Subhajit Roy

Machine Learning & AI Development

• Created a Scalable Interpolant Transformer with SE(3)-invariant coordinates to generate equilibrated 3D conformations of DNA/RNA from sequence inputs, achieving atomic-level structural accuracy.
• Extended the transformer architecture to produce relaxed 3D DNA origami structures from line diagrams, enabling biophysics-informed design of complex nanostructures.
• Designed oxView LLM, fine-tuning the Qwen model for end-to-end DNA/RNA structure generation, relaxation, simulation, and analysis through natural language prompts.
• Improved DNA-PAINT super-resolution microscopy to 1nm resolution using DBSCAN clustering algorithms for 3D origami geometry deconvolution.
• Integrated oxDNA simulation with ML algorithms to uncover detailed nucleic acid dynamics and folding pathways.
• Accelerated AFM image analysis of DNA origami using OpenCV-based automated feature detection.
• Developed an LLM-based autograder using LangChain for ASU Physics experimental coursework, reducing grading time by 60% while improving feedback consistency.
• Built fastDNA: a flow-matching generative model (SE(3)-equivariant transformer, Clifford geometric algebra attention) for generating equilibrium 3D DNA/RNA conformations; trained on HPC with PyTorch Lightning and Weights & Biases; model sizes 2M–70M parameters.
• Developed PhysicsGrader / Graviton: a production LLM grading system (LangChain, Next.js, Canvas LMS API) serving 395 students across 10 lab courses; implemented hallucination/factuality checks, RBAC, MFA, and data encryption (ASU compliance).
• Built NanoCanvas: a cadnano2-inspired DNA origami design tool with a full LLM agent REST API (FastAPI, React, Three.js) enabling programmatic structural design via Model Context Protocol (MCP).
• Developed RAG-backed retrieval pipelines using LangChain, FAISS vector database, and Hugging Face embeddings for domain-specific scientific document search and context-aware LLM responses.

Molecular Simulation & Force Field Development

• Active core developer of oxDNA and oxView: implemented non-singular mass support, modern GPU compatibility (CUDA), and ARM architecture optimization.
• Co-developed algorithms for inverse design of 3D polycube architectures, determining minimal building block sets and interaction rules for target geometries.
• Built custom CUDA-optimized code to simulate unprecedented system sizes, scaling molecular dynamics simulations to ~6 million particles.
• Formulated protein-DNA hybrid system (PROTAC) denaturation pathways using quantum-classical all-atom simulations for targeted protein degradation.
• Developed a novel coarse-grain force field for irregular origami shapes using rigid body physics and patchy particle models, achieving a 1000× speedup over oxDNA2.
• Enhanced the patchy particle force field for DNA-mediated crystal growth with improved dynamics and phase behavior prediction.
• Achieved experimental validation: increased DNA origami crystal growth temperature by 12°C with higher yield based on simulation-guided design.
• Designed DNA icosahedron-helix crystal systems with optical properties at visible wavelengths for photonic applications.
• Built oxCloud: a unified platform for simulation, analysis, and hosting of DNA origami structures, simplifying the complete simulation pipeline for experimentalists.
• Developed a Babylon.js-based 3D visualization platform for non-standard coarse-grain molecular models.
• Fitted SAXS experimental results with simulation using constrained GROMACS simulation.

Computational Infrastructure & Experimental Validation

• Set up and maintain Linux HPC clusters with x86-64 and ARM64-v8 CPUs and NVIDIA GPUs using the SLURM workload manager.
• Hands-on experience with experimental techniques: DNA gel electrophoresis, PCR, DNA-PAINT super-resolution microscopy, AFM imaging.
• Validated computational predictions through wet-lab experiments, closing the loop between simulation and experimental design.

Binding Affinity Prediction

• Generated the PLAS-5k dataset: 5,000 protein-ligand complexes with binding affinities for ML training (published in Scientific Data).
• Developed a CNN model achieving a Pearson correlation coefficient of 0.96 in 10-fold cross-validation.
• Automated all-atom MD simulations and extracted binding parameters (MM-PBSA/MM-GBSA), outperforming AutoDock Vina in accuracy and speed.
• Published an open-access dataset enabling broader community research in drug discovery.

Viral Protein Dynamics

• Investigated allosteric destabilization of the Zika virus NS1 protein β-ladder domain through disulfide bond reduction.
• Performed extensive NAMD simulations, analyzing conformational entropy, enthalpy, SASA, and structural persistence.
• Identified biochemical pathways for NS1 destabilization, informing structure-based drug design strategies (published in Biophysical Journal).