Development Status, Not Potency: A Data Audit of the Public ADCNet Antibody–Drug Conjugate Evaluation Set

Background. Deep learning models for antibody–drug conjugate (ADC) activity are evaluated against several distinct prediction tasks: recognizing constructs assembled from familiar components, predicting unfamiliar combinations of known components, and extrapolating to unfamiliar antibodies or payloads — which a single reported accuracy does not separate. A second difficulty arises when the activity label itself combines experimental potency with clinical development status. ADCNet, a published framework integrating ESM-2 protein embeddings and FG-BERT small-molecule embeddings, reports 87.12% accuracy, balanced accuracy 0.8689 and AUC 0.9293 on its test set, and releases both its data and its code, making it a tractable case for examining both issues. Methods. We audited the released ADCNet evaluation workbook (50 constructs) and modelling workbook (435 constructs). Antibodies were identified by exact heavy- and light-chain sequence rather than by name; payloads and linkers by canonical SMILES. We cross-tabulated the supplied binary activity labels against development status, compared labels across the supplied activity thresholds, quantified overlap between training and evaluation components, examined the representation of drug-to-antibody ratio (DAR), and evaluated single-component reference predictors. Results. The 50-construct evaluation split is a subset of the 435-construct modelling file. Measured against the 385 held-in constructs, 33 of 50 evaluation constructs (66%) contain no component absent from training; 40 of 50 (80%) share a payload–linker pair with a training construct, and 10 of 50 (20%) share an exact antibody–payload–linker triple, differing only in DAR. Reference predictors fitted on the 385 training constructs and scored on the evaluation split reach 0.778 (antigen identity) and 0.773 (linker identity) against a 0.680 majority baseline, compared with the 0.8712 reported for the full model. Labels are near-invariant to the activity threshold: 415 of 435 constructs (95.4%) carry identical labels at all four supplied cutoffs spanning 1 nM to 1000 nM. Antibody names are not injective with respect to sequence: 154 names map to 149 heavy/light chain pairs, and one identical sequence pair appears under six names across 89 constructs, 20.5% of the corpus. DAR is interval-censored for 141 of 435 constructs (32.4%) and silently replaced by range midpoints. Within the evaluation split, where a development-status field is supplied, all 17 constructs with any clinical status are labelled positive without exception while investigative constructs are positive in 17 of 33 cases; the modelling file contains no status column, so this result cannot be extended corpus-wide. Conclusions. A single-component lookup requiring no learned representation reaches 0.778 on the evaluation split against a reported 0.8712, on a split two-thirds of which contains no component unseen in training. Labels are nearly invariant across three orders of magnitude of activity threshold, and identity harmonization by antibody name would place one sequence, under six aliases spanning a fifth of the corpus, on both sides of a held-out partition. These are properties of the released data rather than of any architecture, and each is verifiable from the public repository. We recommend grouping on sequence rather than name, reporting component-overlap and component-identity baselines alongside model metrics, reporting threshold sensitivity, recording censoring explicitly, and releasing the measurement-level data from which labels were derived.