Research map / Probabilistic and scientific modeling / Molecules and proteins
Active Flow Expansion for Out-of-Distribution Discovery: from Theory to Molecules
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 Active Flow Expansion (ActFlow), a continued pre-training method that uses verifier feedback and uncertainty-driven active exploration in a flow model's intermediate representation to expand the model's generable set to new valid regions, with reachability-based guarantees.
Problem
Standard flow and diffusion pre-training matches data distribution, covering only a small part of the valid design space, so out-of-distribution valid designs get negligible probability and are hard to generate for discovery tasks.
Approach
Views a generative model through its generable set (region with non-negligible probability) and expands it iteratively by actively sampling, via a GP-style linear-kernel uncertainty over noised flow hidden features, points where a black-box verifier is uncertain, then fitting the flow to verifier-accepted samples (optionally unlearning rejected ones).
Key results
- On GEOM-Drugs, ActFlow reaches 144.3 valid clusters vs 89.33 (RecF) and 44.67 (RecNF) from a pre-trained baseline of 35.89, with Vendi 303.1.
- On QM9, ActFlow reaches 88.40 valid clusters vs 45.40 (RecF) and 42.80 (RecNF), with Vendi 306.08 and validity 95.90%.
- On therapeutic peptides, coverage rises from 44.33 to 358.33 clusters, while RecNF collapses to 0% validity.
Surprising findings
- Recursive verifier-filtered self-generation (RecF) can decrease coverage even while increasing validity.
- Unfiltered recursive self-generation (RecNF) collapses to 0% validity on therapeutic peptides.
Limitations
- Guarantees are reachability-based and rely on strong assumptions (well-specified logistic verifier in RKHS, approximate uncertainty maximization, EBM abstraction of flow updates).
- Evaluates coverage via cluster counts and validity via proxy verifiers rather than real-world discoveries.
- Real-world discovery gains are not yet assessed.
Experimental setup
Datasets: QM9 GEOM-Drugs therapeutic peptides CreiLOV
Benchmarks: QM9 GEOM-Drugs therapeutic peptide design protein sequence design (CreiLOV)
Metrics: coverage (number of valid clusters) diversity (Vendi score) validity (%) FID
Baselines: RecNF (recursive self-generation, unfiltered) RecF (recursive self-generation, verifier-filtered)
Builds on: flow matching FlowMol SGPO PepTune SISS data unlearning
Closest papers by meaning
- TFG-Flow: Training-free Guidance in Multimodal Generative Flow (ICLR 2025 · similarity 0.90)
- Flow Density Control: Generative Optimization Beyond Entropy-Regularized Fine-Tuning (NeurIPS 2025 · similarity 0.90)
- Pretraining Generative Flow Networks with Inexpensive Rewards for Molecular Graph Generation (ICML 2025 · similarity 0.89)
- Multi-domain Distribution Learning for De Novo Drug Design (ICLR 2025 · similarity 0.89)
- Generative Flows on Synthetic Pathway for Drug Design (ICLR 2025 · similarity 0.89)
- Navigating Chemical Space with Latent Flows (NeurIPS 2024 · similarity 0.89)
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.