Computational Antibody Papers

Filter by tags
language models
Filter by published year
2026
TitleKey points

  • 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.

  • 2026-08-10

    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.

  • 2026-06-23

    GermRL: Alleviating The Germline Bias In Autoregressive Antibody Language Models Through Reinforcement Learning

    • language models
    • New language model addressing the germline bias in NGS training data.
    • GermRL fine-tunes an autoregressive ProGen2-OAS base model using a modified, outcome-supervised Group Relative Policy Optimization (GRPO) framework that features epoch-end weight syncing and automated "prefix grafting." So instead of using an absolute baseline, GRPO generates a local batch of sequences simultaneously; since the reward function favors mutations, any safely mutated sequence scores higher than its germline-heavy peers, resulting in a positive relative z-score that "up-votes" the mutated path while "down-voting" the underperforming germline ones.
    • This is the first framework to fix germline bias in generative, autoregressive models (unlike previous efforts that targeted masked models), and it acts as a lightweight, modular plugin requiring no scratch training or data pre-processing.
    • It was evaluated on one-shot generation success (pass@1) across specific mutation bounds (LD5 to LD35), measuring structural plausibility via ESMFold, sequence/V-gene diversity, developability metrics (GRAVY/instability), and semantic overlap with natural human antibodies using AntiBERTy UMAP embeddings.

  • 2026-05-29

    Protein language models trained on biophysical dynamics inform mutation effects

    • language models
    • Protein Language model that understands protein dynamics.
    • The authors leveraged existing structural data and datasets like mdCATH to gather equilibrium fluctuations for 64,403 proteins. Instead of raw time-series trajectories, they extracted calculated biophysical properties like root-mean-square fluctuations (RMSF) and Normal Mode Analysis (NMA) to serve as training labels.
    • They trained two models, SeqDance 'from scratch' and ESMDance as an extension of ESM2. ESMDance: Built by fine-tuning the pre-trained ESM-2 transformer, teaching it to map its existing evolutionary knowledge to these new physical flexibility profiles. SeqDance: Trained completely from scratch using only raw sequences and the target dynamics data, forcing it to learn pure, unbiased residue co-movement and physics.
    • To test zero-shot mutation prediction, the models compare the wild-type flexibility against the mutated sequence's flexibility. A large mathematical discrepancy flags a highly disruptive, damaging mutation. These predicted shifts were correlated against deep mutational scanning (DMS) lab data measuring actual cellular fitness and stability changes. ESMDance is the go-to for mutation prediction (especially on viral and de novo designed proteins with no evolutionary history), while SeqDance wins at modeling highly flexible Intrinsically Disordered Regions (IDRs).

  • 2026-05-29

    Language Modeling Materializes a World Model of Protein Biology

    • language models
    • protein design
    • structure prediction
    • New versions of ESM and ESMFold, ESMC (ESM Cambrian) and ESMFold2 to model protein sequence, structure, and function.
    • Like previous versions, ESMC is trained entirely on sequences using a masked language modeling (MLM) objective. However, it scales up to 2.8 billion metagenomic sequences, nearly a 100-fold increase over the 50 million sequences used for ESM2.
    • ESMFold2 achieves state-of-the-art atomic resolution and directly outperforms AlphaFold3 on complex antibody-antigen predictions (even when AlphaFold3 is given multiple sequence alignments and ESMFold2 operates from sequence alone).
    • To design therapeutic antibody fragments (scFvs), they input the target sequence and lock in a stable, known antibody framework template, leaving only the target-recognizing loops (CDRs) to be filled in.
    • AI-guided optimization: The system uses mathematical backpropagation through both ESMC and ESMFold2 to iteratively optimize the CDR sequences. It automatically mutates the loops to maximize structural interface confidence scores.
    • They synthesized and tested these computational designs in the wet lab, achieving high experimental hit rates and discovering entirely novel binders with therapeutically relevant nanomolar affinities.

  • 2026-04-30

    How to make the most of your masked language model for protein engineering

    • protein design
    • language models
    • Method to select mutants computationally for lab testing.
    • "Stochastic beam search," a sequence-centric method that evaluates masked language models (MLMs) via pseudo-log-likelihood, rather than using costly mutation-centric approaches.
    • This technique is computationally efficient and produces higher-quality sequences by better balancing likelihood and diversity.
    • The method was extensively validated through both in silico evaluations across various models and direct head-to-head in vitro antibody campaigns.
    • In wet-lab testing, the optimized models effectively screened for synthesizability and binding, with supervised guidance achieving a 100% success rate in the experiments

  • 2026-03-13

    Separating selection from mutation in antibody language models

    • language models
    • Novel training scheme for antibody language models, modeling phylogenetic relationships rather than pure mutational MLM - called DASM.
    • Unlike AbLang2’s standard masked language modeling , DASM uses a mutation-selection framework that factors out nucleotide-level biases (like the codon table and SHM rates) to isolate purely functional selection effects.
    • The model was trained on approximately 2 million parent-child sequence pairs derived from reconstructed B cell phylogenies , using datasets such as JaffePaired, Tang, and Vanwinkle.
    • Model is a compact 4-million-parameter Transformer-encoder featuring 5 layers, 8 attention heads , and a custom "wiggle" activation function to stabilize output selection factors.
    • DASM was validated on the FLAb collection (Koenig and Shanehsazzadeh datasets) and MAGMA-seq high-throughput binding assays for influenza and SARS-CoV-2 antibodies. It was better than ABlang2, progen2 and esm2.

  • 2026-01-09

    NbBench: Benchmarking Language Models for Comprehensive Nanobody Tasks

    • language models
    • nanobodies
    • Benchmarking of pretrained protein, antibody, and nanobody language model representations on a comprehensive suite of nanobody-specific tasks.
    • The authors introduce eight tasks spanning variable-region annotation, CDR infilling, antigen binding prediction, paratope prediction, affinity prediction, polyreactivity, thermostability, and nanobody type classification (e.g. VHH, VNAR, conventional antibody chains).
    • They evaluate generic protein LMs, antibody-specific LMs, and nanobody-specific LMs under a unified and standardized benchmark.
    • All backbone models are kept frozen, with task-specific lightweight heads trained on top to isolate representational quality.
    • No single model consistently outperforms others across all tasks, showing that nanobody-specific pretraining alone does not guarantee superior performance over antibody-specific or generic protein language models.