Protein language models have largely treated each sequence in isolation, learning the statistics of single chains and inferring interaction from the result. Researchers at NUS argue that interaction is better learned directly. Their paired protein language model analyses two proteins together as a single input, on the reasoning that the features determining whether and how two chains bind are properties of the pair rather than of either partner considered alone.
The model was trained on more than three million protein pairs and improved interaction-prediction accuracy by up to about 17 per cent over leading methods on benchmark datasets. It outperformed both sequence-based and structure-based approaches in difficult cases, antibody and antigen interactions prominent among them, where shape complementarity is subtle and homologous examples are scarce. The work was led from the Cancer Science Institute of Singapore and published in Nature Communications.
Three tools were built on the shared representation, reflecting the distinct questions practitioners actually ask. One predicts whether two proteins interact at all, a second estimates binding strength, and a third identifies the interaction interface. Separating these allows the same learned embedding to serve candidate screening, affinity estimation and structural hypothesis generation without retraining from scratch for each separate question.
Protein interactions underpin nearly every cellular process and are central to understanding diseases including cancer, which makes reliable prediction valuable well beyond structural biology. The team plans to integrate structural and experimental data and to extend the approach to host and pathogen interactions, a domain where experimental data are especially sparse and computational prioritisation correspondingly useful.