Poster
in
Workshop: 2nd Generative AI for Biology Workshop
SMICE: enhanced conformational sampling using AlphaFold and coevolutionary information
Yongkai Chen · Samuel Wong · Samuel Kou
Keywords: [ Markov random field model ] [ fold-switching protein ] [ multiple sequence alignment ] [ sequential sampling ] [ contact map. ] [ structure prediction ]
Protein structure prediction has been revolutionized by AlphaFold, yet a key limitation remains:its inability to characterize the multiple conformations of fold-switching proteins. Current approaches to address this limitation within theAlphaFold framework rely on subsampling themultiple sequence alignment (MSA) input, either through random sampling or clustering, butthese methods are statistically inefficient and failto utilize coevolutionary information betweenresidues. We introduce SMICE, a sequential sampling framework that systematically explores theMSA space by incorporating residue-specific frequencies and coevolutionary patterns inferred viaMarkov random fields. On a benchmark set of92 fold-switching proteins, SMICE outperformsexisting methods and substantially improves structural diversity.