NUS AI Model 'Reads' Protein Pairs, Unlocking Key Insights into Disease and Drug Discovery

Singapore's National University of Singapore (NUS) has achieved a groundbreaking advancement in biomedical research with the development of the Paired Protein Language Model (PPLM), an artificial intelligence tool that analyzes pairs of interacting proteins simultaneously. This innovative approach promises to revolutionize our understanding of cellular processes, paving the way for novel treatments in diseases like cancer and accelerating drug discovery efforts. Led by Professor Zhang Yang from the Cancer Science Institute of Singapore (CSI Singapore), NUS Yong Loo Lin School of Medicine, and NUS School of Computing, the model addresses longstanding challenges in protein-protein interaction (PPI) prediction, a cornerstone of biological function where proteins collaborate to regulate everything from gene expression to signal transduction.

Proteins are the workhorses of life, executing essential tasks within cells. Their interactions form intricate networks that control health and disease. Disruptions in these networks underlie conditions such as cancer, neurodegenerative disorders, and infectious diseases. Traditional experimental methods to map PPIs, like yeast two-hybrid screening or co-immunoprecipitation, are labor-intensive, low-throughput, and prone to false positives or negatives. Computational models have emerged as vital complements, but prior tools often examined proteins individually, missing the nuanced, partner-specific dynamics of interactions.

NUS PPLM: A New Era in Protein Interaction Prediction

PPLM marks a paradigm shift by jointly encoding sequences from two proteins at once, capturing both intrinsic features and context-dependent interaction patterns. Trained on over three million protein pairs sourced from the Protein Data Bank (PDB) and STRING database, the model leverages transformer-based architecture with rotary position embeddings to process paired inputs efficiently. This self-supervised pretraining allows PPLM to discern subtle relational cues that single-sequence models overlook.

The NUS team extended PPLM into three specialized applications: PPLM-PPI for binary interaction prediction (yes/no), PPLM-Affinity for binding strength estimation, and PPLM-Contact for pinpointing interaction interfaces. Step-by-step, the process begins with inputting amino acid sequences of candidate proteins. PPLM generates embeddings that integrate pair-wise context, followed by task-specific heads for output: probability scores for interactions, affinity values in kilocalories per mole, or contact maps highlighting residue-residue proximities below 8 angstroms.

Benchmarked against state-of-the-art tools like ESM2 (a single-sequence language model), AlphaFold-Multimer, and RoseTTAFold-All-Atom, PPLM demonstrated superior performance. On diverse datasets spanning human, yeast, and bacterial species, PPLM-PPI boosted accuracy by up to 17 percent in interaction prediction. PPLM-Affinity excelled in quantifying binding affinities, particularly for transient complexes like antibody-antigen pairs, where it surpassed structure-based predictors such as PRODIGY. PPLM-Contact achieved higher precision in interface residue identification, even when validated against AlphaFold-derived structures.

These gains stem from PPLM's ability to learn biologically meaningful patterns, such as correlated mutations signaling physical contacts. For instance, in T-cell receptor-peptide-major histocompatibility complex (TCR-pMHC) interactions critical for immune responses, the model predicted affinities with unprecedented reliability, opening doors to immunotherapy design.

In disease research, PPIs are implicated in nearly every pathology. Cancer, for example, often arises from aberrant interactions in signaling pathways like RAS-RAF or p53-MDM2. Accurate PPI mapping enables identification of oncogenic hubs. At CSI Singapore, Prof. Zhang's work aligns with cancer-focused missions, where PPLM could reveal novel therapeutic vulnerabilities. Statistics underscore the stakes: over 80 percent of approved drugs target proteins involved in PPIs, yet mapping the human interactome—estimated at hundreds of thousands of edges—remains incomplete. PPLM's scalability supports proteome-wide screens, potentially uncovering disease-specific networks.

Real-world validation came from cross-species benchmarks, confirming generalizability. The open-source availability via GitHub (PPLM repository) democratizes access, fostering global collaboration.

Drug discovery benefits immensely from precise PPI modeling. Binding affinity predictions guide small-molecule inhibitors disrupting pathological interactions, while interface maps inform antibody engineering. PPLM's edge in challenging cases like antibody-antigen binding could streamline vaccine and monoclonal antibody development. In Singapore, where precision medicine initiatives thrive, this tool integrates with national efforts like the Singapore Population Health Studies, enhancing genomic-to-phenotypic linkages.

Consider mutant affinity changes: PPLM accurately forecasted shifts post-single-residue alterations, vital for personalized therapies. For host-pathogen battles, predicting viral-host PPIs could yield antivirals. The full study details are in Nature Communications, affirming its rigor.

Singapore's ascent in AI-biomed fusion powers such innovations. The National AI Strategy 2.0 allocates S$1 billion (2025-2030) for public AI research, with AI for Science receiving S$120 million for interdisciplinary projects. NUS, a hub via its Centre for Data Science and Machine Learning, leads alongside A*STAR. Prof. Zhang's lab, renowned for I-TASSER and D-I-TASSER—outperforming AlphaFold in some metrics—exemplifies this synergy. D-I-TASSER blends AI with physics for multi-domain proteins, complementing PPLM's sequence focus.

Government platforms like National Supercomputing Centre provide compute power, while initiatives like SERIUS nurture talent. NUS's interdisciplinary ethos—spanning computing, medicine, and biochemistry—fosters such breakthroughs.

Prof. Zhang Yang, a pioneer in AI-driven structural biology, heads a lab advancing from single-protein to interaction modeling. With tools like DeepFold for RNA and de novo design suites, his group bridges computation and experiment. Quote: “This work highlights AI's role in life sciences, shifting to interaction-aware modeling for multi-protein predictions and AI-guided therapies.” His dual appointments embody NUS's cross-faculty strength.

Looking ahead, PPLM evolves with multimodal integration—structures, dynamics, experiments—for quaternary complexes and host-pathogen modeling. In Singapore, it bolsters the biomedical cluster, attracting talent amid global AI races. Challenges like data scarcity and validation persist, but open models mitigate them.

For Singapore higher education, PPLM spotlights booming AI-biomed careers. NUS programs in bioinformatics, computational biology, and AI produce experts for pharma giants like GSK and Roche hubs here. Demand surges: roles in structural modeling, drug design yield high salaries (S$100,000+ entry-level). Universities like NTU and SMU complement with AI labs, creating an ecosystem for PhDs and postdocs. Explore research positions driving such innovations.

Photo by Lily Banse on Unsplash

NUS's PPLM exemplifies Singapore's vision as an AI powerhouse, transforming protein science into actionable health solutions. By decoding protein dialogues, it heralds precise interventions against diseases, cementing NUS's global stature and inspiring the next generation of researchers.