Machine Learning for Protein Biology

During my PhD, I worked on machine-learning methods that infer biological and structural properties directly from protein sequence.

The work covered several related problems: predicting intrinsically disordered regions and their functions, DNA-binding residues, protein-binding residues, crystal-structure quality, and sequence-derived signals associated with druggability.

A recurring theme was learning useful representations from sequence when direct structural evidence is sparse or unavailable. That meant thinking carefully about labels, evaluation, biological priors, and the gap between benchmark performance and useful scientific inference.

Representative work includes:

• flDPnn — intrinsic disorder prediction with putative propensities of disorder functions (Nature Communications, 2021).
• XRRpred — prediction of crystal-structure quality from protein sequence (Bioinformatics, 2021).
• DNAgenie — DNA-type-specific binding residue prediction (Briefings in Bioinformatics, 2021).
• Work on protein-binding residues, disordered-region function, and sequence-derived markers of drug targets.

This period is also where I developed a lasting interest in the boundary between representation learning, scientific structure, and engineered evaluation — themes that continue into my current work on foundation models and AI systems.

Google Scholar → · XRRpred code → · DNA/RNA binding experiments →