Neuro-Symbolic Control of Structural Queries for TCR–Peptide Selection

We study sequential selection of peptide variants using structural predictions. Our neuro-symbolic system combines AlphaFold 3 features, learned peptide scores and a stopping rule that bounds the shortlist’s worst-case score loss relative to the fully evaluated predictor. We evaluate the system using saved predictions for 1,197 receptor–variant pairs from three previously inspected peptide families. Adding structural features to two matched sequence predictors increases the number of selected pairs with positive activation labels from 21 to 31 and from 24 to 31 out of seventy; random selection yields 19.59 in expectation. Across 420 numerical trajectories, upper-bound querying reduces the mean number of blocks required for an exact score certificate from 844.0 to 613.0 for the primary predictor, compared with uniform querying. Yet at 171 candidate queries its precision at ten, averaged equally across families, is slightly lower, 0.484 versus 0.491. Candidate-specific language-model advice shows no consistent gain. A language model that passes tool-use checks follows the same query rule as a fixed upper-bound policy. The stopping guarantee concerns predicted scores and does not by itself ensure that selected variants activate T cells. Code and verification scripts are available at https://anonymous.4open.science/r/neurosymbolic-tcr-selection-5CFD/README.md.