DrugFlow Design + Boltz-2 Affinity

Design new molecules for a binding pocket with DrugFlow, then score every design with Boltz-2 affinity and get a ΔG table ranked best-first.

by Temple Compute

Run DrugFlow Design + Boltz-2 Affinity on Horus

W-32 · DrugFlow + Boltz-2 Affinity

Design new molecules for a binding pocket with DrugFlow, then score every design with Boltz-2 affinity and get a ΔG table ranked best-first.

Domain: Drug Discovery

Overview

Generates novel molecules for a binding pocket with DrugFlow, then scores each with Boltz-2 affinity and ranks them by estimated ΔG. The pocket is defined by a protein and a reference ligand. The workflow has 6 tasks.

Pipeline

download_checkpoint  Zenodo checkpoint            ──► drugflow.ckpt
fetch_source         DrugFlow @ pinned commit     ──► drugflow_src/
generate             protein + ref ligand + ckpt  ──► results/samples.sdf
prepare_inputs       one Boltz YAML per molecule  ──► results/boltz_inputs/, results/smiles.json
predict              boltz predict                ──► results/predictions/affinity_*.json
rank                 affinity -> ΔG, sort         ──► results/deltaG_table.csv

Inputs / Outputs

ΔG = 1.364 · (affinity_pred_value − 6) kcal/mol, treating IC50 as a Kd proxy at 298 K. It is meant for ranking, not as an exact free energy.

Parameters

| id | Default | Meaning | |---|---|---| | protein | examples/kras.pdb | Target protein, same frame as the reference ligand | | ref_ligand | examples/kras_ref_ligand.sdf | Bound reference ligand | | n_samples | 10 | Molecules DrugFlow generates (1-1000) | | pocket_cutoff | 8.0 | Pocket residues within this distance (Å) of the ligand (4-12) | | seed | 42 | Seed for DrugFlow and Boltz-2 | | use_gpu | on | Run DrugFlow and Boltz-2 on GPU; off runs both on CPU (slow) | | use_msa | on | Boltz-2 MSA server; off is single sequence | | diffusion_samples | 1 | Boltz-2 structure samples per molecule (1-10) |

References