My Works

I am a Postdoctoral Associate in Chemical Engineering at Virginia Tech, with a Ph.D. in Physics from University of Central Florida. My work lies at the intersection of AI, molecular simulation, and biomolecular engineering, where I develop intelligent systems to accelerate discovery in drug design, protein engineering, and advanced biomaterials. I specialize in combining large-scale molecular dynamics simulations with machine learning and generative AI to solve complex, multi-scale problems in biology and chemistry.

My research focuses on building agentic AI frameworks for end-to-end scientific workflows, enabling autonomous setup, execution, and debugging of molecular simulations. I develop AI-driven data pipelines that transform scientific literature into machine-learning-ready datasets ( GlycoData), and design high-throughput drug discovery platforms capable of screening millions of compounds. My work also includes transferable coarse-grained modeling and enhanced sampling techniques for accelerated simulations of proteins, polymers, and hybrid biomaterials, alongside generative AI systems for novel oligomer and protein design—bridging physics-based modeling with next-generation AI-driven discovery.

Areas of Expertise

Research Overview

🤖
Agentic AI for End-to-End Scientific Workflows
Multi-agent AI systems that autonomously handle the full lifecycle of molecular simulations: from model construction and force-field parameterization to execution, debugging, and validation. Integrates LLMs with physics-based reasoning and chemistry aware for self-correcting, scalable simulation pipelines.
  • Agentic AI driven CG model development for polar solvents
  • EvoAI: Evolutionary AI agents for material design
🧬
AI-Powered Scientific Data Infrastructure (GlycoData)
AI agents for automated literature mining and data curation, powering GlycoData.org. Extracts structured, ML-ready datasets from scientific papers, including text, tables, and figures, enabling large-scale, FAIR-compliant data pipelines for material discovery.
  • Curated ML ready datasets with 11 carbohydrate properties (10000+ datapoints)
  • GlycoBot: An AI chatbot for answering question about carbohydrates, trined on GlycoData
⚗️
Advanced Molecular Dynamics & Enhanced Sampling
Equilibrium and non-equilibrium MD simulations of proteins, polymers, and hybrid biomaterials using state-of-the-art techniques to uncover rare events, binding pathways, and conformational landscapes at microsecond timescales.
  • Metadynamics & free energy calculations
  • Accelerated sampling methods
  • Microsecond-scale conformational analysis
💊
AI-Driven Drug Discovery at Scale
End-to-end high-throughput virtual screening (HTVS) platform evaluating millions of compounds against viral protein targets with high predictive accuracy.
  • Multi-engine docking: AutoDock Vina, PLANTS, LeDock, rDock
  • MM/GBSA binding energy refinement
  • MD validation with knockout algorithms
🧪
Transferable Coarse-Grained (CG) Modeling
Next-generation CG force field with 1000+ cross-interaction parameters, optimized via hybrid AI/physics approaches. Enables transferable, constraint-free simulations of proteins and biomaterials while preserving structural fidelity.
  • Advanced optimizer with dimensionality reduction for force field parameterization
  • Capable of modeling of heterogeneous proteins with alpha-helices, beta-sheets, and coils
🧠
ML + Generative AI for Molecular Design
Machine learning and LLM-based systems for closed-loop molecular design combining deep learning, generative models, and simulation feedback.
  • Novel oligomer & protein discovery
  • Property prediction & optimization
  • Agentic exploration of chemical space
🚀
Research Vision
Building autonomous scientific discovery systems that unify AI reasoning, simulation, and experimental data.
  • Drug discovery
  • Biomaterials engineering
  • Protein & polymer design
Molecular simulation

Restraint-free coarse-grained protein models

Coarse-grained protein models that retain secondary structure without structural restraints.

CGAgentX From molecular structure to a coarse-grained model

A master agent coordinates six specialists to build coarse-grained models from atomistic structures and iteratively optimize force fields through hypothesis generation.

DMSO agentic AI optimization

Recent Research

Agentic AI, materials discovery, and molecular dynamics simulations.

CGAgentX workflow and coarse-grained molecular models

CGAgentX: Agentic AI Framework to Develop Transferable Coarse-Grained Models

  • An autonomous, end-to-end agentic AI framework was developed for molecular model development.
  • Months of work were reduced to days through hypothesis-driven force-field optimization and automated error diagnosis.
  • DMSO, DMA, and Tersoff force fields for 2D materials were developed, reproducing targeted experimental properties within 5% error.
Diagram of random hetero-oligomers binding to a green fluorescent protein and enhancing fluorescence

Tuning Polyacrylate Composition to Recognize and Modulate Fluorescent Proteins

  • Mapped GFP–oligomer binding landscapes using well-tempered metadynamics.
  • Identified six deep binding minima for stronger binders, compared with two to three localized wells for weaker binders.
  • Achieved nanomolar GFP binding and over 100% fluorescence enhancement with random-sequence polyacrylates.

De Novo Autogenic Engineered Living Functional Materials

  • Evaluated β-solenoid protein stability using AlphaFold2 and all-atom molecular dynamics simulations.
  • Identified stabilizing hydrogen bonds and electrostatic interactions through molecular analysis.
  • Engineered living materials with threefold higher storage modulus and fivefold longer cell survival under UV-C exposure.

Earlier Research on Nanopores

Polymer capture and translocation snapshots
Article
Capture to Translocation - The Journey of a Polymer
  • Polymer capture time distribution depends on the initial release locations.
  • Capture follows the Poisson process.
  • Even in the nano-pore limit, a fully flexible chain has a finite probability of hairpin-loop capture.
  • Single-file capture can be enhanced in two ways
    • By artificially stiffening the polymer.
    • By adding a charge tag at one of the end terminus.
DNA barcode scanning through a cylindrical nanopore
Article
Barcodes using a Cylindrical Nanopore
  • Eight protein tags are attached to a 16.6 µm long λ-phage ds-DNA.
  • Altering the bias force, we scan the DNA multiple times through the cylindrical nanopore.
  • We observe that heavier protein tags make the DNA velocity non-uniform that leads to inaccurate barcodes.
  • We introduce a novel interpolation scheme by incorporating both the tag and scan velocities that rectifies the barcodes.
Charged protein tags passing through a double nanopore
Article
Discriminating Protein Tags using a Double Nanopore device
  • Using the BD simulation we identify that the charge interaction of the tags with the electric field leads to an asymmetric dwell time distribution.
  • The average dwell time and the degree of asymmetries due to opposing and favoring local fields follow power laws as a function of the charge as well as the length of the protein tags.
  • Lastly, we compare the P'eclet number and shows a close agreement with the experiment.

My Academic Timeline

Swarnadeep Seth's academic timeline