Predicting Binders and Conformational Changes in Metamorphic and Switch Proteins for Therapeutic Design

Sam O’Brien

Mentor: Dr. David Bell, Computational Chemistry and Modeling Group, Advanced Biomedical and Computation Sciences, Frederick National Laboratory for Cancer Research.

Date/Time: August 22nd, 2024 at 2:20pm.

Abstract: Protein structure prediction programs have shown markedly improved accuracy in recent years. Despite this progress, model limitations prevent these programs from accurately capturing the structure of all proteins. AlphaFold2 struggles to predict multiple alternative conformations for metamorphic and switch proteins. Here, I explored a strategy known as AFCluster, which feeds AlphaFold2 limited multiple sequence alignment data to alter its predictions. I tested AFCluster to determine its effectiveness for predicting multiple conformations of known metamorphic and switch proteins. I found that AFCluster accurately predicted experimentally-verified alternative conformations for three out of six metamorphic and switch proteins. However, AFCluster failed to predict alternate conformations for a disordered loop region of EpCAM, a protein of interest with relevance for colorectal cancer. Consequently, I applied RFDiffusion to predict protein binders for EpCAM’s canonical conformation. I assessed and ranked protein binders using unfolding free energy values estimated from an Inverse-Folding-trained protein language model, Evolutionary Scale Model (ESM-IF). These explorations into, and refinement of, structure and binding prediction methods and pipelines will inform future therapeutic design and target identification.