Biology subjects

Singh, H.

Publications and source records attributed to Singh, H..

At least 19 recordsLinked to original sources

bioRxiv2026

Active Learning framework for Antibody Binding Affinity Prediction

Antibody engineering campaigns must prioritise variants from large combinatorial libraries under limited assay budgets. We present a Gaussian process (GP) active-learning benchmark for antibody-antigen binding affinity prediction across eight public AbBiBench assays spanning influenza hemagglutinin antibodies CR6261 and CR9114 and AAYL-series scFv libraries against the SARS-CoV-2 Spike HR2 peptide. The framework embeds variable heavy and light (VH/VL) sequences with ESM-2 650M, trains GP surrogates over compressed protein-language-model representations, and evaluates kernel, acquisition, representation, classical featurisation, and labelling-budget choices under a unified budget of 10\% of each full pool with a label-blind seed and a fixed held-out evaluation split. Across the benchmark, Tanimoto+UCB has the highest mean Recall@2\% point estimate among tested kernel-acquisition configurations, achieving \textbf{57.8\%} mean Recall@2\% across five random seeds (68.3\% when scored against the acquisition-eligible pool), compared with 9.6\% for random selection -- a gap that is significant by a paired Wilcoxon test across datasets ($p=0.0078$); its point estimate is not significantly different from RQ+UCB or Tanimoto+EI at five seeds. We adopt the continuous Tanimoto kernel of \citet{gessner2024}. The same configuration recovers the largest fractions of top binders in the influenza libraries (74.2\% on \texttt{3gbn\_h1}, 82.7\% on \texttt{3gbn\_h9}, and 84.9\% on \texttt{4fqi\_h3}) while exposing harder AAYL combinatorial landscapes where recall remains lower and more variable. An additional three-seed representation ablation, run under the original protocol and not repeated after the evaluation corrections, ranked ESM-2 650M highest (mean Recall@2\% 0.673 $\pm$ 0.28) among ProtBert, AntiBERTy, AbLang2, ProGen2, and classical encodings. On BindingGYM, the same configuration matches or exceeds ALLM-Ab on two of three datasets using the 100-sample held-out test metric. Structure-mapped Integrated Gradients and KernelSHAP analyses recover CDR-region attribution patterns consistent with contact-derived paratope geometry, though threshold-free (AUROC) analysis shows this qualitative agreement is not matched by strong residue-level discrimination. These results support Tanimoto-kernel GP active learning as a simple, reproducible strategy for label-efficient antibody affinity optimisation with qualitative structural validation.

bioinformatics↗

bioRxiv2026

Do LLM Proposers Improve Budgeted Search for Ionizable Lipids?

Efficient lipid design under limited experimental budgets is a central challenge in computational chemistry and drug discovery. We introduce LEAF (Lipid Evolutionary Adaptive Framework), a matched-budget protocol for testing whether language-model proposers improve combinatorial molecular search relative to classical baselines. The study uses a 30,044-candidate ionizable-lipid space, a frozen Gaussian-process oracle fitted to experimental potency labels, and a separate active-learning surrogate. Four language models evaluated over eight seeds and three splits show no significant improvement over genetic-algorithm or random-search baselines. For Llama 8B, 79.1% of acquisition-selected queries repeat previously evaluated archive members. Removing duplicates reduces the performance gap on some splits but increases proposal cost and can prevent budget exhaustion. Larger models reduce duplication without establishing superiority or equivalence. Exploratory scored prompts and stricter novelty controls yield positive median differences against the genetic algorithm, yet wide intervals and shared-seed dependence preclude stronger conclusions. A validated single-assay reanalysis preserves the chemotype-cold deficit at a smaller numerical scale, and exploratory expected hypervolume improvement does not reverse the result. Follow-up checks address whether proposers genuinely use injected evidence. These checks include a zero-shot leakage probe, rank-only and chain-of-thought-rationale prompts, a score-shuffle perturbation, an evidence-conflict probe, and a genuine reasoning-mode proposer. They show measurable engagement with supplied information without a corresponding performance gain. LEAF isolates proposal quality as the experimental variable and documents a concrete failure mode of zero-shot LLM proposers under controlled budgets.

