Computational Antibody Papers
Filter by tags
All
Filter by published year
All
TitleKey points
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.
2026-08-10
A systematic evaluation framework for universal antibody-antigen binding affinity prediction and candidate recommendation
- binding prediction
- Benchmark framework for universal antibody-antigen binding affinity prediction evaluated on the AlphaBind dataset.
- Introduces MochiBind, a sequence-based pairwise predictor that compares two antibodies targeting the same antigen to determine which binds stronger.
- Uses pretrained ESM-2 embeddings for antibody and antigen sequences, projecting pooled representations into an MLP to predict relative binding affinity.
- Compared against structural proxies (Boltz-2, GeoDock) evaluated zero-shot without AlphaBind training, and Graphinity fine-tuned on AlphaBind.
- Across the 4 AlphaBind target antigens, average pairwise accuracy reaches ~0.70 for MochiBind, ~0.65 for fine-tuned Graphinity, ~0.58 for Boltz-2, and ~0.50 (random performance) for GeoDock.
- All training, fine-tuning, and out-of-sample evaluation in this study were conducted exclusively on the AlphaBind dataset.
2026-08-10
High-throughput machine learning-aided antibody discovery for cell surface antigens
- databases
- experimental techniques
- ngs
- ML-Optimized Paratope Representation (ARM).
- The authors engineered a synthetic yeast-display Fab library centered on a compact "Antigen Recognition Module" (ARM), a <100 nt sequence pairing heavy-chain CDRH3 sequence diversity with a light-chain barcode, providing a computationally lightweight paratope representation tailored for high-throughput sequencing and ML workflows.
- To overcome physical display biases (such as yeast growth rates or expression disparities), a k-mer-based logistic regression (LR) model was trained on sequence enrichment data from early sorting rounds (MACS to FACS1) to score clones based on sequence motifs.
- Applying the LR model successfully rescued functional, high-affinity binders for ROBO2N (11 binders) and PD-L2 from early sequence pools that had been depleted or overshadowed by dominant clones in later experimental cell-sorting rounds.
- The LR model trained on ROBO2N accurately predicted binding potency and epitope cluster preference for the closely related paralog ROBO1, while reliably scoring non-binding clones as low probability.
- The paper provides a publicly available dataset of over 68,000 unique target-associated ARM sequences alongside comprehensive biophysical characterizations for 486 antibodies, offering a structured foundation for downstream predictive models, affinity maturation, and zero-shot antibody design.
2026-08-10
Speaking the Language of Science: Toward a General-Purpose Generative Foundation Model for the Natural Sciences
- language models
- Autoregressive model for sequence design, with applications to antibody-antigen design.
- LOGOS was trained autoregressively on 44.87 billion scientific tokens across seven modalities, learning interaction constraints entirely from discrete sequences without natural language text or explicit 3D coordinate inputs (so UniRef, OAS, Smiles etc etc, not scientific ‘text’).
- Cleaned antibody sequences from the Observed Antibody Space (OAS) were clustered at 70% sequence identity and encoded using a native scientific grammar with chain- and CDR-specific boundary tokens.
- Evaluated on the SAbDab benchmark for antibody CDR design using Amino Acid Recovery (AAR), refolded structural self-consistency, and AntiBERTy sequence plausibility.
- LOGOS-8B outperformed specialized 3D structure-conditioned inverse-folding baselines on conserved CDR1 and CDR2 loops, achieving over 85% AAR on CDR-L1 and CDR-L2 from sequence context alone - but if you train on OAS, this is hardly a surprise.
- Performance on CDR-3 was far worse for LOGOS.
2026-07-13
Folding, Reasoning, and Scaling with Open-source Drug Discovery Engine
- structure prediction
- Novel structural co-folding method that rivals AF3 architectures whilst being completely open sourced.
- OpenDDE replaces direct coordinate regression with atomic latent reasoning over granular structural tokens and scales the standard Pairformer hidden dimension from 128 to 384, expanding parameter capacity.
- The model implements a differentiable shape-complementarity objective to physically fit interfaces and naturally unifies structure prediction and de novo design inside a single conditional diffusion framework via masking.
- The model uses a 4-stage precision-breadth-precision data distribution to master geometry , consuming ~414k GPU-hours handled efficiently via Fold-CP context parallelism.
- The model sets state-of-the-art antibody-antigen success rates (e.g., 70.0% on FoldBench-AB and 66.4% on 2026ARK-AB) , easily outperforming Protenix-v1 (beating/equalling AF3!), and climbs to nearly 90% oracle accuracy when scaled with test-time seeds.
