A Neutrosophic Assay-Consensus Gate for Class-Balanced Selective Molecular Bioactivity Prediction

Abstract

Tox21 molecular bioactivity prediction is strongly class-imbalanced, so a reject option based only on classifier confidence can preferentially retain abundant inactive 
compounds while discarding rare active cases. NACG-Tox is a post-hoc neutrosophic assay consensus gate that separates prediction direction from the decision to accept or defer a prediction. For each compound-endpoint pair, endpoint-specific neighbors drawn only from the fitting set define three interpretable memberships: truth T for local evidence supporting activity, falsity F for local evidence supporting inactivity, and indeterminacy I for insufficient chemically relevant reference evidence. The primary consensus score, G={\left(1-I\right)}^{2}\left|T-F\right|, is large only when the query has adequate local support and the neighborhood is directionally coherent. The study uses the complete public Tox21 Challenge training archive. Deterministic curation reduced 11,764 SDF records to 6,442 canonical single-component molecular identities across 12 in vitro assays, with endpoint activity prevalence ranging from 2.86% to 15.87%. A leakage-controlled hybrid structural split combined Bemis-Murcko scaffolds for cyclic molecules with fingerprint clustering for acyclic molecules and was evaluated under three fixed allocation seeds. With an ECFP-based logistic predictor, full-coverage macro average precision was 0.437 ± 0.034 and macro ROC-AUC was 0.798 ± 0.028. Under validation-locked deferral policies, NACG achieved macro balanced accuracies of 0.587, 0.597, and 0.620 at realized mean coverages of 0.796, 0.626, and 0.442, respectively, compared with 0.517, 0.509, and 0.512 for probability margin confidence. In the equal-coverage diagnostic, NACG achieved balanced accuracies of 0.587, 0.601, and 0.631 at 80%, 60%, and 40% coverage and retained 71.1%, 45.4%, and 23.1% of active cases; the probability-margin selector retained 37.0%, 22.4%, and 12.8%. The same qualitative pattern was observed with a fixed Random Forest backbone. NACG 
therefore improves the class balance of the retained prediction set rather than uniformly reducing ordinary classification error. The framework concerns assay-specific in vitro activity and must not be interpreted as a direct probability of human toxicity. 

DOI: 10.5281/zenodo.22804639