bioinformatics↗

bioRxiv2026

Probing isn't Causality in Molecular Models

Linear probes establish whether a chemical variable can be recovered from a molecular representation; they do not establish how that variable contributes to subsequent predictions. We replace a concept-associated scalar component at an intermediate block and measure the change through a separately fitted downstream outcome readout. Comparisons with random directions and architecture-matched random encoders distinguish recoverability, intervention selectivity, and a pretraining-associated selectivity advantage. In Uni-Mol, block-12 total-distortion edits predict the observed energy-change sign for 97.6\% of held-out pairs, with Spearman $\rho=0.787$. Structure-clustered intervals for the primary direction and encoder contrasts exclude zero. A stricter 20-direction sensitivity analysis does not survive correction across six cells. A post-review analysis finds substantial concept--outcome alignment; intervention effects persist after removal of the same-block outcome direction, but this does not establish that subsequent encoder computation is necessary. In MolCLR, pretraining improves dipole recovery by $0.080\,R^2$, whereas concept-minus-random sign-accuracy contrasts are $0.029$ and $-0.015$ at the two tested blocks. Their paired uncertainty is unresolved, so they do not establish an absence of selective effects. We propose an operational evidence ladder whose conclusions are conditional on the concept, intervention, control family, and fitted readout.

bioinformatics↗

bioRxiv2026

CLARA: Compound Library Active Learning with Reasoning Agents

Batch active learning for ligand discovery requires a recurring decision about whether to exploit high predicted affinity, reduce model uncertainty, or broaden chemical coverage. We test whether a language model can make this strategy-level decision while candidate ranking and label revelation remain controlled by a deterministic Gaussian-process environment. Across four protein targets, ten paired seeds, and a 360-compound assay budget, a GPT-5.6 Sol agent outperformed random sampling in all 40 matched campaigns. Its final top-2\% recall was 0.412 on D2R, 0.374 on Mpro, 0.448 on TYK2, and 0.683 on USP7. It was statistically tied with the strongest fixed policy on Mpro and USP7 but underperformed target-specific leaders on D2R and TYK2. The agent shifted from exploration-oriented early batches toward exploitation and produced 686 testable threshold claims with an 86.3\% atomic success rate, although 19.8\% of forecasts were not numerically scorable. In 200 prespecified information-ablation campaigns, removing uncertainty, diversity, or memory caused only small endpoint changes, while shuffling explanatory evidence reproduced the full allocation path in 39 of 40 paired campaigns. A deterministic 80\% availability constraint reduced recall on every target but preserved the target-dependent pattern of wins and losses. These results support language-model strategy control as a feasible and auditable layer.

bioinformatics↗

bioRxiv2026

GRASP: Graph Representation Learning with Assay Supervision for Molecular Properties

Molecular structures are abundant, while experimental bioactivity is sparse and distributed across assays. We introduce GRASP, a 93.5M-parameter graph Transformer that learns from these sources in sequence. GRASP first learns molecular structure through replaced-token detection on 1.54 billion ZINC20 molecule presentations, then adapts to sparse bioactivity across 642 ChEMBL assays. Across 23 cluster-held-out OpenADMET endpoints, GRASP achieves a mean MAE of 0.374, compared with 0.383 for CheMeleon, 0.408 for ECFP4-LightGBM, and 0.440 for Chemprop. ChEMBL adaptation reduces mean MAE from 0.407 to 0.374 and improves 21 of 23 endpoints, showing the value of intermediate bioactivity supervision. On 22 TDC ADMET tasks, GRASP gives a better point estimate than CheMeleon on 13 tasks. These results support staged molecular representation learning, where abundant molecular structure and sparse experimental measurements provide complementary supervision at different stages of training.

