We argue that Physics-Informed ML must evolve to meet the unique challenges of biological modeling, and that this evolution represents not a limitation, but a major opportunity, giving rise to Biology-Informed Machine Learning (BIML).

We propose a 3D all-atom flow matching model with prior interaction guidance and a learnable atom number predictor for target-aware molecule generation.

We present a theoretical framework for mask-based pretraining using high-dimensional statistics and introduce R²MAE, a novel pretraining scheme that enhances self-supervised learning across diverse data domains.