Protein-protein interaction language model that embeds whole sets of interacting chains through cross-chain attention, not one sequence at a time.
No providers recorded yet. Browse all providers
Protein language models learn from sequences in isolation, a poor match for the biology that matters most in disease: proteins act on each other. Practitioners work around this by concatenating two chains into one long string, or by embedding each partner separately and gluing the vectors together. Both distort the problem. Concatenation invents a covalent link that does not exist and lets positional encodings bleed across the interface; separate embedding discards the inter-residue context that determines whether the two proteins bind at all.
MINT (Multimeric INteraction Transformer) makes the set of interacting chains the native unit of modeling. Developed by Varun Ullanat, Bowen Jing, Samuel Sledzieski, and Bonnie Berger at MIT CSAIL, it takes ESM-2 at the 650-million parameter scale as its starting point and adds cross-chain attention blocks alongside the existing within-chain self-attention. Each chain keeps its own rotary positional encoding for intra-sequence relationships; the cross-chain blocks operate without rotary encoding, so a residue in one chain can attend to residues in another without any implied ordering between them. Attention and embedding weights are warm-started from ESM-2, preserving general sequence knowledge while the model specializes.
The training signal comes from interactions rather than from a protein database. MINT is pretrained on 96 million curated protein-protein interactions from STRING, using a masked language modeling objective in which token prediction is conditioned on the partner chain as well as the local sequence — a formulation that surfaces the co-evolutionary constraints binding partners impose on each other. The preprint appeared in March 2025 and the peer-reviewed version in Nature Communications.
The STRING starting corpus of 2.4 billion physical interactions across 59.3 million sequences was clustered with MMseqs2 at 50% identity into 15.6 million clusters, then filtered to one interaction per cluster pair, leaving 382 million. After holding out 250,000 interactions for validation with no cluster shared across the split, training used 95.8 million interactions over 16.4 million unique sequences. Feature dimension is 1,280, matching ESM-2 650M; 15% of tokens are masked with the standard 80/10/10 replacement scheme; optimization used Adam with weight decay 0.01, a peak learning rate of 4e-4 warmed up over 2,000 steps, and an effective batch size of 64, distributed across NVIDIA A100 80GB and RTX A6000 GPUs.
On the Bernett et al. gold-standard PPI dataset — designed to remove the data leakage that inflates older PPI benchmarks — MINT reaches an AUPRC of 0.69. On SKEMPI v2 it improves prediction of binding affinity changes upon mutation by 29% over the strongest sequence-only baseline, and on the FLAb antibody fitness benchmark it exceeds antibody-specific language models by more than 10%. TCR-epitope binding is the narrowest margin: an average AUROC of 0.581 against 0.576 for the best specialized baseline. Applied to SARS-CoV-2 antibodies, MINT identified antibodies with consistent neutralization across Omicron sub-variants at roughly 80% accuracy.
MINT serves teams that need to score many interactions cheaply. Antibody engineers can rank affinity-maturation variants and estimate breadth against emerging viral lineages before committing to expression and binding assays; cancer researchers can triage variants of uncertain significance by asking whether a substitution disrupts an oncogenic interface; immunologists can screen TCR-epitope pairs. Because inference costs a forward pass rather than a structure prediction, proteome-scale screens are practical, and the authors recommend parameter-efficient fine-tuning for protein classes such as peptides that STRING covers thinly.
MINT establishes that the gap between protein language models and interaction modeling is architectural rather than a matter of scale: adding cross-chain attention to a 650M backbone outperforms 3B-parameter single-sequence models across binary interaction classification, affinity prediction, and mutational effects. Its limits are the ones its authors name — pretraining saw only pairs, performance inherits STRING's coverage biases, and the model reads sequence alone, where structure-aware approaches such as SaProt and ESM-3 suggest a complementary signal remains untapped. Code is MIT-licensed and the pretrained checkpoint is public, so the model can be used as a drop-in embedder wherever a project currently calls ESM-2 on interacting chains one at a time.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.