bioinformatics↗

bioRxiv2026

Interpretable machine learning coupled to gene regulatory networks uncovers subcircuits underlying cell fate decisions

Gene regulatory networks (GRNs) model causal linkages that control cell fate decisions and differentiation transitions. Prioritizing regulatory subnetworks underlying cell state differences is of critical importance, but current methods including those reliant on topological metrics introduce circularity as the metrics prioritizing TFs are computed from the same networks whose assumptions they inherit. Separately, interpretable machine learning methods can identify latent factors (LFs) that discriminate cellular states with formal statistical guarantees but do not model regulatory linkages. Here, we present FOCAL (Factor-Outcome Coupling for Assessment of Linkages), a paradigm to prioritize regulatory subnetworks by coupling state-specific and dynamic GRNs with outcome-supervised LFs learned using interpretable machine learning without reference to network topology. This shifts GRN focus from macroscopic TF nodes to state-specific and dynamic TF-gene linkages. In B and T cells, FOCAL identified GIFs (GRNs coupled to Interpretable latent Factors), prioritized regulatory subnetworks underlying established states as well as transient regulatory episodes preceding them. By coupling LFs learnt from perturbation experiments of lineage-defining TFs, FOCAL identified transcriptional predisposition to alternative fates within progenitor cell populations before overt differentiation. This uncovered a novel NFATC2-IRF8 interplay in activated B cells, that was validated by in-vitro and in-vivo genetic perturbations. The two transcription factors act cooperatively to restrain extrafollicular plasmablast differentiation and promote germinal center B cell fate.

systems biology↗

bioRxiv2026

FlowMap: Geometry Dynamics Consistent Embedding of RNA Velocity for Interpretable Cellular Trajectories

Single-cell RNA sequencing reveals how cells vary across states, but these measurements capture only static snapshots of dynamic biological processes. RNA velocity addresses this limitation by estimating how gene expression is changing over time, providing directional information about cell state transitions. However, current approaches treat cellular state and dynamics separately, leading to representations that violate fundamental geometric consistency and obscure biological interpretation. Here, we present FlowMap, a framework that jointly models cellular states and their dynamics within a unified geometric representation. FlowMap simultaneously reconstructs a smooth low-dimensional manifold of gene expression and constrains RNA velocity to follow its local geometry, producing coherent and denoised representations of cellular trajectories. Across simulated and real datasets, FlowMap recovers interpretable dynamical patterns, including continuous progressions, branching events, cyclic behaviors, and stable-like states. It highlights key genes and gene programs involved in development that play distinct roles in these dynamics, and extends naturally to spatial transcriptomics, where it captures spatially organized developmental processes. Together, FlowMap establishes a principled geometric framework for modeling cellular dynamics, bridging representation learning and dynamical inference in single-cell analysis.

bioinformatics↗

bioRxiv2026

Keratinizing Squamous and Intestinal Metaplasia in Urothelium of Long-term Catheterized Patients with Chronic Urinary Tract Infections

