A Smaller AI Model Learns How Proteins Interact, Challenging the Need for Massive Protein Language Models

Understanding how proteins interact is crucial for decoding cellular functions and developing targeted therapeutics. However, current AI models struggle to predict these complex interactions without relying on massive, computationally expensive systems. Now, a highly efficient computational model called MSA Pairformer accurately captures the coevolution of protein-protein interactions using a fraction of standard computing power. The work, led by senior author Sergey Ovchinnikov at the Massachusetts Institute of Technology, with first author Yo Akiyama, is reported in Cell.

The researchers built MSA Pairformer to process multiple sequence alignments (MSAs)—collections of evolutionarily related protein sequences. Unlike traditional protein language models that scale up parameters to store evolutionary statistics internally, this architecture learns to extract physical constraints directly from the aligned sequences. In benchmark tests on bacterial complexes, the model demonstrated a nearly three-fold improvement over existing methods in predicting the physical contact points between interacting proteins. It also successfully distinguished between binding and non-binding mutations at protein interfaces, such as those found in toxin-antitoxin systems.

Remarkably, MSA Pairformer achieves these state-of-the-art predictions using only 111 million parameters, less than 1% the size of frontier single-sequence models. By utilizing a "query-biased" attention mechanism, the model selectively weighs sequences based on evolutionary relevance, allowing it to discover distinct, subfamily-specific binding modes that other models average out. For researchers, this provides an accessible, biologically grounded alternative to the prevailing AI trend of simply scaling up model sizes, allowing advanced protein interaction modeling to run on consumer-grade hardware.

The study authors note important limitations to this approach. Because the model fundamentally relies on deep sequence alignments, it may struggle with newly evolved proteins, synthetically designed de novo proteins with few known evolutionary relatives, or sequences containing complex evolutionary rearrangements like duplications, inversions, and circular permutations.

By demonstrating that biologically informed architectures can outperform sheer computational scale, this development provides a powerful and accessible new tool for engineering molecular assemblies and mapping complex interaction networks.

Reference:
Akiyama, Y., Zhang, Z., Tang, O., et al. "Expanding the scope of protein language modeling to protein-protein interactions with MSA Pairformer." Cell, 2026.