PRIMA: a bidirectional state-space architecture and training approach for sequence modelling of protein-protein interactions

Protein-protein interactions (PPIs) govern essential biological processes, yet building protein language models (PLMs) that directly capture the complexity of interacting multi-chain sequences remains an open challenge. Transformer-based PLMs rely on self-attention whose quadratic complexity makes processing long interacting complexes resource-intensive, leaving an emerging gap between a rich diversity of current single-chain PLMs and scalable PPI PLMs. We present PRIMA, a proof-of-concept PLM for PPIs built on bidirectional Mamba (BiMamba), a selective state-space model that scales linearly with sequence length and processes arbitrarily long inputs without positional encodings. PRIMA is trained in two phases: first on diverse single protein sequences to learn general representations, then on concatenated PPI pairs to capture interaction-specific context. At an 8 million parameter scale and equal count to a transformer baseline, PRIMA (i) outperforms transformer-based PLMs on standard-length inputs, (ii) outperforms them on longer sequences and length extrapolation tasks, and (iii) does so at significantly lower computational cost, establishing BiMamba as a strong backbone for scalable PPI PLMs.