The performance gains over previous FlowMol versions are due to 3 techniques which are cheap and architecture agnostic. We hypothesize that these techniques operate synergistically to reduce a common pathology in transport-based generative models.

I'm excited to share FlowMol3! The 3rd (and final) version of our flow matching model for 3D de novo, small-molecule generation. FlowMol3 achieves state of the art performance over a broad range of evaluations while having ≈10x fewer parameters than comparable models.

But that's not the whole story. We introduce methods to quantify molecule quality at the level of functional groups + ring systems. "Valid" generated molecules tend to contain significantly more reactive functional groups than in the training data.

We test a handful of discrete flow matching methods for 3D de novo molecule design and provide some explanations for their differing performance. The result of this is a version of FlowMol with CTMC flows that achieves SOTA validity with fewer learnable parameters.

I'm presenting a new paper "Exploring Discrete Flow Matching for 3D De Novo Molecule Generation" at @workshopmlsb.bsky.social this week! More info in this thread but reach out if want to chat at NeurIPS about generative models or molecular design. arxiv.org/abs/2411.16644