Risk-controlled candidate screening drops conformal p-values: UNEC calibrates the cutoff directly, taking about 1.3 seconds per repetition over the full simulation grid versus 35 to 37 for SCoRE
Synopsis
Machine learning screens far more candidates than can be experimentally validated, so selection must control risk before validation. UNEC normalizes predictive margins by local uncertainty and calibrates the selection cutoff directly on held-out losses, without building conformal p- or e-values, and supports general bounded losses and observation-specific thresholds. On solubility, toxicity and lipophilicity screening it holds risk-power trade-offs competitive with conformal selection, the NP method and SCoRE variants; in simulation the full grid took about 1.3 seconds per repetition versus 35 to 37 seconds for SCoRE.
Interpretation
Risk-controlled candidate selection can now set its cutoff directly from predictions and held-out losses, without first constructing conformal p- or e-values and running a multiple-testing step. Conformal selection combined conformal p-values with multiple testing to control the false discovery rate, and SCoRE used risk-adjusted conformal e-values for general bounded losses. Across simulations on five data-generating processes, UNEC's empirical false discovery rate tracked the nominal level over the tested target range; on the two Gaussian-mixture processes that depart from the location-scale structure, conformal selection became conservative while the NP method and the unnormalized UNEC-Raw could exceed the nominal level at small targets.
The procedure divides each predicted margin by a local predictive uncertainty to form a standardized score, then picks the smallest cutoff on a held-out calibration set whose sorted scores and cumulative losses satisfy the risk constraint. Normalization separates candidates with similar fitted means but different predictive uncertainty, and calibration reads an empirical loss curve instead of computing e-values at every candidate threshold. Each simulation repetition used training, calibration and test samples with random forests for both prediction and scale estimation; the molecular tasks draw on TDC datasets, encode SMILES strings with AttentiveFP through DeepPurpose, and cover a common one-sided threshold, molecule-specific thresholds and a two-sided interval target.
Under a location-scale model with a common one-sided binary target, UNEC asymptotically attains the oracle power envelope among observation-wise selection rules. For general bounded losses the optimal ranking depends on the loss, so power optimality is stated only for binary loss, where selective discovery risk reduces to the false discovery rate. The result rests on showing that the oracle standardized margin orders observations like the posterior success probability, and that the estimated score and cutoff converge to their oracle counterparts.
Perspective
The demonstrated scope is five simulated data-generating processes and three molecular-property screening tasks from TDC, covering a common one-sided threshold, molecule-specific thresholds and a two-sided interval target; the molecular results only prioritize candidates and do not establish the safety or efficacy of individual compounds, so downstream use still requires experimental validation.
Selective discovery risk control is asymptotic, and the unadjusted rule can show mild finite-sample exceedance at small calibration sizes; power optimality is limited to a common one-sided binary target under a location-scale model, and extending finite-sample selective-risk control and power characterization to more general losses and models remains future work.