eSIG-Net: an interaction language model that decodes the protein code of single mutationsDownload PDF Download PDF Brief CommunicationOpen accessPublished: 29 April 2026Xingxin Pan ORCID: orcid.org/0000-0002-4429-48781,2,3 na1,Aditya Shrawat4 na1,Sidharth Raghavan ORCID: orcid.org/0009-0000-6891-56271,2 na1,Chuanpeng Dong5,6,Yuntao Yang7,Zhao Li ORCID: orcid.org/0000-0002-6154-98797,W. Jim Zheng ORCID: orcid.org/0000-0001-7411-60477,S. Gail Eckhardt8,Erxi Wu ORCID: orcid.org/0000-0002-1680-36391,2,3,9,10,11,Juan I. Fuxman Bass ORCID: orcid.org/0000-0001-9457-120712,13,Daniel F. Jarosz ORCID: orcid.org/0000-0003-3497-588814,Sidi Chen ORCID: orcid.org/0000-0002-3819-50055,6,Daniel J. McGrail ORCID: orcid.org/0000-0002-6669-606915,16,Gloria M. Sheynkman ORCID: orcid.org/0000-0002-4223-994717,Jason H. Huang ORCID: orcid.org/0000-0002-4426-01681,2,Nidhi Sahni ORCID: orcid.org/0000-0002-9155-588218,19 &…S. Stephen Yi ORCID: orcid.org/0000-0003-0047-81031,2,3,8,11 Nature Methods (2026)Cite this articleSubjectsComputational modelsProgramming languageProtein function predictionsProteome informaticsSequence annotationAbstractMost proteins act through interactions with other molecules, yet predicting how single mutations perturb these interactions—defined as ‘protein codes’—remains a central challenge in computational biology. Here we introduce eSIG-Net, the edgetic mutation sequence-based interaction grammar network, a language model that integrates protein sequence embeddings with syntax-aware and evolution-aware mutation encoding and contrastive learning to predict mutation-driven interaction changes. eSIG-Net outperforms state-of-the-art sequence-based and structure-based methods, nominates causal variants and provides mechanistic insights. Together, eSIG-Net is a mutation-centric interaction language model that accurately predicts interaction-specific network rewiring from sequence information alone and generalizes across biological contexts.MainSubstantial improvements in genome and exome sequencing technology in the past 15 years have identified a surfeit of human genetic variation orders of magnitude more extensive than what was previously appreciated. However, how most variants influence the molecular properties and functions of molecules they encode, as well as their impacts on disease initiation and progression remain largely unknown1. Among these genetic variants, missense variants are the most common type of protein-coding mutations. Even single missense variants can drastically change protein–protein interactions or PPIs2,3 (referred to as ‘protein code’ for interactions), and therefore rewire protein signaling4. Similar to the ‘activity cliff’5 problem in chemistry machine learning, where small structural changes often lead to large or unpredictable changes in activity, single mutations pose an ‘interaction cliff’ grand challenge, causing computational models to mispredict mutation-mediated PPIs (Supplementary Note 1).Applying protein language models is a potential solution to these limitations and has been implemented in methods such as ESM1b6, ESM-27, ProtT58, ESM39, D-SCRIPT10 and AlphaMissense11. However, these methods also face at least two substantial challenges. First, they do not explicitly learn the sequence distinctions between mutant proteins and their corresponding wild-type (WT) counterparts. Second, they fail to capture the inherent complexity of PPIs, which are critical for PPI-related tasks.Here we introduce a mutation-centric interaction language model named eSIG-Net (edgetic mutation sequence-based interaction grammar network). In contrast to conventional PPI prediction methods (Extended Data Fig. 1a), eSIG-Net focuses on the discrepancy between WT and mutant proteins, as well as their PPI profiles with a specific interaction partner. As shown in Fig. 1a, the framework of eSIG-Net consists of two encoder modules: (1) the first is a PPI ‘protein encoder’ module, which is commonly employed in classical PPI prediction tasks. It typically involves separately obtaining the encodings of a protein and its interactor and then merging them to predict PPIs. By constrast, in our PPI perturbation prediction pipeline, we obtained the merged encodings of both the WT with its interactor and the mutant with its interactor. These merged encodings were then fed into a constrained discrepancy module to attempt to discern the differences between them. (2) The second module is a mutant ‘protein language model’ encoder. Extensive empirical evidence has demonstrated that leveraging protein language models could capture the evolutionary information of proteins, thereby facilitating various downstream protein-related tasks. To accentuate the differences between mutant and WT proteins, we exclusively used the residue-level embeddings of the mutation sites. This was processed through channel-wise learning to obtain a merged mutation-site encoding (Fig. 1a). Finally, the