X-PAIR

Whether two proteins interact and where they touch are two faces of the same biological question, but computational methods have largely treated them as separate problems: one family of tools scores pairs for interaction, another annotates interface residues on a single protein without reference to the partner. X-PAIR is a sequence-based multitask framework that, for a pair of protein sequences, predicts both whether they interact and which residues form their partner-specific interface — the residues used with that partner, not a generic binding propensity.

X-PAIR was developed by Sara Rescalli and Alessandra Carbone at Sorbonne University and is described in a bioRxiv preprint. It combines representations from a pretrained protein language model with a lightweight cross-attention module that lets the two sequences condition one another, and it requires neither structural templates nor multiple-sequence alignments. That places it alongside a growing set of sequence-only interaction predictors built on language-model embeddings, such as PPLM, PUMBA, FlashPPI, and PLMDA-PPI, and apart from structure-based approaches like AlphaFold-Multimer whose per-pair cost rules out proteome-wide screens.

The distinguishing result is that coupling the two tasks costs almost nothing. X-PAIR returns both outputs at close to single-task compute, and analyzes one million protein pairs in under two hours — roughly 500-fold faster than existing approaches for interface prediction and about 20-fold faster for interaction prediction — which is what moves residue-level interface analysis from case studies to whole interactomes.

#Key Features

• Joint interaction and interface prediction: A single multitask model answers both questions for a protein pair, and the shared training preserves single-task accuracy rather than trading one objective off against the other.
• Partner-specific interfaces: Interface residues are predicted conditional on the specific partner, capturing that many proteins engage different surfaces with different binders.
• No alignments or structures: Inference runs from amino acid sequence alone, with no MSA construction and no template or predicted structure, removing the two most expensive steps in conventional pipelines.
• Proteome-scale throughput: One million pairs are processed in under two hours, with reported speedups of approximately 500-fold for interface prediction and 20-fold for interaction prediction over current methods.
• Leakage-controlled evaluation: Benchmarks are constructed to control for train/test similarity leakage, and X-PAIR outperforms existing methods on both tasks, with the largest gains in interface localization.

#Technical Details

X-PAIR pairs protein language model representations with a lightweight cross-attention block that exchanges information between the two input sequences, producing both an interaction score for the pair and per-residue interface predictions on each chain. Because the cross-attention module is small and no alignment or folding step is involved, multitask inference costs little more than a single task — which is what enables the million-pair, sub-two-hour throughput and the roughly 500-fold (interface) and 20-fold (interaction) speedups. On leakage-controlled splits, X-PAIR improves on existing methods for both tasks, with substantial gains on interface localization. Cross-species analyses expose different evolutionary dependencies for the two: interaction prediction benefits from multispecies training data, whereas interface localization stays robust across taxonomic scales. The work is a preprint released under CC BY-NC, a license that covers the manuscript; no public code repository or model weights accompany it.

#Applications

The primary use is proteome-scale interactome reconstruction: scanning large pair sets to propose candidate interactions and, for each one, the residues that mediate it. That combination is directly actionable in the lab — a predicted interface nominates mutations for alanine scanning or interface-disruption experiments, and prioritizes which pairs are worth committing to co-immunoprecipitation, two-hybrid, or structural follow-up. Because it needs only sequence, X-PAIR applies to organisms where MSAs are shallow or structures are unavailable, and its cross-species behavior supports comparative interactome work.

#Impact

X-PAIR argues that interaction detection and interface localization are better learned together than bolted together, and it cuts the cost of residue-level interface annotation by orders of magnitude — the regime where whole-interactome, residue-resolved analysis becomes tractable. Its leakage-controlled benchmarking is a welcome methodological choice in a subfield where optimistic splits have historically inflated reported accuracy. Two caveats bound its current standing: the work has not completed peer review, and with no code or weights released, independent reproduction and third-party benchmarking are not yet possible.