Purpose: Many spinal cord-injured (SCI) patients are recurrently catheterized to manage urinary retention and incontinence. Microbial pathogens colonize urethral catheter surfaces, persist in form of biofilms and trigger inflammatory responses in the bladder. Our objective was to determine bladder tissue changes due to long-term inflammation using LC-MS proteomics. Experimental Design: We examined longitudinal proteomic profiles from eight SCI patients, sampled over 2 to 6 months, in a patient-specific manner as well as compared with trauma patients diagnosed with acute urinary tract infections (UTI). The sample sources were urinary pellets (UP) from all subjects, and catheter biofilm extracts (CB) from SCI patients. Bioinformatic protein network and western blot analyses served to corroborate evidence of protein and cell differentiation changes in the urinary tract of SCI patients. Results: More than 50 proteins with functional or structural roles in the epidermal cornified envelope were quantitatively increased in proteomic profiles from three SCI patients. This included stress response keratins and proteins involved in keratin crosslinking, e.g. transglutaminases and small proline-rich proteins, suggesting occurrence of keratinizing squamous epithelial metaplasia. Protein co-expression network analyses supported epithelial trans-differentiation changes. Nearly thirty proteins with gastro-intestinal epithelial cell expression specificity were enriched in proteomic profiles from a single SCI patient. Among those were mucin-2 and sucrase-isomaltase, secreted proteins detected in samples for only that patient via western blots. This finding was consistent with intestinal epithelial metaplasia. Conclusions and clinical relevance: Without using invasive cystoscopy and histopathology, proteomic analyses from easily available UP and CB samples allowed us to infer cell differentiation changes in urothelial tissues of SCI patients with chronic infections, supporting a concept of clinical relevance for urothelial metaplasia screening, a precursor stage of urothelial neoplasms.

molecular biology↗

bioRxiv2026

Solvent-Isotope Effects in Biomolecular Phase Separation and Fibrillation of Disordered Proteins and Peptides

Heavy water (D2O) is widely used in biomolecular spectroscopy and imaging, often under the assumption that it is an inert replacement for H2O. However, D2O differs subtly in hydrogen-bonding, viscosity, and dielectric properties, which can alter biomolecular interactions and self-assembly. Here, we test how solvent isotope substitution modulates protein/peptide phase separation and amyloid formation in multiple intrinsically disordered systems. Using turbidity-based phase diagrams and microscopy, we quantify how D2O shifts protein-RNA complex coacervation boundaries and alters condensate morphology. Droplet recovery measurements indicate significant solvent-dependent changes in condensate material properties. We further evaluate amyloid formation kinetics, in the presence or absence of a cofactor, supported by orthogonal structural characterization, and assess the functional consequences of tau fibrils using a tau biosensor seeding assay with explicitly defined seed delivery conditions. Together, these results show that D2O can systematically bias liquid-liquid phase separation and aggregation readouts and should be treated as an active experimental variable rather than a neutral solvent substitute.

biophysics↗

bioRxiv2026

Continuous pharyngeal endoderm links external and internal gills

Amphibians develop both external and internal gills during ontogeny, offering an opportunity to investigate the developmental relationship between these positionally distinct respiratory organs. Although internal gills of vertebrates are widely accepted to arise from pharyngeal endoderm, external gills have long been regarded as purely ectodermal outgrowths, obscuring their relationship to other vertebrate gills. Here, we combine histological analysis with direct lineage tracing in the Mexican axolotl (Ambystoma mexicanum) and the African clawed frog (Xenopus laevis) to resolve the embryonic origin of amphibian gills. We show that the external gill develops as a continuous epithelial extension of the pharyngeal endoderm, which forms its basal epithelium and reaches the distal gill tip. In the frog, this extension remains continuous with the epithelium giving rise to the internal gills. Rather than representing separate epithelial structures, external and internal gills therefore arise from a shared epithelial domain of the pharyngeal endoderm. These findings resolve a longstanding question concerning the embryonic origin of amphibian gills and provide a developmental viewpoint for understanding how spatially diverse vertebrate gills can evolve through repeated modification of a conserved endodermal tissue.

developmental biology↗

bioRxiv2026

Extracellular Vesicles Mediate Activation and Trafficking of Splenic Immune Cells to the Heart Post-Myocardial Infarction

