Asymmetric cross-reactivity of nuclear receptors reveals an evolutionary buffer between estrogen and androgen signaling.

Nuclear receptor ligand interaction landscapes remain incompletely characterized, particularly at the scale of entire receptor families. We used Boltz-2, a structure-based deep learning framework for protein-ligand affinity prediction, to perform an unbiased survey of endogenous hormones across all human nuclear receptors. This analysis revealed an asymmetric pattern within steroid hormone receptors: estradiol showed broad predicted compatibility across multiple receptors, whereas other steroid hormones did not exhibit reciprocal compatibility with estrogen receptors. To interpret this pattern, we integrated structural modeling, evolutionary analyses, physiological contextualization, and prior experimental observations, which collectively led us to propose an evolutionary buffering framework for steroid receptor interactions. This study illustrates how large-scale affinity prediction can move beyond individual receptor-ligand pairs to reveal family-level interaction architectures, generate testable hypotheses, and guide future experimental investigation.