appendix-i

Appendix I. Diffusion & Flow Matching with FlowMol3, DiffSBDD, RFdiffusion, & Boltz-2

 

This chapter covers

  • Why de novo design moved beyond latent-space autoencoders to diffusion and flow matching
  • How diffusion models generate by learning to reverse a noising process
  • How flow matching generalizes diffusion and reaches comparable quality in far fewer sampling steps
  • Generating drug-like 3D molecules from a pretrained model with FlowMol3
  • Designing ligands directly inside a protein pocket with DiffSBDD
  • Using RFdiffusion to design a protein around a ligand, then triaging generated molecules with Boltz-2

With de novo design (Chapter 10), we invent molecules that do not exist yet. The space we invent them from is unimaginably large: somewhere around 1060 drug-like small molecules could in principle be written down, and only a vanishing sliver of them has ever been made. In Appendix C, we learned to tame that space by compressing it. We trained an autoencoder to squeeze each molecule down to a point in a smooth, continuous latent space — a kind of ZIP file for molecules — and then to decode any point back into a structure. Once molecules live on a map like that, generation becomes navigation: sample a point, decode it, and you have a candidate; walk the map toward regions your property model likes, and you have optimization.

I.1 From Latent Spaces to Noise

I.2 Diffusion: Learning to Reverse Noise

I.2.1 The Forward Process: Dissolving Structure into Noise

I.2.2 The Reverse Process: Learning to Denoise

I.2.3 Training the Denoiser

I.2.4 Sampling, with a catch

I.3 Flow Matching: From Noise to Data

I.3.1 From Stochastic Reversal to a Learned Velocity Field

I.3.2 The Conditional Flow Matching Objective

I.3.3 ODE Sampling: Following the Current

I.4 Generating real 3D molecules with FlowMol3

I.4.1 Multi-modal Flow Matching

I.4.2 Loading and Sampling from a Pretrained Model

I.4.3 Validity, QED, and synthetic accessibility

I.4.4 Validity plateaus by fifty steps

I.5 Structure-based Design: Generating inside the Pocket (DiffSBDD)

I.5.1 Pocket-conditioned Diffusion

I.5.2 Generating ligands in the adenosine A2A receptor (3RFM)

I.5.3 Profiling the ligands: drug-like, synthesizable, and novel?

I.5.4 One checkpoint, many tasks: optimizing QED without retraining

I.5.5 Keeping a scaffold: inpainting

I.6 RFdiffusion: Two directions of SBDD

I.7 From generation to triage: Boltz-2 as a connector

I.8 Summary

I.9 References