Reddy, S. T.
A single transformer-based architecture, inspired by the mathematical equivalence between attention and the Boltzmann distribution, is developed as a Specificity Foundation Model (SFM) for six molecular recognition domains.
Molecular recognition specificity prediction relies on domain-specific experimental screening or computational tools that do not generalize across different binding modalities.
Leveraging the identity between transformer softmax attention and the Boltzmann distribution, the authors design a universal architecture with dual sequence encoders, symmetric contrastive learning, and a learned physical temperature. This identical architecture is applied without modification to six domains: transcription factor-DNA, enzyme-substrate, peptide-MHC, CRISPR gRNA-off-target DNA, microRNA-mRNA, and small molecule-protein.
All six SFMs achieve high cross-modal retrieval performance (R@1 from 27.7% to 98.0%) using only sequence data. The mir-SFM retrieves miRNA targets at 98.0% R@1, including ~80% of validated interactions missed by seed-matching tools. The crisprSFM improves CRISPR off-target prediction precision to 94.0% from 33.2%. The work demonstrates that a single, physics-derived architecture can achieve high accuracy across diverse molecular recognition tasks without structural information.