TriFit: Trimodal Fusion with Protein Dynamics for Mutation Fitness Prediction

TriFit addresses a known blind spot in protein AI: the reliance on static structures (snapshots) rather than conformational ensembles (dynamics). While models like ESM-3 and AlphaFold 3 are moving toward multi-modal inputs, TriFit specifically targets 'allosteric coupling' and 'residue flexibility' as supervised signals, which are often neglected in standard pLM or structure-based workflows. With 0 stars and a 4-day age, the project currently has zero market defensibility and exists purely as a research artifact. However, the technical premise—fusion of sequence, structure, and dynamics—is highly relevant to specialized therapeutic engineering where static models fail to predict long-range functional impacts. The primary risk is that frontier labs (e.g., EvolutionaryScale, DeepMind) are already training on broader datasets that implicitly capture some of this through massive scale, potentially rendering specialized 'trimodal' architectures obsolete if the data scale can overcome the need for explicit dynamics modeling. From an investor perspective, this is a 'wait and see' on whether the dynamics-specific architecture yields a significant delta in Spearman correlation over standard ESM-2 zero-shot predictors.