2026-07-13
Decoding Bispecific Antibody Developability: Design Rules and Predictive Models from a 160-Member Library
- developability
- Developability characterization of multispecific antibodies, for the purposes of training machine learning models
- To map out developer behavior, the researchers used advanced diversity-sampling algorithms to select 160 unique bispecific combinations from 65 parental antibodies, capturing a vast spectrum of biophysical profiles.
- Developability traits don’t transfer uniformly from parent to child; while hydrophobicity and charge are highly predictable, critical flaws like self-association often emerge unexpectedly as "outliers" when the two arms interact.
- The study revealed that thermostability is strictly format-driven, meaning it cannot be reliably predicted from parental data and must be measured directly on the final multi-specific construct.
- Very fast antibody modeling, FlashABB, with applications to developability screening.
- FlashABB accurately performs sota antibody antibody structure prediction in just 5 ms, making it faster than protein language models can even generate sequence embeddings, by completely discarding traditional pair representations in favor of an inner-product-based, linear-memory "Flashpoint Attention" (FPA) algorithm.
- Featuring roughly 7.2 million parameters trained on 8,395 SAbDab structures, the model is strictly forced to run on float32 precision to avoid the catastrophic numerical cancellation errors that heavily degrade 16-bit float implementations during distance calculations.
- Eliminating the structural compute bottleneck allows developers to scale downstream analysis to millions of sequences; using the variant tool FlashTAP, teams can filter out liability flags and predict therapeutic developability metrics for 100 antibodies per second.
- Having said that, CDR-H3 is still sota - so in the region of 2.5A, but it gets there way faster.
2026-06-24
BoltzProt-1: Towards Efficient De Novo Binder Design with Good Developability
- protein design
- binding prediction
- Novel de novo protein binder design pipeline from Boltz that pairs an optimized candidate generation engine with BoltzPPI, a novel, interaction-aware scoring model built to rank designs based on binding confidence rather than just geometric plausibility.
- To train BoltzPPI, the authors created a set of positive labels (PDB and patent complexes) and negative labels (synthetic non-interacting protein pairs co-localized via Boltz-2). This input pool (token, pair, distance, and mask features) is processed through a Pairformer stack using specific optimization tricks (co-trained focal loss, multi-view representation dropping, and Gaussian noise injection) to predict a binary binding-confidence score.
- Significantly improves experimental VHH design performance, boosting the confirmed-binder hit rate from 3.3% to 8.0% on novel targets and successfully discovering screening hits for 7 out of 10 benchmark targets from the Chai-2 dataset.
- Yields highly manufacturable binders, with 58% of its confirmed candidates passing a comprehensive panel of strict biophysical filters (such as thermal stability, purity, and low aggregation Propensity), outperforming both BoltzGen (40%) and clinical-stage VHH controls (21%).
- Generates structurally distinct binders whose CDR loops remain highly distinct from any known entries in the SAbDab database, with the entire pipeline made available to researchers via the Boltz API and Lab platforms.
- Case study of JAM-2, a generative biomolecular design model that engineered drug-like VHH antibodies against five distinct peptide-MHC class I (pMHC-I) targets across two HLA alleles using only target amino acid sequences as input.
- When reformatted into bispecific T-cell engagers (TCEs) without experimental optimization, the designs mediated potent, sub-nanomolar T-cell activation and successfully directed primary human T cells to kill target-presenting cells.
- Binders achieved stringent selectivity, showing a minimum 216-fold preference for NY-ESO-1 over highly similar human self-peptides and safely discriminating mutant KRAS variants (G12V and G12C) from wild-type sequences. A cryo-EM structure verified this atomic accuracy with a whole-complex Ca RMSD of 0.93 Å.
- Bypassing the need for downstream engineering, 78% of the designed antibodies passed five core industry-standard biophysical developability criteria, exhibiting strong expression titers, favorable monomericity, thermal stability, and low polyreactivity
- Theoretical framework for predicting molecular recognition from sequence alone structures are not required at inference time.
- The selection probability governing which molecules bind which is provably unique, it is the Boltzmann distribution, derived from first principles rather than assumed.
- To use it, we need sequence representations whose dot product approximates relative binding energies, not just higher for binders vs non-binders, but correctly ranking the full competitor pool.
- This separates two problems: learning the embedding geometry (cheap, binary binding pairs) from recovering absolute binding energies (cheap, just two calibration parameters fitted against a small number of experimental KdK_d Kd measurements).