We present a physics-based machine learning approach to predict transcription factor binding affinities in vitro from structural and mechanical DNA properties directly derived from atomistic molecular dynamics simulations.