SYNTERACT

Training a classifier to recognize interacting protein pairs requires non-interacting pairs to contrast them against, and those are scarce: curating a protein pair that has been shown not to interact is slow, unrewarding work. After mapping BioGRID's multi-validated physical interactions to UniProt sequences, Hallee and Gleghorn had 179,018 positive pairs and 3,958 negatives from Negatome. The field's usual workaround is to fabricate negatives by pairing random proteins drawn from different subcellular compartments, on the reasoning that proteins which never share a compartment cannot touch.

That workaround leaks. A classifier handed such a dataset can score well by learning compartment membership rather than binding, and this paper demonstrates it directly: a support vector machine given four features per pair — subcellular compartment and solubility for each protein, and nothing about sequence — reaches 65% accuracy on those compiled datasets. The negatives are recognizable by where they came from.

SYNTERACT (SYNThetic data-driven protein-protein intERACtion Transformer) answers that by manufacturing negatives from real pairs instead of assembling them from unrelated proteins. One partner in a pair is held fixed and the other is mutated through BLOSUM62-guided substitutions — deleterious substitutions to turn a positive into a negative, conservative ones to keep a negative negative — or shuffled outright, with the initiating methionine preserved. That yielded over 170,000 synthetic negatives, and each one matches a positive in provenance and compartment, differing only in sequence content. Built at the University of Delaware in 2023, it is the first generation of what Synthyra later grew into the Synteract line, whose descendants include Synteract-4 and Atlas.

#Key Features

• Negatives built by mutation, not by pairing: Synthetic non-interactors are derived from annotated pairs through substitution or shuffling, so the label cannot be recovered from compartment or provenance.
• Pure sequence input: Two amino acid sequences joined by a [SEP] token are the entire input; no structure, alignment, orthology or functional annotation is used.
• Cross-phylogeny holdout: The 1,000-pair test set spans 50 organisms, none of which appear in training, and the model reports 92% accuracy on it.
• A diagnostic as well as a predictor: Retraining the same architecture on compartment-sampled negatives scores 96% on its own test set and 54% on multi-validated pairs, quantifying how far the two objectives diverge.

#Technical Details

SYNTERACT fine-tunes ProtBERT-BFD from ProtTrans — an encoder-only BERT with 30 layers, hidden dimension 1024, intermediate dimension 4096 and 16 attention heads, 420 million parameters, pretrained by masked language modeling over the two-billion-sequence Big Fantastic Database. Pairs are tokenized as [CLS] Protein A [SEP] Protein B [SEP] within a 1,024-token window; the pooled [CLS] embedding passes through a feed-forward head to a two-way softmax. Training used BioGRID release 4.4.213 and Negatome 2.0 across 160 species, with pairs trimmed to under 1,000 combined residues, and ran 20,000 steps at a global batch size of 70 after an AdamW sweep over 24 initializations. On the held-out set it reports 0.92 accuracy, 0.93 precision and 0.92 F1; on BioGRID release 4.4.221, containing 4,534 interactions absent from the training release, 0.96 accuracy at 1.00 precision; and 100% accuracy on randomly generated vertebrate-mimetic sequences. An SVM fitted only on SYNTERACT embeddings of real positives and synthetic negatives correctly classified 238 genuine Negatome pairs it had never seen, at 73% accuracy against 50% from untuned ProtBERT-BFD embeddings — evidence the synthetic negatives approximate real ones. On the compartment-sampled datasets, SYNTERACT scores 85% on positives but 38% on negatives, calling most of them interactions.

#Applications

Because the input is sequence alone, the model applies wherever structures are unavailable: screening a designed peptide against a target, annotating proteins with no recognizable motif or characterized function, and triaging candidate partners before committing to pull-down or two-hybrid experiments. It is the natural downstream check on generative protein design, where the sequences produced have no structural precedent to look up. The released checkpoint runs as a standard BertForSequenceClassification and returns a probability per pair.

#Impact

SYNTERACT's durable contribution is the negative-sampling critique — the demonstration that compartment-sampled negatives are learnable from four annotation features, which reframes a body of reported PPI benchmark scores as partly measuring dataset construction. Its own limits are stated plainly: the training pairs were not trimmed by homology, so proteins recur across splits even though no pair does; false positives on compartment-sampled data are high enough that the authors read the model as detecting plausible physical contact rather than functional interaction; and the perfect score on random sequences may reflect the shuffled negatives in training. The pairwise design is quadratic in proteome size, the constraint that later generations were built to remove — Synteract-4 and FlashPPI both replace pair scoring with one encoding per protein. Weights are public under a non-commercial license; the preprint has not been peer reviewed.