Abhinav Gupta
Abhinav Gupta has a total of 12 co-authors. Here are the top co-authors by number of shared publications:
Norbert Furtmann
Sanofi
2 shared publications- Pierre F. J. Lermusiaux
Department of Mechanical Engineering Massachusetts Institute of Technology Cambridge MA USA
2 shared publications Ruijiang Li
Sanofi
2 shared publications- Bonnie Berger
Massachusetts Institute of Technology
1 shared publication - Bryan D. Bryson
Department of Biological Engineering, Massachusetts Institute of Technology
1 shared publication Bryan Munoz Rivero
Large Molecule Research Sanofi Cambridge Massachusetts USA
1 shared publication
Latest Publications
- Open Access
Breadth over Depth: Predictor-Agnostic Consensus for Improved Antibody--Antigen Structure Prediction
Antibody–antigen structure prediction commonly relies on sampling a single co-folding predictor deeply and selecting its highest-confidence output using model-specific confidence scores. We introduce AbPoolRank, a predictor-agnostic structural selector that ranks mixed pools using cross-pose agreement, within-source reproducibility and interface geometry, without native confidence, predictor identities or predictor-specific parameters. On 103 cluster-disjoint, post-2021 antibody–antigen complexes from SAbDab2, three structures each from Boltz-2, Promera, Protenix-v2, ESMFold2-Fast and OpenDDE form a 15-structure pool that reaches 58.0% medium-or-better Success@1 (DockQ ≥ 0.49), exceeding confidence-based selection from every individual predictor sampled 100 times (26–52%). Existing ranking methods did not rescue deep single-predictor sampling or surpass our selector on mixed pools, with no method exceeding 51.5% on 100 structures from one co-folder and AbPoolRank matching the strongest single-structure methods while significantly outperforming consensus methods. AbPoolRank trained without one predictor lost at most 3.0 percentage points, and the frozen model accepted full MSA-conditioned ESMFold2, a sixth predictor absent from method development, without retraining or loss of accuracy. Diversification also improved the accuracy–compute frontier, with the 15-structure pool using 1.8-fold less aggregate GPU time than the most accurate single-predictor setting in our compute benchmark and 3.3-fold less than sampling that predictor 100 times. Together, these results show that shallow sampling across complementary co-folders, coupled with a confidence-free selector, provides an accurate and computationally efficient inference strategy for antibody–antigen structure prediction.
Open AccessBreadth over Depth: Predictor-Agnostic Consensus for Improved Antibody--Antigen Structure Prediction
Antibody–antigen structure prediction commonly relies on sampling a single co-folding predictor deeply and selecting its highest-confidence output using model-specific confidence scores. We introduce AbPoolRank, a predictor-agnostic structural selector that ranks mixed pools using cross-pose agreement, within-source reproducibility and interface geometry, without native confidence, predictor identities or predictor-specific parameters. On 103 cluster-disjoint, post-2021 antibody–antigen complexes from SAbDab2, three structures each from Boltz-2, Promera, Protenix-v2, ESMFold2-Fast and OpenDDE form a 15-structure pool that reaches 58.0% medium-or-better Success@1 (DockQ ≥ 0.49), exceeding confidence-based selection from every individual predictor sampled 100 times (26–52%). Existing ranking methods did not rescue deep single-predictor sampling or surpass our selector on mixed pools, with no method exceeding 51.5% on 100 structures from one co-folder and AbPoolRank matching the strongest single-structure methods while significantly outperforming consensus methods. AbPoolRank trained without one predictor lost at most 3.0 percentage points, and the frozen model accepted full MSA-conditioned ESMFold2, a sixth predictor absent from method development, without retraining or loss of accuracy. Diversification also improved the accuracy–compute frontier, with the 15-structure pool using 1.8-fold less aggregate GPU time than the most accurate single-predictor setting in our compute benchmark and 3.3-fold less than sampling that predictor 100 times. Together, these results show that shallow sampling across complementary co-folders, coupled with a confidence-free selector, provides an accurate and computationally efficient inference strategy for antibody–antigen structure prediction.
