/DRLiPS/.