Evidence-aware comparison of sequence-centric machine learning for antibody discovery and optimization

Sequence-centric machine learning is increasingly used across antibody discovery and optimization, from repertoire-scale representation learning to target-aware scoring and generative design. Cross-study comparison remains difficult because methods differ in target conditioning, molecular output, dataset construction, benchmark design, and validation evidence. We therefore conducted a structured mapping of primary antibody machine-learning studies reported from 2020 to 30 June 2026 and organized the literature using a three-layer functional stack-foundation priors, scorer-rankers, and generator-optimizers-and three analytical axes: conditioning interface, output granularity, and validation evidence profile. A standardized method-level synthesis is complemented by representative anchor cases and four framework-guided audits showing how split units, recovery metrics, computational proxies, and sequence novelty can change the interpretation of headline results. The framework separates training supervision and internal evaluation from complementary domains of independent validation and links increasing molecular commitment to broader evaluation needs. We also provide an operational reporting checklist for auditing datasets, splits, negative construction, generative evaluation, experimental attrition, and resource availability. The framework is intended as a comparative audit scaffold for evidence-aware interpretation rather than as a universal performance ranking or formal benchmarking standard.