- 7 June 2026
- Published by Wiley in Protein Science
- Vol. 35 (7) , e70655
SNAC ‐ DB : An ML ‐ready database for antibody and NANOBODY ® VHH –antigen complexes with expanded structural diversity and real‐world benchmarking
Predicting antibody and NANOBODY® VHH–antigen complexes remains a critical challenge for state‐of‐the‐art structure prediction models, limiting their impact in therapeutic discovery pipelines. We introduce SNAC‐DB, an ML‐ready database and curation pipeline enriched with structural biology expertise, designed to accelerate model accuracy and generalization by providing 31–37% expanded structural diversity over existing resources like SAbDab through comprehensive re‐curation that extracts maximum value from available experimental structures. SNAC‐DB expands coverage by capturing often‐overlooked complexes and accurately identifying complete multi‐chain epitopes through improved biological‐assembly based logic. Built for ML practitioners, SNAC‐DB provides standardized formats with multi‐threshold structure‐based clustering to enable principled sample weighting during training. Using a rigorous benchmark of public PDB entries deposited post‐May 2024 plus confidential therapeutic structures, we evaluate seven leading models (Protenix‐v1, OpenFold‐3p2, RosettaFold‐3, Boltz‐2, Boltz‐1x, Chai‐1, and AlphaFold2.3‐multimer) with evaluation methodology tailored to antibody/NANOBODY® VHH–antigen complexes to ensure correct handling of multi‐chain epitopes, revealing systematic performance gaps: success rates rarely exceed 25%, confidence‐based ranking fails to identify best predictions even when accurate structures exist in ensembles, and all models consistently struggle with therapeutically relevant NANOBODY® VHHs. Systematic evaluation of sampling strategies demonstrates that while generating 1000 samples per target substantially increases the likelihood of producing accurate structures (oracle selection improves from 11.9% to 50.5%), confidence‐based ranking remains nearly flat (between 10.9% and 14.9%), revealing that improved ranking mechanisms represent a more tractable path to performance gains. Finally, fine‐tuning GeoDock on SNAC‐DB yields higher success rates than training on SAbDab (11.0% vs. 7.1% for antibodies; 7.0% vs. 4.0% for NANOBODY® VHHs), suggesting that SNAC‐DB's expanded structural diversity translates to improved model generalization.
- 7 June 2026
- Published by Wiley in Protein Science
- Vol. 35 (7) , e70655
SNAC ‐ DB : An ML ‐ready database for antibody and NANOBODY ® VHH –antigen complexes with expanded structural diversity and real‐world benchmarking
Predicting antibody and NANOBODY® VHH–antigen complexes remains a critical challenge for state‐of‐the‐art structure prediction models, limiting their impact in therapeutic discovery pipelines. We introduce SNAC‐DB, an ML‐ready database and curation pipeline enriched with structural biology expertise, designed to accelerate model accuracy and generalization by providing 31–37% expanded structural diversity over existing resources like SAbDab through comprehensive re‐curation that extracts maximum value from available experimental structures. SNAC‐DB expands coverage by capturing often‐overlooked complexes and accurately identifying complete multi‐chain epitopes through improved biological‐assembly based logic. Built for ML practitioners, SNAC‐DB provides standardized formats with multi‐threshold structure‐based clustering to enable principled sample weighting during training. Using a rigorous benchmark of public PDB entries deposited post‐May 2024 plus confidential therapeutic structures, we evaluate seven leading models (Protenix‐v1, OpenFold‐3p2, RosettaFold‐3, Boltz‐2, Boltz‐1x, Chai‐1, and AlphaFold2.3‐multimer) with evaluation methodology tailored to antibody/NANOBODY® VHH–antigen complexes to ensure correct handling of multi‐chain epitopes, revealing systematic performance gaps: success rates rarely exceed 25%, confidence‐based ranking fails to identify best predictions even when accurate structures exist in ensembles, and all models consistently struggle with therapeutically relevant NANOBODY® VHHs. Systematic evaluation of sampling strategies demonstrates that while generating 1000 samples per target substantially increases the likelihood of producing accurate structures (oracle selection improves from 11.9% to 50.5%), confidence‐based ranking remains nearly flat (between 10.9% and 14.9%), revealing that improved ranking mechanisms represent a more tractable path to performance gains. Finally, fine‐tuning GeoDock on SNAC‐DB yields higher success rates than training on SAbDab (11.0% vs. 7.1% for antibodies; 7.0% vs. 4.0% for NANOBODY® VHHs), suggesting that SNAC‐DB's expanded structural diversity translates to improved model generalization.
- Open Access
- 30 June 2023
- Published by Springer Nature in Scientific Reports
- Vol. 13 (1) , 1-21
Generalized neural closure models with interpretability
Improving the predictive capability and computational cost of dynamical models is often at the heart of augmenting computational physics with machine learning (ML). However, most learning results are limited in interpretability and generalization over different computational grid resolutions, initial and boundary conditions, domain geometries, and physical or problem-specific parameters. In the present study, we simultaneously address all these challenges by developing the novel and versatile methodology of unified neural partial delay differential equations. We augment existing/low-fidelity dynamical models directly in their partial differential equation (PDE) forms with both Markovian and non-Markovian neural network (NN) closure parameterizations. The melding of the existing models with NNs in the continuous spatiotemporal space followed by numerical discretization automatically allows for the desired generalizability. The Markovian term is designed to enable extraction of its analytical form and thus provides interpretability. The non-Markovian terms allow accounting for inherently missing time delays needed to represent the real world. Our flexible modeling framework provides full autonomy for the design of the unknown closure terms such as using any linear-, shallow-, or deep-NN architectures, selecting the span of the input function libraries, and using either or both Markovian and non-Markovian closure terms, all in accord with prior knowledge. We obtain adjoint PDEs in the continuous form, thus enabling direct implementation across differentiable and non-differentiable computational physics codes, different ML frameworks, and treatment of nonuniformly-spaced spatiotemporal training data. We demonstrate the new generalized neural closure models (gnCMs) framework using four sets of experiments based on advecting nonlinear waves, shocks, and ocean acidification models. Our learned gnCMs discover missing physics, find leading numerical error terms, discriminate among candidate functional forms in an interpretable fashion, achieve generalization, and compensate for the lack of complexity in simpler models. Finally, we analyze the computational advantages of our new framework.
