NUS scientists devise AI model that “reads” protein pairs, unlocking new insights into disease and drug discovery

NUS scientists devise AI model that "reads" protein pairs, unlocking new insights into disease and drug discovery

An error occurred while preparing your download

Researchers led by Professor Zhang Yang, Senior Principal Investigator from the Cancer Science Institute of Singapore (CSI Singapore) at the National University of Singapore, have developed a new artificial intelligence (AI) model that can more accurately predict how proteins interact with one another-an advance that could accelerate drug discovery and deepen insights into diseases such as cancer.

Published in Nature Communications on 10 March 2026, the study introduces a paired protein language model (PPLM) that learns from two interacting proteins simultaneously, rather than analysing them in isolation. This marks a significant shift in how AI is applied to biology, enabling more accurate prediction of protein-protein interactions that underpin nearly all cellular processes.

A new way to understand protein interactions

Protein-protein interactions are inherently relational, yet most current AI models are trained on single protein sequences. This limits their ability to fully capture how proteins recognise and bind to one another.

To address this, the research team developed PPLM, a model specifically designed to learn inter-protein relationships during training. By jointly encoding paired protein sequences, PPLM captures both individual protein features and partner-dependent interaction patterns within a unified framework. The model was trained on more than three million protein pairs, enabling it to learn interaction patterns at scale.

An error occurred while preparing your download

Strong performance across multiple tasks

Building on this foundation, the team developed three specialised tools: PPLM-PPI for predicting whether proteins interact, PPLM-Affinity for estimating binding strength, and PPLM-Contact for identifying interaction interfaces. Across benchmark datasets, the model improved interaction prediction accuracy by up to about 17 per cent over leading methods, with consistent gains across multiple species.

Notably, the model outperformed both sequence-based and structure-based methods in challenging scenarios such as antibody-antigen interactions. In addition, the model identified patterns that match how proteins interact in real life, indicating that it can capture biologically meaningful relationships between proteins.

"This work highlights the growing role of AI in transforming the life sciences. By moving from single-protein analysis to interaction-aware modelling, the study lays the groundwork for future advances in multi-protein complex prediction, systems-level biology, and AI-guided therapeutic design," explained Prof Zhang, who also has appointments at the Department of Biochemistry at the NUS Yong Loo Lin School of Medicine and the Department of Computer Science at the NUS School of Computing.

Towards scalable and translational impact

By improving the accuracy and scalability of protein interaction modelling, PPLM could support a wide range of applications, including proteome-scale interaction discovery, drug target identification, and therapeutic development.

The NUS team is now working to further enhance the model by integrating structural and experimental data and extending its application to more complex biological systems such as host-pathogen interactions.