Frozen foundation-model embeddings are position as a strong, data-efficient default for affinity ranking in antibody engineering and establish a conservative lower bound that task-adaptive fine-tuning is expected to exceed.
Abstract
We investigate whether representations from AINN-P1—a protein foundation model trained autoregressively on tens of millions of natural protein sequences—transfer to the task of ranking antibody–antigen pairs by binding affinity. Casting affinity maturation as a learning-to-rank problem over the change in binding free energy (ΔΔG), we compare a task-specific sequence model trained end-to-end from scratch against lightweight downstream heads built on top of frozen AINN-P1 embeddings, all evaluated under an identical five-fold cross-validation protocol. A regularized linear probe on the frozen embeddings already surpasses the from-scratch baseline, and an optimized lightweight head raises the mean Spearman rank correlation from 0.42 to 0.53—a relative improvement of approximately 28%— while training in seconds and without any fine-tuning of the foundation model. Because a linear probe alone exceeds the fully trained end-to-end baseline, the gain is attributable to representation quality rather than to added downstream-model capacity. These results position frozen foundation-model embeddings as a strong, data-efficient default for affinity ranking in antibody engineering and establish a conservative lower bound that task-adaptive fine-tuning is expected to exceed.
A many-objective NSGA-III search is built that attaches a compact head to a frozen ESM-2 t33 backbone and jointly optimizes the head architecture, a physicochemical feature mask, and a contrastive-pretraining curriculum against a fitness vector that explicitly contains the cold-start gap, and adds distribution-free calibrated selective prediction via split conformal, turning a frozen-ESM classifier into a deployable cold-start predictor with finite-sample coverage.
Muhammed Tekin, Murat Gök· IEEE Access· 0 citations
Abstract Motivation To enable real-world protein-ligand affinity prediction, not only out-of-distribution generalization but also robustness to variable structural availability and quality should be considered in model design. Results We present AlignNet, a hierarchical representation alignment framework that mitigates intra- and inter-molecular heterogeneity to learn robust protein-ligand embeddings for generalizable affinity prediction, even from sequence-level inputs. Its intra-molecular module projects unimodal and multimodal features into a unified space, aligning augmented multimodal views for feature fusion and unimodal with multimodal embeddings to distill multimodal priors for structure-agnostic inference. Its inter-molecular module aligns protein and ligand embeddings for cross-molecular integration. Extensive experiments show that AlignNet (i) achieves highly competitive performance, with up to a 20.4% gain in SCC on the challenging LBA 30% split under sequence-only settings, suggesting improved out-of-distribution generalization; and (ii) learns well-separated affinity-related clusters, supporting reliable structure-independent prediction. Availability and implementation AlignNet is available at https://github.com/altriavin/AlignNet.
Xiaowen Hu, Hongyi Huang, Hao Sun et al.· Bioinform.· 0 citations
This work designs a biological prior-guided feature fusion framework that integrates pseudo-structural epitope knowledge and CDR-specific attention mechanisms via a mixture-of-experts architecture to effectively capture complex binding landscapes in antibody screening and drug residence time analysis.
G. Luo, Junkai Wang, Sizhe Zhang et al.· Bioinformatics· 0 citations
Abstract Motivation T-cell receptors (TCRs) recognize antigenic peptides presented by major histocompatibility complex (MHC) molecules and are central to adaptive immunity. Computational prediction of TCR–epitope binding (TEB) can accelerate immunotherapy development, yet remains hampered by limited labeled data, false-negative noise in unobserved pairs, and over-smoothing in graph-based models. Results We present DynaTCR, a dynamic graph ensemble learning framework for TEB prediction. DynaTCR encodes TCR and epitope sequences with protein language model embeddings and organizes them into a bipartite interaction graph. A graph regularization-variance-preserving aggregation (GR-VPA) encoder stabilizes message propagation and alleviates over-smoothing, while a global attention layer captures long-range dependencies. Multiple base learners are trained with iteratively updated hard-negative samples to reduce false-negative predictions. Under the StrictTCR evaluation protocol on four public datasets, DynaTCR achieves AUC improvements of 4.0–8.2 percentage points over the strongest existing method and up to 15.8 percentage points in AUPR. On the most stringently curated dataset, DynaTCR attains an AUC of 95.1%. Furthermore, on an independent structure-derived test set, DynaTCR achieves the highest AUC (72.6%) among all compared methods, demonstrating its robustness and effectiveness for TEB prediction and candidate prioritization. Availability Source code and data can be downloaded from: https://github.com/2014402680/TEB/.
Mutation-induced changes in binding free energy (ΔΔG) at antibody–antigen interfaces are important for antibody optimization, mutational scanning, and viral immune escape assessment. However, computational prediction remains challenging because antibodies and antigens have distinct sequence backgrounds, mutation effects are often localized at interfaces, and related complexes may remain across training and evaluation partitions. We present AbAgMut-GNN as a task-oriented paired graph framework that coordinates established sequence and geometric learning components around explicit comparison of wild-type (WT) and mutant (MUT) antibody–antigen complexes. AntiBERTy and ESM2 provide frozen residue-level embeddings for antibody and antigen chains, respectively, while mutation-centered, interface-aware, paired-residue, and contact-delta representations capture local perturbations and interaction remodeling. We evaluate AbAgMut-GNN under four complementary internal settings, including the PDB-based split, the complex-cluster split, the antibody-family preserving validation split, and the antigen-cluster-preserving validation split. Under the complex-cluster split, AbAgMut-GNN achieves Pearson correlation coefficients of 0.5841 on AB-Bind and 0.5480 on SKEMPI v2.0. External validation on SARS-CoV-2 and influenza antibody–antigen systems further shows useful mutation-effect correlation trends, although absolute-error performance varies across target systems. Contact-masking and residue-class enrichment analyses indicate that model-derived importance patterns are associated with biologically relevant interface interactions. Overall, AbAgMut-GNN is best viewed as a task-oriented computational tool for trend-level mutation ranking and pre-experimental candidate prioritization rather than as a high-precision substitute for quantitative biophysical measurement.
Wenchi Ge, Qi-Jia Yu, Jincen Shuai et al.· Journal of Chemical Informat...· 0 citations
It is shown that PLM embeddings encode patterns correlated with biochemical properties and quantify their contribution to predicting protein fitness, and this technique is readily transferable to problem settings beyond protein fitness prediction.
Paulo Yanez Sarmiento, Pia Francesca Rissom, Manuel Pfeuffer et al.· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