- 30 June 2023
- Published by Springer Nature in Scientific Reports
- Vol. 13 (1) , 1-21
Open AccessGeneralized neural closure models with interpretability
Improving the predictive capability and computational cost of dynamical models is often at the heart of augmenting computational physics with machine learning (ML). However, most learning results are limited in interpretability and generalization over different computational grid resolutions, initial and boundary conditions, domain geometries, and physical or problem-specific parameters. In the present study, we simultaneously address all these challenges by developing the novel and versatile methodology of unified neural partial delay differential equations. We augment existing/low-fidelity dynamical models directly in their partial differential equation (PDE) forms with both Markovian and non-Markovian neural network (NN) closure parameterizations. The melding of the existing models with NNs in the continuous spatiotemporal space followed by numerical discretization automatically allows for the desired generalizability. The Markovian term is designed to enable extraction of its analytical form and thus provides interpretability. The non-Markovian terms allow accounting for inherently missing time delays needed to represent the real world. Our flexible modeling framework provides full autonomy for the design of the unknown closure terms such as using any linear-, shallow-, or deep-NN architectures, selecting the span of the input function libraries, and using either or both Markovian and non-Markovian closure terms, all in accord with prior knowledge. We obtain adjoint PDEs in the continuous form, thus enabling direct implementation across differentiable and non-differentiable computational physics codes, different ML frameworks, and treatment of nonuniformly-spaced spatiotemporal training data. We demonstrate the new generalized neural closure models (gnCMs) framework using four sets of experiments based on advecting nonlinear waves, shocks, and ocean acidification models. Our learned gnCMs discover missing physics, find leading numerical error terms, discriminate among candidate functional forms in an interpretable fashion, achieve generalization, and compensate for the lack of complexity in simpler models. Finally, we analyze the computational advantages of our new framework.
- Open Access
- 18 August 2021
- Published by The Royal Society in Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences
- Vol. 477 (2252)
Neural closure models for dynamical systems
Complex dynamical systems are used for predictions in many domains. Because of computational costs, models are truncated, coarsened or aggregated. As the neglected and unresolved terms become important, the utility of model predictions diminishes. We develop a novel, versatile and rigorous methodology to learn non-Markovian closure parametrizations for known-physics/low-fidelity models using data from high-fidelity simulations. The new neural closure models augment low-fidelity models with neural delay differential equations (nDDEs), motivated by the Mori–Zwanzig formulation and the inherent delays in complex dynamical systems. We demonstrate that neural closures efficiently account for truncated modes in reduced-order-models, capture the effects of subgrid-scale processes in coarse models and augment the simplification of complex biological and physical–biogeochemical models. We find that using non-Markovian over Markovian closures improves long-term prediction accuracy and requires smaller networks. We derive adjoint equations and network architectures needed to efficiently implement the new discrete and distributed nDDEs, for any time-integration schemes and allowing non-uniformly spaced temporal training data. The performance of discrete over distributed delays in closure models is explained using information theory, and we find an optimal amount of past information for a specified architecture. Finally, we analyse computational complexity and explain the limited additional cost due to neural closure models.
- 18 August 2021
- Published by The Royal Society in Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences
- Vol. 477 (2252)
Open AccessNeural closure models for dynamical systems
Complex dynamical systems are used for predictions in many domains. Because of computational costs, models are truncated, coarsened or aggregated. As the neglected and unresolved terms become important, the utility of model predictions diminishes. We develop a novel, versatile and rigorous methodology to learn non-Markovian closure parametrizations for known-physics/low-fidelity models using data from high-fidelity simulations. The new neural closure models augment low-fidelity models with neural delay differential equations (nDDEs), motivated by the Mori–Zwanzig formulation and the inherent delays in complex dynamical systems. We demonstrate that neural closures efficiently account for truncated modes in reduced-order-models, capture the effects of subgrid-scale processes in coarse models and augment the simplification of complex biological and physical–biogeochemical models. We find that using non-Markovian over Markovian closures improves long-term prediction accuracy and requires smaller networks. We derive adjoint equations and network architectures needed to efficiently implement the new discrete and distributed nDDEs, for any time-integration schemes and allowing non-uniformly spaced temporal training data. The performance of discrete over distributed delays in closure models is explained using information theory, and we find an optimal amount of past information for a specified architecture. Finally, we analyse computational complexity and explain the limited additional cost due to neural closure models.