BackgroundMyocardial infarction (MI) triggers splenic immune cell trafficking to the heart. Vehicles that carry these signals and mediate this crosstalk are unknown. HypothesisWe hypothesize that extracellular vesicles (EVs) released post-MI mediate splenic immune trafficking to the heart. MethodsMice were treated daily with an EV biogenesis inhibitor (GW4869) or vehicle. Splenic/cardiac immune cells were assessed at 3d while survival, cardiac function, hypertrophy, and fibrosis were evaluated at 8w post-MI. Plasma EVs from 1d MI mice or from the hearts that underwent MI/sham in a Langendorff system induced splenic immune trafficking to the heart within 3d and systolic dysfunction at 8w in naive mice. ResultsGW4869 i) inhibited splenic regression, ii) increased splenic retention of neutrophils, monocytes, dendritic cells (DCs), and CD4 T-cells, iii) decreased cardiac gene expression of pro-inflammatory cytokines/chemokines, and iv) decreased trafficking of immune cells to the hearts at 3d post-MI, and iii) improved systolic function and attenuated hypertrophy at 8w post-MI. MI EVs accumulated in the spleen and promoted egress of matured splenic immune cells upon administration to naive mice. Cardiac pro-inflammatory cytokines/chemokines expression and CCR2+MHC-IIhi infiltrating macrophages, CD11c+ DCs, and CD4+ and CD4+TNF+ T-cell levels were also increased in naive mice at 3d post-injection. Importantly, transfer of MI EVs for 2 days induced systolic dysfunction, cellular hypertrophy, and fibrosis in naive mice at 8 w post-injection. DCs process MI EVs for T-cells activation. ConclusionsEVs mobilize splenic immune cells to the heart post-MI and their inhibition can subdue inflammatory tissue-damage to promote healing post-MI.

immunology↗

bioRxiv2026

Antibody-Antigen Affinity Prediction with Chain-Aware Protein Language Modeling

MotivationAntibody-antigen affinity determines which antibodies advance in therapeutic discovery, repertoire analysis and affinity maturation, but experimental measurements are sparse relative to the scale of sequence libraries. Structure-based predictors can exploit interface geometry when reliable complexes are available, yet early discovery often requires ranking many heavy-light chain pairs against antigens for which no complex structure exists. Existing sequence-based models are scalable, but frequently compress heavy and light chains into a single antibody representation or concatenate antibody and antigen features obscuring the chain-specific and epitope-specific signals that drive binding. ResultsWe present AbAffinity, a sequence-only chain-aware three-stream architecture that maintains heavy chain, light chain and antigen as distinct streams. It integrates frozen ESM-2 embeddings with heavy-chain CDR-focused pooling, heavy-light self-attention, adaptive fusion gating and gated cross-attention, training only a compact interaction module. On the SAAINT-DB benchmark, AbAffinity achieves strong predictive performance under ten-fold cross-validation and maintains robust accuracy on novel antigens. It consistently outperforms recent sequence-based models across external benchmarks including SAbDab, AB-Bind and SKEMPI 2.0. Ablation studies highlight the contributions of chain-specific representations, CDR-focused pooling and the gated interaction pathway. Integrated Gradients attributions recover known paratope and epitope residues at structurally validated interfaces. AbAffinity provides a lightweight, explainable sequence-first framework for antibody triage and prioritisation when structural information is limited or unavailable.

bioinformatics↗

bioRxiv2026

Sequential and coordinated control of human plasma cell differentiation by IRF4 and BLIMP1 utilizing a discriminating ISRE/EICE motif lexicon

The transcription factors IRF4 and BLIMP1 (PRDM1) promote plasma cell (PC) differentiation by repressing B cell identity genes while inducing the unfolded protein response, ER/Golgi biogenesis, and immunoglobulin secretion. How IRF4 and BLIMP1 partition and coordinate their genomic activities during plasma cell differentiation remains unresolved. Using naive human B cells and a stepwise in vitro differentiation system, we performed CRISPR/Cas9 perturbations of IRF4 or PRDM1 in plasma cell precursors followed by single-cell RNA sequencing. Despite their mutually reinforced expression and shared recognition of related IRF-family motifs (ISREs and EICEs), loss of IRF4, but not of BLIMP1, yielded a stunted intermediate with incomplete silencing of B cell identity genes and defective induction of the secretory program. Multiome profiling, base-pair-resolution motif modeling, CUT&RUN, and DNA-binding assays identified non-conserved nucleotides within ISRE/EICE motifs that discriminate IRF4 from BLIMP1 binding. These findings reveal a motif-lexicon-dependent IRF4-BLIMP1 interplay, in which IRF4 first acts independently and then in concert with BLIMP1. This regulatory logic may also underlie programming of additional lymphocyte effector states.

