Dr. Sudip Das

Research

ML for Biophysical Sciences Lab   Our research focuses on understanding chemical and biological systems of medical, industrial, and environmental relevance using advanced computational approaches grounded in statistical physics, quantum mechanics, and molecular simulations. We elucidate the structure, function, free energy landscapes, and kinetics of reactive and conformational events in proteins and catalytic systems. Our group integrates molecular dynamics (MD), enhanced sampling, and machine learning (ML) to overcome time- and length-scale limitations in conventional simulations. Recently, we contributed to a physics-informed ML approach (an ML-driven, probability-based framework) for blind exploration of enzymatic catalytic space, enabling mechanistic discovery and characterisation of accurate transition-state (TS) ensembles that govern the thermodynamics and kinetics of the entire catalytic turnover with minimal prior intuition. Building on this foundation, our goals centre on developing scalable, explainable, and transferable AI-assisted quantum models — capable of capturing quantum-level insights even for large biomolecular systems. We aim to make impactful contributions to healthcare, industrial catalysis, and sustainable development. Our research theme is TS-based design principle. In this regard, our current research directions are:   1. Development of Dynamic & Covalent Docking Platform for High-throughput Drug Design Covalent inhibitors are emerging as promising alternatives to the traditional non-covalent and reversible inhibitors, as the former require less number of doses and have less off-target activity. Though there are several sophisticated non-covalent docking tools for high-throughput computer-aided drug discovery, such tools for these emerging covalent inhibitors are lacking. In this regard, our group will develop an automated high-throughput covalent docking pipeline that integrates:   a) Both kinetics and thermodynamics parameters as scoring functions.   b) TS-informed rational design of the warhead of the inhibitor and the target. c) More accurate Quantitative Dynamics-Activity Relationship (QDAR) than traditional Quantitative Structure-Activity Relationship (QSAR).   d) Transfer learning to accelerate screening of drug candidates. e) Explainable AI for speed and interpretability. 2. Modelling Entropy-Driven Assembly Processes The traditional methods for calculating TS depend on the potential energy surface, and hence, they are limited to only enthalpy-driven processes. However, the ML-based probabilistic method, to which we recently contributed, searches the TS ensemble dynamically, hence, is applicable to the entropy-driven processes as well. We aim to reformulate this method for its broader applicability to study entropy-driven processes with significant application towards material and pharmaceutical industries, such as polymer collapse, polymeric aggregation, amyloid fibril formation, protein-protein aggregation, liquid-liquid phase separation (LLPS) and biomolecular condensation. 3. Condensate-membrane Interactions Liquid–liquid phase separation (LLPS) of proteins on biological membranes regulates key cellular processes such as endocytosis, autophagy, and membrane remodelling. Although condensates in bulk solution are well characterised, the molecular details of condensate–membrane interfaces remain poorly understood. Recent findings suggest that proteins and membranes undergo coupled phase behaviours, including prewetting transitions, where dense protein layers form at membrane surfaces. Understanding these phenomena is crucial for revealing how cells use LLPS to organise functional structures and remodel membranes. This involves simulations of large systems composed of condensate and membrane over a long period of time. We aim to develop ML-based coarse-grained models to perform molecular dynamics simulations at these long time and length scales. In this regard, we will employ integrative modelling, combining both atomistic and coarse-grained molecular dynamics simulations and enhanced sampling, supported by biophysical experiments (through collaborations). 4. Modelling Challenging Catalytic Events   Although Density Functional Theory (DFT), hybrid Quantum Mechanics and Molecular Mechanics (QM/MM), and ab initio Molecular Dynamics (MD) provide electronic-level accuracy, their high computational cost limits applications to large and/or slow systems. ML potentials offer efficiency but require extensive quantum training data. Semi-empirical approaches such as Density Functional Tight Binding (DFTB) are computationally efficient, yet lack accurate transition-state (TS) parameterisation, particularly for polarised or high charge-transfer TSs. To address this challenge, we will develop a Transition-State-Informed DFTB (TSI-DFTB) potential by reparameterising DFTB with TS ensemble data generated by our recently developed ML-based probabilistic method. We will employ this newly developed TSI-DFTB potential for high-throughput modelling of challenging catalytic events involving long-range electron/proton transfer and/or strongly polarised TSs, such as: a) Allosteric inhibition targeting metalloenzymes to treat bacterial infection. b) Enzyme-mimicking heterogeneous catalysis within supramolecular cavities and metal-organic or covalent-organic frameworks (MOFs & COFs) with application in industrial catalysis (C-H bond activation) and clean energy production and sustainable development (CO2 capture, utilisation and storage).