Work

Protein–protein interactions mediated by peptide-binding domains (PBDs) and short linear motifs (SLiMs) play a key role in regulating cellular signaling pathways. Malfunctions in these pathways are implicated in a range of diseases, including cancer and neurological disorders. While high-throughput experimental methods have uncovered binding preferences within individual PBD families, they fall short of revealing the broader molecular principles that govern binding specificity across all PBD families found in the human proteome. In this work, we apply mechanistic interpretability techniques to protein language models to investigate the biophysical determinants of SLiM recognition across 24 human PBD families. Using sparse autoencoders, we extract interpretable features from protein representations and identify two distinct classes of features: (i) family-specific and (ii) features shared across two families. Our findings take the first steps towards revealing common recognition strategies and diverse specificity mechanisms that underlie PBD–SLiM interactions, offering a unified framework for understanding how signaling proteins achieve both selectivity and promiscuity. Our work highlights the potential of interpretable machine learning to uncover mechanistic insights into protein function and could serve as a future foundation for rationally engineering signaling interactions or diagnosing the effects of disease-associated mutations.

Generative protein-design models produce candidate binders without revealing what internally distinguishes a strong binder from a weak one, limiting rational improvement and failure diagnosis. We apply sparse autoencoders (SAEs), a mechanistic-interpretability technique, to residue-level ESM-C activations of Boltz-designed binder candidates against Vilip-1, a blood biomarker of acute neurological injury, to identify human-interpretable features in these latent representations without modifying any binder sequences. Mixing real, evolutionarily distinct natural-protein sequences into SAE training was the largest lever for generalization beyond synthetic designs alone, raising held-out natural-protein reconstruction fidelity from 0.39 to 0.60 fraction of variance explained. To test whether learned features reflect genuine protein structure rather than training artifacts, we compared two independently trained SAE dictionaries: binder-alone encoding versus encoding with full binder-target cross-attention before target positions were discarded. Despite different encodings, each dictionary's most generic features converge on the same real-protein residues at a rate chance cannot explain (hypergeometric test against the ~7M-residue candidate space, p ≈ 7 × 10⁻²⁸), confirming that cross-attention preserves the learned signal. This convergence survives a shared-seed confound check (5 of 6 agreeing cases use different learned feature indices per dictionary) and is corroborated by InterPro domain annotations and calibrated LLM-assisted labeling that withholds a description when evidence is scattered. Together, these results show that SAE features recovered from protein language model activations carry real, reproducible structural signal, a necessary validation step before using such features to inform or steer binder design.

Parkinson's disease is a form of progressive neurodegeneration that primarily affects dopaminergic neurons. It is characterized by misfolded α-Synuclein (α-Syn) proteins clumped together in Lewy bodies. More recently, it has been proposed that α-Syn toxicity may increase during interactions with fatty acids. There have been several studies linking stearoyl-coenzyme A desaturase 1 (SCD1), the rate-limiting enzyme for the conversion of saturated fatty acids (SFAs) to monounsaturated fatty acids (MUFAs), to the increased toxicity of α-Syn. Consequently, SCD1 inhibition is shown to decrease the toxicity and aggregation of α-Syn. However, the precise interactions of SCD1 inhibitors and SCD1 are unclear. This project compared seven novel analogs of SCD1 inhibitors, which we hypothesized to compete with SCD1's coenzyme stearoyl coenzyme A, decreasing SFA conversion into their respective MUFAs. The analogs shared the same general pharmacophore with varying R groups (p-toluoyl, 4-fluorobenzoyl, 3-trifluoromethyl benzoyl, o-anisoyl, 3,4-difluorobenzoyl, 2-trifluoromethyl benzoyl, and 2-chlorobenzoyl). We hypothesized that analogs with the least steric hindrance would perform best. We drew a structure-activity relationship from in silico studies, with molecular docking results showing that four analogs were just as or more effective than MF-438, a commercially available SCD1 inhibitor. These results imply that the most effective R group was least sterically hindered, guiding further analog development in the field of small molecule Parkinson's disease cures.

Alzheimer's Disease is caused by an aggregation of amyloid beta and tau proteins in the brain. Polyphenols, a broad class of naturally-existing compounds, have been shown to inhibit the aggregation of those proteins. This project aims to focus on expressing different combinations of those proteins, as well as assaying those proteins for aggregation inhibition using polyphenols such as curcumin, caffeic acid, epigallocatechin gallate (EGCG), and more to determine which polyphenol is most effective in doing so. We chose to use these polyphenols because of their past precedence in other work, along with their widespread prevalence. However, this project focused more on the biological and in-vitro aspect of polyphenols inhibiting amyloid beta, such as conducting multiple assays including Congo Red, Avoidance, and Dynamic Light Scattering in order to receive tangible results. Through our studies, we found that polyphenols do produce an inhibitory effect on the aggregation of amyloid-beta.

Investigated synthesis of analogs of rivastigmine to treat Alzheimer's. Presented at 2022 SCCUR.

Led a project using LCMS to quantify neurotransmitters in Parkinson's. Presented at 2023 WCBSURC.

Spearheaded a project investigating fluorophore-labeled peptides. Presented at 2023 Fall ACS.