two merged encodings were integrated and fed into a discriminator for the prediction of PPI perturbations. Compared with the conventional PPI prediction methods, eSIG-Net thus uses an innovative discrepancy strategy to effectively discern the effects of single amino acid changes on proteins and predict ensuing perturbations in their interaction profiles.Fig. 1: Overview of the eSIG-Net model, and benchmarking with state-of-the-art sequence-based prediction methods.The alternative text for this image may have been generated using AI.Full size imagea, The prediction framework of eSIG-Net: WT and mutant sequences are processed by a protein language model to obtain residue-level embeddings. These embeddings are then merged through mutation encoding and passed through a channel-wise mutation-site learning module. Concurrently, WT–interactor PPI and mutant–interactor PPI pairs are encoded by protein encoder, and their merged encodings are used for both constrained discrepancy assessment and traditional PPI prediction. Finally, the combined encodings—mutation-site, WT–interactor PPI and mutant–interactor PPI—are input into a discriminator to predict potential PPI perturbations caused by the mutation. b, Prediction accuracy comparison for the disease mutation PPI dataset, showcasing the performance of eSIG-Net against other PPI prediction models, with statistical significance denoted by asterisks. Error bars denote the standard deviations (n = 1,633 PPIs for each prediction model plotted). When compared to eSIG-Net, P = 1.3 × 10−6 for DeepFE, P = 7.9 × 10−7 for D-SCRIPT, PIPR, PLM-interact and SDNN. c, ROC curves for the disease mutation dataset, comparing the AUC metrics for eSIG-Net and other models, highlighting eSIG-Net’s superior performance. d, Precision–recall curves for the disease mutation PPI dataset, with eSIG-Net outperforming other models in terms of both precision and recall. Line color scheme is the same as c. e, Prediction accuracy comparison for the gnomAD–ExAC population variant PPI dataset, with eSIG-Net achieving the highest accuracy. Error bars denote the standard deviations (n = 4,020 PPIs for eSIG-Net, DeepFE, PIPR, PLM-interact, SDNN; n = 4,002 PPIs for D-SCRIPT). When compared to eSIG-Net, the P = 3.8 × 10−4 for DeepFE, P = 3.6 × 10−4 for D-SCRIPT, P = 3.6 × 10−4 for PIPR, P = 2.6 × 10−4 for PLM-interact and P = 8.7 × 10−4 for SDNN. f, ROC curves for the population variant PPI dataset, with AUC values for each model, indicating that eSIG-Net maintains a high performance on this dataset as well. g, Precision–recall curves for the ExAC population variant PPI dataset, detailing the precision and recall performance of each model, with eSIG-Net providing a competitive precision–recall balance. Line color scheme is the same as f. P values are calculated by two-sided paired t-tests, with Holm–Bonferroni correction. ***P 0.5; right). d, Density distribution plot of pDockQ scores with correct (green) and incorrect (red) cases, for WT protein-mediated interactions. e, Density distribution plot of pDockQ scores with correct (green) and incorrect (red) cases, for mutant protein-mediated interactions. f–h, Prediction evaluation using mutation-centric, structure-based methods. The disease mutation PPI dataset is used here. f, Bar chart summarizing the mean accuracy of different structural algorithms, with the length of each bar representing the mean accuracy and the error bars denoting the standard deviations (n = 1,633 PPIs for eSIG-Net, BeAtMuSiC, GeoPPI, PIONEER; n = 1,612 PPIs for MutaBind2; n = 1,157 PPIs for TopNetTree). When compared to eSIG-Net, the P = 1.3 × 10−8 for BeAtMuSiC, P = 2.3 × 10−7 for GeoPPI, P = 2.7 × 10−7 for PIONEER, P = 9.5 × 10−8 for MutaBind2 and P = 2.3 × 10−7 for TopNetTree. g, ROC curves displaying the comparative AUC values for various structural algorithms. Shading indicates standard deviations. h, Precision–recall curves for structural algorithms. Shading indicates standard deviations. The centers for the error bands indicate mean true or false positive rate (in g), and mean precision or recall (in h), respectively. i, Bar chart showing the mean accuracy of ESM variants (cyan; a mutation-centric disease-causing prediction tool), compared with eSIG-Net (blue). Bar length represents the mean accuracy and error bars denote the standard deviations. Dashed line indicates a random classifier (n = 1,027 ‘Perturbed PPIs’; n = 606 ‘Nonperturbed PPIs’). j, eSIG-Net-predicted interaction profiles of two disease mutations in the pleiotropic gene TPM3. k, eSIG-Net-predicted interaction profiles of a disease mutation and a population variant in the gene COQ8A (also known as ADCK3). P values are calculated by two-sided paired t-tests, with Holm–Bonferroni correction. ***P 0.5 indicated the presence of an interaction according to the baseline method’s definition. We defined PPI state change to include both loss of interaction (pWT ≥ 0.5 and pMut