Research map / Probabilistic and scientific modeling / Molecules and proteins

MF-LAL: Drug Compound Generation Using Multi-Fidelity Latent Space Active Learning

ICML 2025 · Probabilistic and scientific modeling › Molecules and proteins · approach: Structure-based 3D molecule generation · 2 citations (OpenAlex)

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 MF-LAL, a multi-fidelity latent-space active learning framework with hierarchical latent spaces and per-fidelity decoders that generates drug compounds optimized for absolute binding free energy (ABFE) oracles.

Problem

Generative drug design typically relies on cheap but inaccurate docking as the oracle, while accurate binding free energy (ABFE) calculations are too expensive to use within generative loops; this paper seeks to generate compounds with real-world binding activity under a fixed compute budget.

Approach

Integrates the generative model and multi-fidelity surrogates into one framework using a hierarchy of latent spaces (one per fidelity, each with its own decoder and mapping networks), performs Bayesian query-synthesis active learning with UCB, and adds a likelihood term that restricts high-fidelity generation to compounds scoring well at lower fidelity.

Key results

• MF-LAL achieves mean ABFE of -6.2 kcal/mol on BRD4(2) and -6.7 kcal/mol on c-MET over 15 compounds, versus -3.1 and -2.9 for the best-performing or similar baselines in those target-specific comparisons (about -1.6 and -2.8 kcal/mol better than the top baseline).
• MF-LAL generates 6 active scaffolds for BRD4(2) and 4 for c-MET out of 15 compounds, more than any baseline.
• With 40 compounds, MF-LAL mean ABFE is -6.3 (BRD4(2)) and -7.1 (c-MET) kcal/mol, significantly better than the most competitive baselines (p<0.05).

Measured effects reported in the paper

Surprising findings

• Multi-fidelity baselines other than MF-LAL performed mostly similarly to single-fidelity methods, suggesting architecture matters for exploiting multiple fidelities.
• Removing the likelihood term from the generation objective greatly reduced performance (mean ABFE -3.4 vs -6.2 on BRD4(2)).

Limitations

• Limited set of oracles (four fidelity levels).
• Potential lack of synthesizability since SAscore is imperfect; SVGP may overestimate variance far from inducing points, biasing toward out-of-distribution molecules.
• Fixed 7-day runtime evaluation means methods are not trained to convergence, leading to high variance; may miss high-fidelity scaffolds that lower fidelities scored poorly.

Experimental setup

Datasets: ZINC250k BindingDB DUD-E

Benchmarks: ABFE on BRD4(2) and c-MET generated compounds

Metrics: mean ABFE (kcal/mol) top-3 ABFE number of active scaffolds QED SAscore diversity

Baselines: SF-VAE REINVENT RGA VAE + 4x SF-GP VAE + MF-GP MF-AL-GFN MF-AL-PPO MF-GP + ZINC250k Pocket2Mol DecompDiff TAGMol

Builds on: SF-VAE multi-fidelity GP MF-AL-GFN SELFIES deep kernel learning

Closest papers by meaning

• Monroe: A Molecular Foundation Model for In-context Probabilistic Inference (NeurIPS 2026 · similarity 0.88)
• Controllable 3D Molecular Generation for Structure-Based Drug Design Through Bayesian Flow Networks and Gradient Integration (NeurIPS 2025 · similarity 0.86)
• Generating Novel Leads for Drug Discovery Using LLMs with Logical Feedback (AAAI 2024 · similarity 0.86)
• Active Flow Expansion for Out-of-Distribution Discovery: from Theory to Molecules (NeurIPS 2026 · similarity 0.86)
• TFG-Flow: Training-free Guidance in Multimodal Generative Flow (ICLR 2025 · similarity 0.86)
• Navigating Chemical Space with Latent Flows (NeurIPS 2024 · similarity 0.85)