How marieoestreich/drugdiff works
Project type and purpose
• DrugDiff is a research‑oriented machine‑learning pipeline for de novo molecular generation. It implements a latent diffusion model (DDPM) augmented with predictor‑based guidance to steer generation toward desired molecular properties.
File structure
• Top level:
• data/ – raw and preprocessed datasets (ZINC‑250k, GuacaMol) and symbol/mapping pickles.
• src/ – core library code:
· src/models/ – model definitions (VAE, DDPM variants, predictor guidance).
· src/utils/ – helper functions (SMILES scoring, fingerprinting, I/O).
• scripts/ – CLI entry points for training the affinity predictor and for molecule generation under three modes (unconditional, single‑property, multi‑property).
• notebooks/ – Jupyter notebooks demonstrating evaluation workflows for each generation mode.
• README.md – project overview, usage instructions, architectural figure reference.
README for additional insights
• Confirms the three generation modes (unconditional, single‑property, multi‑property).
• Instructs where to place pretrained checkpoints (/model/).
Final recommendations
• Keep the layout left‑to‑right: data on the left, training in the middle, inference & evaluation on the right.
• Don’t omit any of the three generation modes—each is a distinct flow that uses the same core diffusion engine with different predictor interactions.