Seq2Bind

Seq2Bind identifies which residues drive a protein-protein interaction using only the two amino acid sequences involved. Most tools that rank interface "hotspots" depend on an experimental or predicted complex structure, which restricts them to well-behaved, folded partners and makes screening slow. By moving the problem entirely into sequence space, Seq2Bind sidesteps that requirement and can be applied to intrinsically disordered proteins and to complexes for which no structure exists.

The method was developed by Xiang Ma, Supantha Dey, and colleagues in the Chowdhury Lab at Iowa State University, posted as a preprint in June 2025 and published in NAR Genomics and Bioinformatics later that year. Rather than proposing a new architecture, the work asks how far existing protein language models can be pushed on a quantitative biophysical task: four backbones — ProtBERT and ProtT5 from the ProtTrans family, ESM-2 at 650M parameters, and a BiLSTM baseline — are fine-tuned once on experimental binding affinities, then applied unchanged to new complexes.

Hotspots are read out indirectly. The fine-tuned regressor predicts binding free energy for a sequence pair, and an in-silico alanine scan mutates each residue in turn; residues whose substitution is predicted to be most destabilizing are ranked as interface candidates. The resulting tool is exposed as a free webserver in the StructF suite alongside a Colab notebook.

#Key Features

• No structural input: Predictions run from raw sequence, so complexes with no solved or confidently predicted structure — including disordered regions — remain in scope.
• Alanine-scanning hotspot readout: Every residue is mutated to alanine (and, optionally, in 5- or 10-residue patches) and ranked by predicted destabilization, turning an affinity regressor into a residue-level interface predictor.
• Four benchmarked backbones: ProtBERT, ProtT5-XL, ESM-2 650M, and a BiLSTM are each fine-tuned under the same protocol, making the contribution of the pretrained representation directly measurable.
• Screening-speed inference: On an A100 GPU the models process roughly 46 (ProtBERT), 10.9 (ESM-2), and 5.0 (ProtT5) sequences per second, so a full scan takes minutes rather than the hours a docking run requires.
• Free webserver and notebook: Both binding-energy prediction and alanine scanning are available without installation, with all four models selectable.

#Technical Details

Each model is wrapped in a Siamese configuration: the two partner sequences pass through weight-shared encoders producing 1024-dimensional per-residue embeddings, which feed a regression head of stacked dropout and linear layers with ReLU activation. Training used SKEMPI 2.0, filtered from 7,086 complex pairs down to 5,387 usable pairs (3,770 train / 1,617 validation) with binding affinity expressed as negative ΔG derived from dissociation constants. ProtBERT converged in five epochs and ProtT5 in two. Evaluation used an N-factor metric that asks whether the n true interface residues of a complex appear within the top N×n ranked predictions. Across 6,063 dimers from the Protein Data Bank — some 7.6 million per-residue predictions — ProtBERT recovered 68.2% and ESM-2 67.4% of interface residues at N-factor = 3. On an independent panel of 14 human-health-relevant complexes, ESM-2 reached 37.2% and ProtBERT 35.1% at N-factor = 2, ahead of the structural docking program HADDOCK3 at 32.1%, and ESM-2 recovered 50.5% at N-factor = 3.

#Applications

The intended use is as a fast prior rather than a final answer: Seq2Bind narrows a protein pair down to a short list of candidate interface residues, which can then seed restraints for blind docking, focus a mutagenesis campaign, or triage therapeutic targets before committing to structural work. Groups studying disordered or poorly characterized proteins benefit most, since these are precisely the cases where docking and structure-based hotspot tools have no usable input. The webserver makes the workflow accessible to bench scientists without a GPU or a modeling background.

#Impact

Seq2Bind adds quantitative evidence to an ongoing question in protein modeling — how much interaction information is recoverable from sequence representations alone — and shows that general-purpose language models, fine-tuned on a few thousand affinity measurements, can match or exceed a structural docking baseline at ranking interface residues. The honest caveat is scale: absolute recovery rates in the 30–50% range on the independent therapeutic panel mean the output is a ranked hypothesis set, not a determination, and performance on the large PDB dimer set is substantially higher than on the harder held-out complexes. Reproduction materials are limited to a Colab notebook and the hosted webserver; no standalone repository or downloadable fine-tuned checkpoint has been released, which constrains offline and high-throughput use.