PLM-interact

Whether two proteins physically touch is a property of the pair, yet most sequence-based interaction predictors never let a model see both halves at once. The standard recipe runs each protein separately through a frozen protein language model and trains a small classifier on the two independent embedding vectors. Nothing inside the encoder ever attends across the boundary between the partners, so the evidence for binding has to be reconstructed downstream from two summaries computed as if the other protein did not exist.

PLM-interact removes that separation. The two amino acid sequences are concatenated into a single token stream and passed through one ESM-2 encoder, so self-attention runs across both chains from the first layer; the classification token embedding then feeds a single linear layer that emits an interaction probability. The encoder is retrained rather than frozen, under a combined masked-language-modelling and binary-interaction objective. The analogy the authors draw is BERT's next-sentence prediction — a pair of sequences is scored as a pair, not as two independent inputs.

The model comes from the MRC-University of Glasgow Centre for Virus Research with the University of Glasgow Schools of Cancer Sciences and Computing Science and the Cancer Research UK Scotland Institute, posted as a preprint in November 2024 and published in Nature Communications in 2025.

#Key Features

• Joint encoding of the pair: Both sequences enter one encoder as a single token stream, so inter-protein context is learned inside the language model rather than inferred by a downstream head.
• Masked-language modelling kept in the loop: Training combines an MLM loss over the concatenated pair with the binary interaction loss, so the encoder keeps a general sequence representation instead of collapsing onto the classification shortcut.
• Cross-species transfer from a fixed checkpoint: A model trained only on human PPIs is applied unchanged to mouse, fly, worm, yeast and E. coli, held out at the organism level rather than the split level.
• Mutation-effect fine-tuning: A sibling checkpoint scores the log ratio of interaction probability between a mutant and canonical pair, separating mutations that strengthen an interaction from those that disrupt it.
• Virus-host checkpoint: A separately trained model covers virus-human pairs, carrying the same architecture into pathogen interactome work.

#Technical Details

The backbone is ESM-2 650M (esm2_t33_650M_UR50D, 1280-dimensional embeddings); a 35M version is released as well, and the larger backbone performs better throughout. The pair length limit is a runtime parameter, set to 1603 in the documented inference example. The loss is a weighted sum of masked-LM and cross-entropy terms, and a sweep over mask ratios of 7%, 15%, 22% and 30% found 15% the only setting that consistently beat an unmasked binary baseline. On the cross-species benchmark of Sledzieski et al., built from STRING v11 and redundancy-reduced at 40% identity with CD-HIT, training uses 38,344 positive human PPIs against a 1:10 negative ratio, with 5,000 positive pairs per test species except E. coli, which has 2,000. PLM-interact reaches AUPR 0.706 on yeast against TUnA's 0.641 and 0.722 on E. coli, with gains of 2%, 8% and 6% over TUnA on mouse, fly and worm. On the leakage-free benchmark of Bernett et al. it matches TUnA on AUPR and AUROC while improving recall by 9% at a 0.5 threshold. Full-layer fine-tuning lifts mutation-effect AUPR by 150% over the zero-shot model, and on 22,383 virus-human PPIs from HPIDB 3.0 it improves on STEP by 5.7% AUPR, 10.9% F1 and 11.9% MCC. Training the STRING v12 human checkpoint took 86.4 hours on 16 A100-80GB GPUs.

#Applications

Proteome-scale interaction screening is the direct use: scoring pairs from sequence alone is cheap enough to run over candidate lists that per-pair structure prediction cannot reach, and the output thresholds cleanly into a shortlist for co-immunoprecipitation or two-hybrid follow-up. Because the human-trained checkpoint transfers to other organisms, it reaches species whose interactomes are sparsely mapped. The mutation checkpoint speaks to variant interpretation — whether a missense change breaks a specific interaction — and the virus-host checkpoint supports host-factor discovery for emerging pathogens. Inference runs locally on a GPU, with an A40 or A100 recommended for the 2.6 GB checkpoint.

#Impact

The result argues that the bottleneck in language-model-based interaction prediction was the frozen embedding interface rather than the language model itself, and that retraining the encoder on pairs recovers signal that a downstream head cannot. Code is MIT-licensed, and checkpoints for the human v11 and v12, leakage-free, virus-host and mutation settings are on Hugging Face alongside their training sets, so the comparisons are reproducible. The limitations are worth stating plainly: the deployed head emits one interaction probability rather than a general-purpose pair embedding, so downstream reuse is narrower than a general protein language model; the leakage-free benchmark shows the cross-species margins shrink once train-test protein overlap is controlled; and the model says whether two proteins interact, not where they touch or what the complex looks like.