Restoring the RNA-Ligand Interaction Manifold via Topology-Preserving Contrastive Learning under Epistemic Uncertainty

Targeting RNA with small molecules is severely bottlenecked by epistemic uncertainty in high-throughput screening data, formally manifesting as the "Unknown-as-Negative" (UaN) bias. We demonstrate that naive Empirical Risk Minimization (ERM) under UaN contamination structurally misaligns optimization trajectories, introducing adversarial decoy gradients that force a topological rupture of the chemical manifold and shatter Structure-Activity Relationship (SAR) continuity. To transcend this epistemic barrier, we propose Topology-preserving Robust Adaptive Contrastive Embedding (TRACE), shifting the paradigm from heuristic data-fitting to theoretically guaranteed manifold reconstruction. TRACE neutralizes topological collapse through three mathematically bounded mechanisms: (i) confidence-weighted importance sampling dynamically rectifies the surrogate risk under posterior drift, establishing a uniform convergence bound to selectively rescue unannotated bioactive scaffolds; (ii) thermodynamic temperature regulation strictly bounds the backpropagated Jacobian factor, enforcing local Lipschitz continuity to geometrically shield SAR against stochastic gradient explosions; and (iii) margin-aware contrastive regularization provides explicit exponential tail guarantees on misordering probabilities, ensuring discriminative sub-Gaussian alignment between genuine binders and decoys. Grounded entirely in these theoretical pillars, empirical validations strictly reflect our formal bounds. Beyond establishing state-of-the-art generalization, yielding an 18.2% relative AUPRC improvement on SMRTnet, TRACE sustains 1.44x higher topological resilience under extreme 50% noise corruption, and fundamentally resolves UaN-induced structural overconfidence by slashing Expected Calibration Error (ECE) by 83.9% and reducing Brier score by 3.4%.