Investigators applied a Gaussian-process active-learning pipeline to eight public AbBiBench assays and reported a large improvement over random selection when prioritising binders under a strict labelling budget. The configuration combining a continuous Tanimoto kernel with an upper-confidence-bound (UCB) acquisition rule produced the highest mean Recall@2% point estimate across tested kernel–acquisition pairs. Reported numbers in the preprint: 57.8% mean Recall@2% across five random seeds, and 68.3% when scored against the acquisition-eligible pool. Random selection, by contrast, achieved 9.6% Recall@2% in the same benchmark.
The authors performed a paired Wilcoxon test across datasets and observed a significant gap versus random selection (p=0.0078). They also noted that, at five seeds, Tanimoto+UCB was not significantly different from RQ+UCB or Tanimoto+EI, indicating that some alternative kernel–acquisition combinations produced similar point estimates in these experiments.
The framework embeds variable heavy and light chain (VH/VL) sequences using ESM-2 650M. Investigators compress those protein-language-model representations and train Gaussian process surrogates on the compressed embeddings. The benchmark enforces a unified labelling budget equal to 10% of each full pool, uses a label-blind seed for initialization, and evaluates performance on a fixed held-out split. The source reports that kernel, acquisition, representation, classical featurisation, and labelling-budget choices were evaluated under this unified protocol.
The preprint adopts the continuous Tanimoto kernel described in the cited literature, and compares it against alternative kernels and acquisition functions. The manuscript does not provide further experimental protocol details in the abstract; readers should consult the full methods in the preprint PDF or the linked code repository for stepwise procedure and implementation specifics.
Performance varied by dataset family. On multiple influenza libraries, the Tanimoto+UCB configuration recovered large fractions of the top binders: the authors report recovery percentages for specific assay identifiers in the abstract (for example, values for 3gbn_h1, 3gbn_h9, and 4fqi_h3). These influenza assays showed higher recall values under the same labelling budget.
By contrast, the AAYL combinatorial scFv libraries targeting the SARS-CoV-2 Spike HR2 peptide posed a harder optimisation problem. The authors observed lower and more variable recall in those landscapes, suggesting combinatorial diversity or signal-to-noise characteristics that reduce label-efficiency for the GP active learner.
An additional three-seed representation ablation was performed under the original protocol and not repeated after the evaluation corrections described in the preprint. Within that ablation, ESM-2 650M ranked highest by mean Recall@2% against several other encoders including ProtBert, AntiBERTy, AbLang2, ProGen2, and classical encodings. The authors reported mean Recall@2% 0.673 ± 0.28 for ESM-2 650M in that ablation run. Readers should note that the representation ablation was run three times only and was not repeated after evaluation corrections, as stated in the source.
The manuscript pairs attribution methods with structure mapping to examine residue-level signals. Structure-mapped Integrated Gradients and KernelSHAP recovered attribution patterns concentrated in complementarity-determining regions (CDRs), which the authors report are consistent with contact-derived paratope geometry. However, when the same signals were evaluated with a threshold-free discrimination metric (AUROC), residue-level discrimination was not strong. In short, qualitative agreement appears between attribution maps and known contact geometry, but quantitative residue-level predictive power was limited in the analyses the authors report.
The preprint includes statistical comparisons across datasets: the paired Wilcoxon test reported a significant advantage of Tanimoto+UCB over random selection (p=0.0078). The authors also observed variability across random seeds and datasets—particularly within the AAYL series—emphasising heterogeneity in label-efficient recovery across different antibody libraries. The manuscript reports that some analyses (notably a representation ablation) were run under an earlier protocol and were not repeated after evaluation corrections; those procedural details and their potential effect on results are explicitly acknowledged in the preprint.
This work is presented as a preprint and has not been peer reviewed. The authors note evaluation corrections and specify that an additional ablation was not re-run after those corrections. The abstract does not report additional experimental details such as full hyperparameter sweeps, exact compression procedures used before GP training, or per-dataset sample sizes in the body text; those details must be taken from the full manuscript or the project repository. The code and supplementary materials are available at the project link provided in the footnotes: https://github.com/harshitsinghsnu/AntibodyAL. Future verification and replication should consult those resources and the full preprint PDF for complete methods and data provenance.