Research map / Probabilistic and scientific modeling / Molecules and proteins
Navigating Chemical Space with Latent Flows
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.
Proposes ChemFlow, a dynamical-systems framework that learns latent flows (gradient, wave, Hamilton-Jacobi, Langevin/Fokker-Planck) over a pretrained molecule VAE to traverse chemical space for molecule manipulation and single- and multi-objective optimization, in supervised and unsupervised settings.
Problem
Efficiently exploring and understanding the vast chemical space by traversing the latent space of pretrained molecule generative models toward desired molecular properties or structural diversity, under both supervised (property labels) and unsupervised settings.
Approach
Formulates latent traversal as learning a vector field (gradient of a learned energy network) that transports the latent distribution, with different PDE-inspired dynamics (wave, Hamilton-Jacobi, Fokker-Planck/Langevin, heat) unifying gradient-based optimization and linear (ChemSpace) traversal; unsupervised variant uses a Jacobian-vector product structure-change objective plus a disentanglement c
Key results
- Unsupervised Wave flow reaches plogP top-1 of 5.30 versus 3.52 for random and 3.74 for ChemSpace in unconstrained optimization.
- Gradient flow (supervised) achieves the highest average manipulation strict success rate of 7.62% across properties, ahead of ChemSpace at 6.17%.
- Langevin dynamics achieves the best or competitive results in multi-objective QED-SA optimization at all similarity cutoffs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use unsupervised Wave flow rather than Gradient Flow | ChemFlow · Unconstrained plogP maximization | plogP, 1st: 4.06 -> 5.30 | Helped |
| use supervised Wave flow rather than Gradient Flow | ChemFlow · Unconstrained QED maximization | QED, 1st: 0.944 -> 0.947 | Helped |
| use supervised Wave flow rather than ChemSpace | ChemFlow · Unconstrained ESR1 docking minimization | ESR1 docking, 1st: -11.66 -> -11.05 | Hurt |
| use supervised Wave flow rather than ChemSpace | ChemFlow · Unconstrained ACAA1 docking minimization | ACAA1 docking, 1st: -9.81 -> -10.48 | Helped |
| use unsupervised HJ flow rather than ChemSpace | ChemFlow · Similarity-constrained plogP, δ = 0 | absolute plogP improvement: 12.13 ± 6.41 (99.8) -> 20.64 ± 12.93 (98.0) | Helped |
| use LD rather than ChemSpace | ChemFlow · Similarity-constrained plogP, δ = 0.2 | absolute plogP improvement: 9.07 ± 6.80 (90.2) -> 9.70 ± 6.21 (94.4) | Helped |
| use LD rather than ChemSpace | ChemFlow · Similarity-constrained plogP, δ = 0.4 | absolute plogP improvement: 7.52 ± 6.29 (59.4) -> 6.14 ± 5.99 (70.9) | Mixed |
| use LD rather than ChemSpace | ChemFlow · Similarity-constrained plogP, δ = 0.6 | absolute plogP improvement: 5.70 ± 5.84 (20.2) -> 2.94 ± 4.34 (35.4) | Mixed |
| fine-tune the VAE with PDE regularization and use K=1 instead of K=10 (ablation) | Wave (unsup) · Unsupervised plogP maximization | plogP, 1st: 5.30 -> 3.71 | Hurt |
| fine-tune the VAE with PDE regularization and use K=1 (ablation) | Wave (unsup) · Unsupervised QED maximization | QED, 1st: 0.905 -> 0.936 | Helped |
| use supervised Gradient Flow rather than ChemSpace | ChemFlow · Molecule manipulation, plogP | strict success rate: 5.20 -> 6.90 | Helped |
| use supervised Gradient Flow rather than ChemSpace | ChemFlow · Molecule manipulation, QED | relaxed success rate: 18.10 -> 16.70 | Hurt |
Surprising findings
- A random latent direction performs well on molecule optimization and manipulation, which the authors attribute to latent vector norms correlating with molecular properties.
- Unsupervised relaxed success rates for JNK3 and GSK3B exceed supervised ones.
Limitations
- Preliminary study restricted to small molecules; transfer to larger molecular systems is untested.
- Unsupervised flows rely on a disentanglement heuristic and selection by correlation with properties.
Experimental setup
Datasets: MOSES ZINC250K ChEMBL ZINC
Benchmarks: plogP optimization QED optimization ESR1 docking ACAA1 docking molecule manipulation success rate (DRD2, JNK3, GSK3B, QED, SA, plogP) similarity-constrained plogP optimization QED-SA multi-objective optimization
Metrics: top-3 property score strict success rate relaxed success rate success rate with similarity constraint
Baselines: Random Random-1D ChemSpace Gradient Flow (LIMO) Evolutionary Algorithm
Builds on: LIMO ChemSpace JT-VAE Wasserstein gradient flow Wave-like potential flows
Closest papers by meaning
- Pretraining Generative Flow Networks with Inexpensive Rewards for Molecular Graph Generation (ICML 2025 · similarity 0.90)
- Active Flow Expansion for Out-of-Distribution Discovery: from Theory to Molecules (NeurIPS 2026 · similarity 0.89)
- Projecting Molecules into Synthesizable Chemical Spaces (ICML 2024 · similarity 0.89)
- Molecule Design by Latent Prompt Transformer (NeurIPS 2024 · similarity 0.88)
- Navigating the Design Space of Equivariant Diffusion-Based Generative Models for De Novo 3D Molecule Generation (ICLR 2024 · similarity 0.87)
- Generative Flows on Synthetic Pathway for Drug Design (ICLR 2025 · similarity 0.87)
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.