Computational Antibody Papers
Filter by tags
All
Filter by published year
All
TitleKey points
- Novel dataset of 160 VHH-Fc profiled across 10 biophysical assays.
- They demonstrated that tabular neural networks (TabICLv2, TabPFN v2.5) trained solely on 559 IgG heavy chains outperform intra-format VHH-Fc models in zero-shot predictions, showing that training data scale outweighs structural divergence between scaffolds.
- Surface-driven properties transfer with high accuracy, heparin binding (rho=.82), hydrophobicity (HIC, rho=.63), and self-association (AC-SINS, rho=0.62), whereas thermostability (Tm2, rho=0.16) remains scaffold-dependent and requires direct measurement.
- Augmenting models with simple surface property inputs (HIC and HAC) boosts prediction accuracy for complex liabilities like polyreactivity (PR-CHO, rho = 0.40 to 0.51).
- Evaluated ten co-folding protocols on a benchmark of 412 human monomeric antigen complexes
- They demonstrate that recent architectures like Protenix v2 (52% top-1 success on post-cutoff Fv complexes) substantially outperform earlier methods like AlphaFold-Multimer (20%) and Protenix v0.5, which serves as the open-source AlphaFold3 reproduction baseline.
- Modeling accuracy scaled inversely with CDR-H3 loop length. Short loops (<11 residues) were predicted accurately across all methods (0.4–0.7 Å Calpha RMSD), whereas long loops (>16 residues) remained challenging but showed distinct improvement in Protenix v2 (median 2.67 Å RMSD vs. 3.22–3.88 Å in older methods).
- CDR-H3 accuracy was identified as the primary structural feature distinguishing successful complex predictions DockQ \ge 0.49 from failures, demonstrating a strong inverse correlation with overall DockQ scores (Spearman rho = -0.74) and high metric discrimination (AUROC = 0.91).
- Prediction failures in unconstrained models were dominated by sampling limitations (failure to generate native-like poses) rather than ipTM ranking failures; providing idealized epitope constraints or increasing seed depth successfully rescued many of these failures by guiding the search space.
- Training specificity models across several domains.
- Models were trained strictly on public 1D sequence data (amino acids, DNA/RNA nucleotides, and chemical SMILES) across six biological domains, requiring no 3D structural information.
- Base sequence encoders were kept frozen while small 5–7 million parameter projection heads were trained in under an hour per fold on a single GPU using thermodynamic contrastive learning.
- Development was driven entirely via natural-language prompts by a domain expert with zero coding experience, with all numerical claims verified by an independent AI auditor.
- Top-1 target retrieval accuracy reached up to 98.0% from pools of 512 candidates, successfully generalizing to unseen rare HLA alleles and non-canonical binding targets that rule-based tools miss.
- Used as re-ranking filters alongside existing computational workflows, the models dramatically boosted precision, such as raising CRISPR off-target prediction precision from 33.2% to 94.0%.
2026-08-28
Aligning protein-generative models to experimental fitness with ProteinDPO
- developability
- language models
- Novel way (ProteinDPO) to apply pre-trained models to biophysical readouts.
- ProteinDPO is the first framework to apply Direct Preference Optimization (introducing a novel scalar-weighted DPO objective) to align protein generative models with experimental biophysical data, without overfitting like traditional supervised fine-tuning.
- Despite training exclusively on small monomeric stabilities, ProteinDPO generalizes zero-shot to accurately rank the thermal melting temperatures of multichain antibodies and score antibody-antigen binding affinities.
- Applied to H5N1 influenza hemagglutinin, it generated stabilized variants yielding up to a 32C improvement in thermal stability while retaining strong nanomolar binding affinity to broadly neutralizing anti-HA antibodies.
2026-08-28
OpenGerminal: an open-source implementation of the Germinal antibody design pipeline
- protein design
- generative methods
- A restriction free reproduction of the antibody design workflow germinal
- Replaces proprietary dependencies (PyRosetta, IgLM) with an open-source toolchain (OpenMM, AbLang1, sc-rs) and fixes multi-chain bugs, enabling unrestricted academic and commercial deployment.
- Demonstrates that AbLang1-guided hallucination significantly increases initial cofolding pass rates (e.g., 33.7% vs. 18.6% for PD-L1) with equal or higher structural confidence, at the cost of a ~1.5x increase in per-trajectory compute time.
- Uses hard-coded placeholder values for three energy metrics which degrades ensemble selection and disables the interface hydrogen-bond filter and lacks wet-lab experimental validation of the generated binders.
2026-08-21
Benchmarking Antibody Modeling Tools across Structure Prediction, Docking, and Paratope–Epitope Interface Analysis
- structure prediction
- Benchmarking of some structural prediction methods on antibody-related tasks.
- Benchmarked five computational tools across two core tasks: AlphaFold3 (AF3), ImmuneBuilder (ABodyBuilder2/ABB2), and IgFold for antibody variable fragment (Fv) structure prediction, as well as AF3, dyMEAN, and GRAMM for antigen–antibody complex structure prediction and docking.
- Tools were tested on 50 non-redundant humanized antibody–antigen Fab complexes from the Protein Data Bank (PDB), filtered for resolution < 3.0Å and released after December 19, 2023, to eliminate training data overlap.
- All Fv predictors achieved high backbone accuracy (mean TM-score > 0.97), but AF3 demonstrated statistically significant advantages in global Fv geometry and hypervariable CDR-H3 loop modeling (median CDR-H3 RMSD of 0.86 Å, compared to 1.34 Å for ABB2 and 1.52 Å for IgFold). In complex prediction, AF3 substantially outperformed dyMEAN and GRAMM, generating reliable docking poses for 46% of complexes, whereas classical rigid-body docking (GRAMM) and epitope-guided modeling (dyMEAN) almost completely failed.
