Boosting Drug Discovery via Paired Protein Language Model

Researchers at the National University of Singapore (NUS) launched the Paired Protein Language Model (PPLM) and its associated diagnostic toolset. The April 20, 2026 rollout introduces a relational architecture specifically built for modeling protein-protein interactions (PPIs). If you build machine learning pipelines for drug discovery, this model shifts the baseline for predicting how complex biological molecules bind and interact.

Relational Architecture Design

Traditional protein language models typically process single protein sequences in isolation. PPLM encodes paired protein sequences jointly. This approach captures partner-dependent interaction patterns across the sequences. The model utilizes a hybrid intra-/inter-protein attention mechanism to process these relationships.

The architecture applies Rotary embeddings for intra-protein positional encoding. It pairs this with non-positional embeddings for inter-protein residue pairs. This split embedding strategy prevents the model from assuming spatial priors when learning relational features between two separate molecules. The NUS team, led by Professor Zhang Yang alongside co-authors Jun Liu and Hungyu Chen at the Cancer Science Institute of Singapore, trained PPLM on a composite dataset of over three million protein pairs drawn from the Protein Data Bank and the STRING database.

Performance Benchmarks

PPLM demonstrates measurable improvements over existing single-chain models like ESM2. Jointly processing pairs allows the model to predict interaction behaviors with higher precision.

Metric	PPLM Performance Gain	Comparison Baseline
Perplexity Reduction	20% to 23%	ESM2
Interaction Prediction Accuracy	Up to 17% increase	Leading sequence and structure methods

Specialized Pipeline Tools

The release includes three downstream tools optimized for specific stages of therapeutic development. PPLM-PPI classifies binary interactions to determine if two proteins will interact at all. PPLM-Affinity estimates the exact binding strength between interacting proteins. This specific tool models highly complex structural interactions, including antibody-antigen and TCR-pMHC (T-cell receptor and peptide-major histocompatibility) complexes. PPLM-Contact maps the precise interaction interfaces to identify residue-level contacts.

These tools support targeted cancer therapies by enabling more precise optimization of Complementarity-Determining Regions (CDRs) in antibody design. They also allow researchers to identify previously undruggable targets by processing flat PPI surfaces at a proteome scale.

Deployment Requirements

The model weights and codebase operate under a PolyForm Noncommercial License. PPLM requires an x86_64 Linux environment to run. The system dependencies include HH-suite3 and the Uniclust30 database for full functionality.

Evaluate your current sequence-based prediction steps to see where joint encoding can replace isolated chain processing. Integrating PPLM-Affinity into your early-stage screening will yield more accurate binding strength estimates for complex antibody-antigen pairings before moving to physical synthesis.