immunology↗

bioRxiv2026

Multi-centre laboratory study to determine discriminating concentrations for broflanilide and isocycloseram resistance monitoring in mosquitoes

1In 2024-2025 a multi-centre study involving seven international laboratories was conducted with the support of the World Health Organization (WHO). The aim of the study was to establish and validate discriminating concentrations (DCs) in WHO bottle bioassays for monitoring susceptibility to broflanilide and isocycloseram in Anopheles gambiae s.s., An. funestus, An. stephensi and Aedes aegypti. The following values are recommended for adoption as DCs for broflanilide: 10 {micro}g/bottle for An. gambiae and Ae. aegypti, 15 {micro}g/bottle for An. funestus, and 25 {micro}g/bottle for An. stephensi. The recommended DCs for isocycloseram are 15 {micro}g/bottle for Ae. aegypti, 30 {micro}g/bottle for An. gambiae, 50 {micro}g/bottle for An. stephensi, and 60 {micro}g/bottle for An. funestus. Based on the experiences of conducting this study, which represents the application of a generic protocol for establishing discriminating concentrations produced by the WHO, technical recommendations are made on the generation and analysis of DC data for insecticides in future.

zoology↗

bioRxiv2026

Interconnecting ADC Structure with Tumor Cell Biology with Multimodal Learning

Antibody-drug conjugates (ADCs) represent a significant advancement in cancer therapy, yet their development remains constrained by high attrition rates driven by an incomplete understanding of how ADC chemical design interconnects with tumor biology. Compounding this challenge, the field has converged on a narrow set of redundant structural components, and current linker-payload systems that do not share a single mechanism of action. Existing drug response prediction frameworks cannot resolve this multidimensional complexity, relying predominantly on genomic inputs while protein-level biology is challenging to integrate. To address this, we developed a multimodal machine learning platform interconnecting ADC structural parameters with tumor cell biology across thousands of curated structure-activity datapoints, including multi-omics profiles from 1,479 human tumor cell lines and protein-level inputs from a unique model (GENCEP) that derives complete proteomic signatures. The model was validated through blinded retrospective evaluation and, critically, large coverage prospective prediction of cytotoxicity across 159 ADC-cell line combinations spanning five antigens, four mechanistically distinct linker-payload systems, and eight tumor types, most with no prior published associated ADC data. Performance surpassed industry benchmarks established for small molecule therapeutic modalities, demonstrating that protein-informed multimodal integrated framework is effective at capturing cytotoxic determinants at scale.

pharmacology and toxicology↗

bioRxiv2026

BK Channels Orchestrate Cardiac Homeostasis Through Mitochondrial Uncoupling Proteins

BK channels, coded by the Kcnma1 gene, integrate voltage and intracellular Ca2+ signals and are recognized for their roles in smooth muscle and neuronal excitability. However, their contribution to baseline cardiac physiology remains poorly defined. Here we uncover a fundamental function for BK channels in maintaining normal cardiac performance, independent of pathological stress. Using non-invasive echocardiography, transcriptional profiling, and mechanistic analyses, we demonstrate that Kcnma1 deletion disrupts ventricular function, and remodels metabolic and stress-response pathways. Transcriptomic profiling revealed selective downregulation of mitochondrial uncoupling proteins (UCPs) and suppression of the PGC-1/FOXO3a axis, without broad loss of oxidative phosphorylation components. Enhancing UCP expression restored cardiac performance, indicating that mitochondrial uncoupling and redox control constitute key downstream effectors of BK signaling. Together, these results identify a physiological role for BK channels in maintaining myocardial function and define a mitochondrial BK-UCP axis, critical for cardiac homeostasis.

cell biology↗

bioRxiv2026

Understanding Substance Dependence: What Differentiates Addictive from Non-Addictive Drugs?

