Loading…
Sequence-only protein-protein interaction prediction at proteome scale. Each protein is embedded once, so a pair score compares two stored vectors.
No providers recorded yet. Browse all providers
A protein interaction predictor is normally a classifier over pairs: hand it protein A and protein B together and it returns a probability. That fits a benchmark, whose test set is a list of pairs somebody wrote down. It fits badly the question a biologist asks, which is rarely about one pair — "what does this protein touch?" ranges over a proteome, and a proteome is quadratic. Scoring every pair among the 19,226 human reference proteins means roughly 185 million forward passes.
Synteract-4 changes what the model produces. Instead of a verdict about a pair it emits one vector per protein, and the interaction score is a comparison between two such vectors. Each protein is encoded once whether it appears in one candidate pair or twenty thousand, so the same human screen costs 19,226 encoder passes and then arithmetic. Amino acid sequence is the only input — no structures, no genomic context, no per-species tuning at evaluation time — and the system trains and runs on a single GPU.
Built by Synthyra with Jason Gleghorn's lab at the University of Delaware, it is the fourth generation of the Synteract line: the first asked whether a protein language model could learn the interaction objective at all, the second added affinity and interface-residue prediction, and the third was never released. Atlas is the productized successor, retrained on a larger encoder and extended with ligand and annotation scoring.
On the Bernett gold-standard split, which partitions proteins so no sequence above 40 percent identity crosses between train, validation and test, Synteract-4 reports a Matthews correlation coefficient of 0.34 against 0.30 for the next best published method, with 0.73 ROC AUC and 0.67 F1 — a benchmark most prior methods do not clear at 0.70. The proteome-wide comparison is against ProteomeLM, which trains a logistic-regression head on each evaluation proteome and uses orthology information Synteract-4 never sees: 0.968 against 0.830 on the full human evaluation at STRING combined score 990 or above, and 0.897 against 0.898 as a mean over nineteen bacterial pathogens, leading on ten of them. On the strict zero-shot subset — 43.5 of 184.8 million human pairs, both proteins outside every training cluster — it reaches 0.953. Cross-species is thinner: on a homology-controlled human–SARS-CoV-2 surface a human-only variant reaches 0.58 balanced ROC AUC and a multi-species variant with a virus-side homology cut reaches 0.601, where a prior audit placed seven published human-trained methods at chance. The production checkpoint's 0.767 is an upper bound: its training distribution was never given a SARS-CoV-2 homology cut.
The cost structure decides which screens are thinkable. One expensive pass per pair confines a method to a curated shortlist; one pass per protein runs on every pair in a proteome, and then across two. That supports a neighborhood query around one protein, a full intra-actome map for an organism, and a host-pathogen inter-actome screen — searches rather than single predictions. For a lab with a target and no partner list, the output ranks where to spend the expensive experiment.
The strongest evidence sits outside the databases. Cardiac HSP90-beta partners were measured by immunoprecipitation mass spectrometry across 24 human left-ventricular samples, yielding 92 partners, 29 of which appear in none of the Picard, STRING or BioGRID partner sets. At a threshold calibrated so exactly 20 percent of the human proteome scores positive, the model calls 15 of those 29, a rate of 0.517 against a 0.200 background; the misses are dominated by immunoglobulin chains, keratins and mitochondrial enzymes, classes an immunoprecipitation pulls down for reasons other than direct binding. The limits are stated plainly: each variant was trained with a single seed, so run-to-run spread is unbounded; full-surface precision-recall AUC on the controlled cross-species variant is 0.038; and in the unfiltered top 702 Escherichia coli predictions, 95.6 percent are already-known physical interactions and 0.3 percent novel — a model rediscovering the textbook. The diversity-controlled retrieval pipeline used by FlashPPI shifts that mix to 20.4 percent physical, 42.3 percent functional and 37.3 percent novel, where "novel" means absent from every STRING evidence channel, a statement about the database rather than the biology. Synthyra has not released Synteract-4 weights and the manuscript is preparing for journal submission; the capability reaches users through Atlas.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.