Other-Modality Evidence for Prioritizing CAR-T Target Genes: A Label Audit and Benchmark in Two Input Regimes

Background: Ranking genes as CAR-T cell targets is limited by the small number of validated positives (168 of 3,220 candidate genes). Antibody-drug conjugates (ADCs), bispecific antibodies and monoclonal antibodies pursue the same kind of surface antigen, so evidence that another modality already targets a gene may help prioritize it for CAR-T therapy. Results: We built an audited evidence table for 3,220 genes that combines single-cell tumor expression for five cancers, protein topology, normal-tissue expression and CAR-T labels from three trial registries, a CAR-construct library and a curated list, together with ADC, bispecific and monoclonal antibody labels. A model with four outputs, one per modality (CAR-T, ADC, bispecific, monoclonal antibody), was scored in an evidence regime and a biology-only regime under nested, gene-grouped cross-validation with fresh-split re-checks. A label-free count of other-modality evidence reached an end-to-end Matthews correlation coefficient (MCC) of 0.516, matching the best learners. Biology alone ranked curated targets well above chance (ROC-AUC 0.78; best MCC 0.283), with logistic regression and XGBoost ahead of a multilayer perceptron (0.207). A two-stage screen-then-classify design reduced class imbalance from 1:16 to between 1:3.8 and 1:8.4 and produced a tiered gene list; with evidence inputs it reached MCC 0.503 with 32 false positives, against 68 for single-stage XGBoost. Conclusions: Other-modality evidence is the strongest available signal for prioritizing CAR-T target genes, and compact, interpretable learners are sufficient with limited data. The audited evidence table and the two-regime design give a transparent basis for ranking candidate genes.