Addiction is a global health challenge, yet the molecular features that distinguish addictive from non-addictive drugs remain incompletely understood at the pathway and circuit levels. Here, we present a systematic computational framework that integrates drug-target binding predictions (provided by FINDSITEcomb2.0) with brain-region-specific protein expression to compare addictive and non-addictive compounds. We analyzed 457 addictive and 1,774 non-addictive blood-brain barrier permeable drugs and mapped their predicted targets and associated pathways onto proteins expressed across 120 addiction-relevant brain regions. This analysis reveals widespread convergence between the two classes (addictive and non-addictive drugs) on shared molecular pathways, accompanied by distinct patterns of target and pathway engagement. Functional annotation of differentially engaged targets highlights biases toward plasticity-associated components for addictive drugs. In contrast, non-addictive drugs interact with both plasticity-associated proteins and proteins within the same molecular complex that have addiction suppression, regulatory, and homeostatic functions. Notably, both target classes co-localize within the same addiction-relevant circuits and form an integrated protein-protein interaction network. Together, these results define a differential engagement landscape that links chemical interactions to pathway-level utilization in the brain, revealing molecular features associated with differences in addiction propensity. Significance StatementAddiction is a global health crisis, yet the molecular features that distinguish addictive from non-addictive drugs remain poorly defined. By systematically comparing drug-target and pathway engagement across shared addiction-relevant brain circuits, this study identifies distinct signatures in which addictive compounds preferentially engage plasticity- and stress-associated targets, while non-addictive compounds engage these as well as addiction suppression, regulatory, transport, and metabolic pathways. These results provide a circuit- and pathway-level framework for interpreting addiction liability and guiding the design of safer therapeutics.

biophysics↗

bioRxiv2026

Explainable protein-protein binding affinity prediction via fine-tuning protein language models

Protein-protein interactions underpin virtually every aspect of cellular life, and the precise quantification of their binding affinity is fundamental to understanding immune recognition, disease mechanisms, and the rational design of therapeutic antibodies. Yet predicting binding affinity at scale remains an unsolved challenge: reliable experimental assays are low-throughput and expensive, while computational methods that depend on three-dimensional complex struc-tures cannot be applied to the vast majority of clinically relevant targets where structural data are absent. Here we present BALM-PPI, a framework that predicts protein-protein binding affinity from amino acid sequence alone. Both proteins are encoded by a protein language model trained on evolutionary sequence data and projected into a shared representational space, where their distance directly reflects binding strength. Fine-tuning this protein language model requires updating fewer than 1% of its parameters, and we show that this targeted adaptation steers the model toward interface-relevant sequence signals rather than spurious background correlations. On a curated benchmark of over 12,000 protein complexes, BALM-PPI matches or exceeds the accuracy of structure-based methods and retains predictive power for proteins with less than 30% sequence identity to the training set. Using only a subset of project-specific assay data, BALM-PPI outperforms a recent method trained on three times the data, suggesting that the model has already encoded the underlying interaction signals and requires only minimal supervision to specialise to a new target. BALM-PPI further provides residue-level attribution maps that pinpoint the amino acid positions driving each affinity prediction, consistently re-covering experimentally validated interaction hotspots across enzyme-inhibitor, signalling, and antibody-antigen systems without any structural input during training. This allows predictions to be cross-validated against structural and mutagenesis evidence, providing a mechanistic basis for candidate shortlisting ahead of experimental follow-up. BALM-PPI is freely accessible via an interactive web server.

bioinformatics↗