Research map / Probabilistic and scientific modeling / Molecules and proteins
Multi-domain Distribution Learning for De Novo Drug Design
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces DrugFlow, a structure-based drug design model combining flow matching (coordinates) with Markov bridges (atom/bond types), plus OOD uncertainty, adaptive size via virtual nodes, side-chain flexibility (FlexFlow), and DPO-style multi-domain preference alignment.
Problem
Structure-based de novo drug design needs generative models that accurately learn the distribution of protein-binding molecules (including 3D poses, chemistry and size), rather than only optimizing docking scores, and that can handle uncertainty, varying ligand size and protein flexibility.
Approach
Jointly models discrete atom/bond types with Markov bridges and continuous coordinates with flow matching in one heterogeneous equivariant GNN; adds a learned per-atom uncertainty head trained with a Gaussian likelihood loss, virtual nodes for end-to-end size selection, torsional flow matching for side chains, and a DPO-style multi-domain preference alignment loss applied across both frameworks.
Key results
- DrugFlow achieves the best FCD of 4.278 versus 11.637 for the next-best DiffSBDD on CrossDocked distribution learning.
- DrugFlow obtains the lowest JSD over joint QED, SA, logP and Vina efficiency distributions (0.099 vs 0.223 for Pocket2Mol), and best or second-best on most Wasserstein metrics.
- Preference-aligned models consistently outperform fine-tuned counterparts across QED, SA, Vina efficiency and REOS, at some cost in validity.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DrugFlow instead of the strongest baseline | CrossDocked distribution learning · CrossDocked | Wasserstein distance across molecular geometry and properties: 0.017 vs 0.017 (C--C); 0.016 vs 0.019 (C--N); 0.016 vs 0.028 (C=C); 0.952 vs 2.173 (C--C=C); 2.269 vs 2.936 (C--C--C); 1 | Helped |
| Use DrugFlow instead of TargetDiff | CrossDocked distribution learning · CrossDocked | Wasserstein distance, logP: 0.665 vs 0.489 | Hurt |
| Use DrugFlow instead of Pocket2Mol | CrossDocked distribution learning · CrossDocked | Wasserstein distance, hydrophobic interactions: 0.036 vs 0.027 | Hurt |
| Use DrugFlow instead of the strongest baseline | CrossDocked distribution learning · CrossDocked | Wasserstein distance, binding efficiency: 0.028 vs 0.034 (Vina); 0.013 vs 0.030 (Gnina) | Helped |
| Use DrugFlow instead of Pocket2Mol | CrossDocked distribution learning · CrossDocked | Wasserstein distance, H-bond interaction types: 0.019 vs 0.031 (H-bond); 0.012 vs 0.021 (acceptor); 0.007 vs 0.010 (donor) | Helped |
| Use DrugFlow instead of Pocket2Mol | CrossDocked distribution learning · CrossDocked | Wasserstein distance, π-stacking interactions: 0.006 vs 0.007 | Helped |
| Use DrugFlow instead of the strongest baseline | CrossDocked distribution learning · CrossDocked | Jensen-Shannon divergence, joint molecular properties: 0.099 vs 0.223 | Helped |
| Use DrugFlow instead of the strongest baseline | CrossDocked distribution learning · CrossDocked | Fréchet ChemNet Distance: 4.278 vs 11.637 | Helped |
| Use DrugFlow instead of the strongest baseline | CrossDocked distribution learning · CrossDocked | Jensen-Shannon divergence, atom types: 0.043 vs 0.050 | Helped |
| Use multi-domain preference alignment rather than fine-tuning on winning samples only | DrugFlow preference optimization · Property optimization: REOS, QED, SA, and Vina efficiency | target property performance | Helped |
| Add preference alignment | DrugFlow · Property optimization: REOS, QED, SA, and Vina efficiency | molecular validity | Hurt |
| Add virtual-node size adaptation | DrugFlow · CrossDocked pocket-conditioned molecule generation | atoms removed when sampling with ground-truth sizes plus 5 atoms: 5 | Helped |
Surprising findings
- Uncertainty score correlates with ligand efficiency, not only molecule size, which the authors attribute to learning to avoid steric clashes.
- Reported that absolute docking/property values are unreasonable to expect from pure distribution learners without fine-tuning or sampling strategies.
Limitations
- Virtual-node approach still requires pre-specifying an upper bound on the number of computational nodes; removal fails beyond N_max=10 extra atoms.
- Preference alignment improvements come at a cost of reduced molecular validity.
- Uncertainty values concentrate in a narrow range (0.85 to 0.92), limiting the score's granularity.
Experimental setup
Datasets: CrossDocked
Benchmarks: CrossDocked test set (100 proteins)
Metrics: Wasserstein-1 distance Jensen-Shannon divergence Fréchet ChemNet Distance validity PoseBusters pass rate
Baselines: Pocket2Mol TargetDiff DiffSBDD
Builds on: flow matching Markov bridge models direct preference optimization diffusion DPO geometric vector perceptrons
Closest papers by meaning
- Active Flow Expansion for Out-of-Distribution Discovery: from Theory to Molecules (NeurIPS 2026 · similarity 0.89)
- Prior-Guided Flow Matching for Target-Aware Molecule Design with Learnable Atom Number (NeurIPS 2025 · similarity 0.89)
- Controllable 3D Molecular Generation for Structure-Based Drug Design Through Bayesian Flow Networks and Gradient Integration (NeurIPS 2025 · similarity 0.89)
- Pretraining Generative Flow Networks with Inexpensive Rewards for Molecular Graph Generation (ICML 2025 · similarity 0.88)
- Enhancing Ligand Validity and Affinity in Structure-Based Drug Design with Multi-Reward Optimization (ICML 2025 · similarity 0.88)
- Energy-Based Flow Matching for Generating 3D Molecular Structure (ICML 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.