Enabling late-stage drug diversification by high-throughput experimentation with geometric deep learning - Nature Chemistry