Controlling the Diffusion Denoising Process: A Molecular Show

This blog post is supporting my poster at Young Modellers Forum and makes things way easier to see and understand. Underneath each GIF, is the explanation of what you should look for as things denoise throughout the diffusion trajectory. Click the GIFs for higher quality viewing!

Reward: Increase Oxygen Count

Oxygen atoms are coloured in red.

Trajectory 1: PRISM with easily interpretable, proof-of-concept reward:

Notice at the beginning a diverse set of atoms (multiple colours). As timesteps progress toward t=0, all atoms transform to oxygens (red).

Reward: Increase Similarity

I ran a similarity reward as an experiment – see below for PRISM’s denoising trajectory after training.

Trajectory 2: On the left PRISM with SuCOS similarity reward, Trajectory 3: On the right, DiffSBDD for the same target.

Notice the shape overlap and the inclusion of a dichlorobenzene ring generated by PRISM. This is contained in the ground truth shown in white. Although placement is incorrect, the policy has optimised for inclusion of this pharmacophore in line with the ground truth. While DiffSBDD has good shape overlap, it struggles with inclusion of similar pharmacophores.

I hope you enjoyed my poster and the molecular show!

Author