Loading…
Protein binding affinity prediction from sequence alone, returning pKd and per-residue interface labels instead of a yes-or-no interaction call.
No providers recorded yet. Browse all providers
A binary interaction model answers one question — are these two proteins likely to touch? — and a researcher almost never stops there. How strongly? At which residues? Worth the cost of an assay? A yes-or-no label cannot order a candidate list by anything except the model's own confidence, and confidence is not affinity.
Synteract-2 keeps the sequence-only premise of Synteract-1 and hangs two quantitative outputs on it. One is a pair-level regression in pKd, the negative logarithm of the dissociation constant, where an error of 1.0 means being off by a factor of ten in binding strength. The other is a per-residue label marking which positions sit at the interface. Both, together with the interaction probability, come from one model reading two amino acid sequences — no structure, no multiple sequence alignment, no complex to fold first. That constraint is the point: a sequence model runs on anything with a sequence, including a de novo design whose predicted structure would be the pipeline's least trustworthy step.
It was built by Synthyra with Jason Gleghorn's lab at the University of Delaware and announced in April 2025 as the second generation of the Synteract line, which runs 1, 2, 4 — the third was internal and never released. Its sibling Synteract-4 took the opposite design decision, collapsing each protein to a single embedding so that an entire proteome can be screened, and returning one number per pair because that is what a search needs. Synteract-2 is the depth end of the same stack: once a screen has narrowed the field, it says how strong and where. Both sit beneath Atlas, Synthyra's productized search layer.
The affinity track extends a BERT-style protein language model through custom parameter-efficient fine-tuning, trained on the affinity dataset assembled for the APPT project; Synthyra's processed copy holds 11,076 sequence pairs with measured dissociation constants. Neither the backbone checkpoint nor a parameter count is published.
On the Haddock benchmark, trimmed to complexes where exactly two chains participate, mean absolute error is 0.723 pKd against 1.023 for APPT, with Spearman correlation 0.935 against 0.855; PPB-Affinity is weakest of the three at 0.545 Spearman. Both evaluation sets are small: Synthyra's published Haddock table holds 31 complexes and Affinity Benchmark v5.5 holds 162 pairs. The margin does not survive the harder set — the 0.30 pKd gap over APPT narrows to 0.11 on the larger, messier Affinity5.5, and Synthyra warns against any ranking established on one dataset. On interface residues the result is parity with a structure model rather than a win over one: F1 0.904 and accuracy 0.922 against Boltz-1's 0.909 and 0.903, with substantially overlapping error bars and Synthyra's own caveat that Boltz-1 may have trained on the PDB structures in the test set, making part of its score recall rather than prediction.
The natural position is between a screen and an assay. A proteome-wide or library-wide search produces a ranked list far longer than any wet-lab budget; Synteract-2 re-ranks it on a quantity the experiment measures, then says which residues to target. For binder design that means ordering generated sequences by predicted strength rather than by a binary score; for interaction biology, proposing interface regions for mutagenesis on pairs with no complex structure.
The concrete deployment is the DSM binder campaign, where targets and their designed partners went through the Synthyra API for affinity prediction. Predictions from a Synteract-2-family model ranked 1.4 million generated candidates — 100,000 per target and template scheme across seven BenchBB targets — and 40 designs went to biolayer interferometry, where 12 of 13 expressed EGFR variants bound and the best reached 861 pM. Synthyra's own reading is candid: the target those predictions rated hardest produced the only clear success, so the ranking narrowed the pool but was not correct in its details — a prioritization layer, not a measurement. Measured affinity also depends on assay conditions, cofactors, conformational state and cellular context, none of which a sequence model observes. Neither weights nor code have been released and there is no preprint; the HuggingFace repository reserved under the name holds a placeholder README and no checkpoint, and access is through the Synthyra API.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.