AlloShift: Population-Directed Binder Design with State-Selective Molecular Landscapes

At fixed chemical composition, differential affinity determines the direction of equilibrium population redistribution across interconverting receptor states. Designing a binder that favors a desired functional population therefore requires learning state preference from sparse, heterogeneous data with paired measurements, censoring, limited overlap among binders and receptors, and uneven coverage of noncanonical chemistry. In this work, we develop Alloshift to learn directional state preference from these measurements and select peptides from generated pools. Within AlloShift, $Q_{bind}$ combines peptide chemistry, receptor state geometry, and biochemical context without requiring a peptide and receptor complex structure. On 59 receptor-cold pure-conformation pairs, two disjoint five-seed ensembles achieve balanced direction accuracies of 0.703 and 0.756, compared with 0.583 and 0.661 after peptide permutation ($p=0.003$ for both). Across 154 direction tasks spanning 77 receptor-cold pairs, $Q_{bind}$ reranking increases the fraction predicted in the requested direction by the separate evaluation ensemble from 0.348 to 0.526 for PepTune and from 0.347 to 0.538 for TR2-D2. We observe positive enrichment with a distinct structure-conditioned cyclic-peptide generator and after noncanonical diversification. Together, we establish a measurement-grounded computational route for selecting generated binders by predicted state preference. Prospective affinity and population measurements remain necessary to test biological control.