- Paratope and epitope residue recovery is strictly dependent on initial docking accuracy. When docking succeeds, AF3 reliably identifies interface residues (F1 score ~ 0.85/0.87), salt bridges (84% recall), and non-bonded contacts (75% recall).
2026-08-21
High-accuracy structural modeling of antibody-antigen complexes
- binding prediction
- structure prediction
- Novel method, DeepAAAssembly, to model antibody antigen complexes, beating AF3.
- DeepAAAssembly introduces a hybrid framework that pairs deep learning inter-chain distance predictions with flexible Monte Carlo sampling, overcoming the lack of strong co-evolutionary signals in antibody-antigen interfaces.
- The pipeline feeds Voronoi tessellation contact geometry, AntiBERTy embeddings, and a Triangular Awareness module into a multi-column CNN, creating continuous multi-peak distance energy landscapes to guide both global orientation sampling and local CDR loop refinement.
- Evaluated on Docking Benchmark 5.5, it surpassed AF3 with a 12.9% higher average DockQ score on best-generated models (0.454 vs. 0.402) and boosted the overall modeling success rate (DockQ > 0.23) from 47.8% to 56.7%.
- Demonstrates strong generalization on CASP15/16 blind targets by successfully correcting severe initial orientation errors, elevating previously failed structural predictions into acceptable, physically plausible models
2026-08-21
AbAgKer: A Unified Semi-Supervised Framework for Antigen-Antibody Binding Affinity and Kinetics Prediction
- binding prediction
- Novel antibody-antigen affinity predictor.
- Evaluated on SKEMPI v2.0 and SAbDab curated into three subsets: AbAgA (1,953 labeled affinity pairs), AbAgI (7,725 unlabeled interaction pairs), and AbAgKoff (272 kinetic pairs).
- Compared against five PLM-based baseline setups, A2binder, AntiBERTa2, DG-Affinity, ESM-F, and AntiBERTY-F, which adapt pre-trained language models with regression heads for affinity tasks.
- Generates pseudo-labels on 7,725 unlabeled pairs (AbAgI) using a teacher model trained on AbAgA; retraining on this expanded pool serves as the main engine for gains, boosting Pearson correlation from 0.45 to 0.68 and reducing MSE by ~35%.
- Infuses explicit biological priors into sequence representations via an MEI module (sequence-based biophysical epitope features), CDR-masked attention (CAttention), and a Mixture-of-Experts (MoE) network to dynamically model diverse binding modes.
- Extends beyond static equilibrium affinity (K_d) by using few-shot transfer learning on learned interaction features to predict dissociation rates (k_off) for target residence time analysis.
2026-08-21
A blinded, prospective benchmark of in silico antibody discovery anchored to experimental affinity and developability
- binding prediction
- developability
- ngs
- Results from the antibody discovery competition, where ML methods faced off with experimental baselines.
- Participating organizations were evaluated across three tasks on a SARS-CoV-2 receptor-binding domain (RBD) dataset: (1) in silico affinity maturation by modifying non-HCDR3 CDRs using sequencing outputs, (2) affinity ranking of candidates within three HCDR3-clustered sequence datasets, and (3) out-of-library CDR optimization/design to generate novel high-affinity binders not present in the sequencing data.
- For lead optimization, top ML methods matched experimental performance by producing sub-100 pM binders (top ML design: 95 pM vs. best experimental control: 113 pM), though overall ML submissions underperformed relative to experimental sorting libraries. Notably, a simple non-ML statistical consensus baseline also achieved a highly competitive 540 pM affinity without using machine learning.
- For sequence ranking within HCDR3 clusters, almost all ML algorithms performed worse than random clone picking; only 9.8%-13.8% of ML submissions beat the cluster baseline clone compared to 39% from random selection. Additionally, out-of-library ML designs yielded a high proportion of non-binders (30.4%) and top-ranked affinity hits frequently suffered from severe biophysical developability failures.
- ML is genuinely useful for timeline reduction during lead optimization and affinity maturation, replacing 2-3 weeks of physical combination library construction with an in silico step when provided with deep, biologically grounded experimental datasets.
- Pretrained protein language models (PLMs) paired with structure-aware modules (e.g., pairformers) showed utility in exploring novel out-of-library sequence spaces. However, current ML models lack cross-task generalization, struggle with local epistasis, and fail to reliably co-optimize affinity alongside multiparametric biophysical developability.
2026-08-10
Analysing open-source protein folding models for nanobody binding prediction
- binding prediction
- nanobodies
- ngs
- Benchmarking the ability of co-folding models to distinguish nanobody binders and non-binders.
- Evaluated four state-of-the-art structure prediction models (AlphaFold3, Boltz-2, Chai-1, and IntFold) across true binder ranking, out-of-distribution (OOD) sequence detection, and mutational sensitivity in nanobody–antigen complexes.
- Discovered that no single confidence score excels across all tasks: while Boltz-2 achieved the highest median accuracy for true binder identification, local metrics like PLDDT (particularly in AlphaFold3) were far superior at catching OOD alanine-substituted sequences.
- Contributed a original in vivo camelid immunization dataset targeting CD33, revealing that all evaluated models struggle to generalize when discriminating enriched binders from realistic immune repertoire background sequences.
- Demonstrated that commonly used hard filtering thresholds (e.g., pAE < 10) can erroneously discard up to 75% of true binders, underscoring the need to combine complementary global interface and local CDR metrics rather than relying on single confidence scores.