The advent of a new modeling paradigm known as ‘differentiable programming’ makes possible bespoke machine-learned models of biological phenomena that are partly learned from data and partly informed by human-derived biophysical knowledge. In this talk I will describe three instantiations of this new approach for (i) de novo protein structure prediction, (ii) elucidation of the combinatorial grammar underlying metazoan signaling networks, and (iii) design of new protein function. In all cases qualitative improvements in model accuracy or speed, or both, are achieved using differentiable programming, enabling new scientific insights into biological macromolecules and the networks they comprise.