Workflows

This section gives example end-to-end workflows built from the MolCraftDiffusion modules. Unlike the tutorials, which explain individual capabilities one by one, these pages show how those capabilities are combined in practice.

Workflow Families

The examples in this section follow four common patterns:

  1. Generate and filter from a pre-trained diffusion model.

  2. Transfer learning and generate for a more target-relevant chemical space.

  3. Conditioned generation with property or structure guidance.

  4. End-to-end design loop from generation to ranking and validation.

How to Use This Section

Use these pages when you already understand the individual modules, but want a clearer picture of which sequence of steps fits your design goal.

Note

Adapt a pre-trained model; don’t train from scratch. Training a 3D diffusion model from scratch is expensive and rarely necessary. Start from one of the shipped pre-trained bases (GEOM-drugs foundation, or QM9) and fine-tune it toward your target chemistry — this is the default recommended path. The only case that forces training from scratch is chemistry that uses atom types outside the pre-trained vocabulary (H, B, C, N, O, F, Al, Si, P, S, Cl, As, Se, Br, I, Hg, Bi). Pre-trained checkpoints: https://zenodo.org/records/19511401. See Transfer Learning and Generate.

Workflow Comparison

Workflow

Best for

Main idea

Typical next tutorial

Generate and filter

Broad exploration

Sample first, narrow later

Tutorial 5, Tutorial 9

Transfer learning and generate

Target-focused exploration

Adapt the model to a more relevant dataset before sampling

Tutorial 4, Tutorial 5

Conditioned generation

Directed search

Bias generation toward a structure or property target

Tutorial 6, Tutorial 7

End-to-end design loop

Candidate prioritization

Combine generation, filtering, optimization, prediction, and ranking

Tutorial 8, Tutorial 9

Pages