Selected research

Scientific question

Can protein sequences be redesigned to lower predicted immunogenicity while preserving structure and function?

Why it matters

Therapeutic proteins that provoke unwanted immune responses can lose efficacy or safety. Computational deimmunization offers a way to screen for lower-risk sequence variants before wet-lab validation.

Approach

Combined Markov chain Monte Carlo (MCMC) sequence sampling with NetMHC-predicted immunogenicity scores and inverse-folding scores from ESM-IF1 and ProteinMPNN, balancing immunogenicity reduction against structural and functional constraints. Separately, added property guidance to ADFLIP to steer deimmunized sequence design.

My contribution

• Implemented the MCMC optimization loop combining NetMHC and inverse-folding scores under competing constraints
• Added property guidance to ADFLIP for deimmunized sequence design

Evidence & results

Methods & tools

Scientific question

How can geometric deep-learning architectures better capture protein structure, including antibody-antigen interactions and hard-to-predict regions?

Why it matters

Accurate structure prediction underlies most downstream design work. Improving attention mechanisms and reproducing state-of-the-art pipelines makes prediction more useful for design tasks that follow it.

Approach

Reproduced AlphaFold-style monomer and multimer models in PyTorch across CUDA and ROCm environments, extended SE(3)-Transformer self-attention with additional kernels, fine-tuned AlphaFold-derived models for CDR-loop and antibody-antigen prediction, and built a peptide-design network targeting a specified binding location.

My contribution

• Reproduced AlphaFold monomer/multimer training in PyTorch for multi-node, multi-GPU environments
• Extended SE(3)-Transformer attention with additional self-attention kernels
• Fine-tuned models for CDR-loop and antibody-antigen interaction prediction
• Built a deep-learning approach for peptide sequence design targeting a specified binding location

Evidence & results

Methods & tools

Related

• View dissertation (opens in a new tab)

Scientific question

Can pH-sensitive binders be designed computationally to control target engagement across physiological environments?

Why it matters

pH-sensitive binding can enable more selective or tunable therapeutic behavior. This is active, ongoing work rather than a completed and validated result.

Approach

Designing pH-sensitive nanobody and ligand binders for Pdgfr-b using LigandMPNN and RFD3.

My contribution

• Applying LigandMPNN and RFD3 to design pH-sensitive nanobody and ligand candidates

Methods & tools

Scientific question

How can imaging and data-processing pipelines better classify and compress macromolecular crystallography data?

Why it matters

High data-rate crystallography generates large volumes of diffraction data. Better classification and compression make large-scale structural studies more tractable and can reveal subtle structural heterogeneity.

Approach

Built a CNN-based classifier for indexable diffraction images, implemented lossy compression to reduce diffraction-image storage, and used structure-factor clustering to detect movement in binding hotspots.

My contribution

• Developed a CNN in PyTorch to detect and classify indexable diffraction images
• Implemented lossy compression for X-ray diffraction images to reduce storage
• Applied structure-factor clustering to detect binding-hotspot movement, revealing small structural differences in chymotrypsinogen
• Co-developed a technique to classify diffraction data from dynamic proteins by individual polymorph

Methods & tools

Related

• Classifying diffraction data by polymorph (Acta Crystallographica D, 2022) (opens in a new tab)
• Best practices for HDRMX (Structural Dynamics, 2020) (opens in a new tab)

Scientific question

How can research-scale deep learning training be made reproducible and portable across accelerator ecosystems and cloud infrastructure?

Why it matters

Scientific AI models are only useful if they can be trained reliably at scale. Portability across CUDA/ROCm and cloud batch systems reduces friction for research teams reusing these pipelines.

Approach

Built PyTorch training pipelines compatible with both CUDA and ROCm across multi-node, multi-GPU clusters, and set up folding-model workflows on Google Cloud Platform for other users via GCP Batch and Vertex AI.

My contribution

• Made AlphaFold reproduction code compatible with both ROCm and CUDA for multi-node training
• Set up folding models on GCP for internal users
• Pushed training workflows to GCP Batch and Vertex AI

Methods & tools