Research map / Probabilistic and scientific modeling / Molecules and proteins

Generative Flow Networks for Lead Optimization in Drug Design (Student Abstract)

AAAI 2025 · Probabilistic and scientific modeling › Molecules and proteins · approach: Structure-based 3D molecule generation · 0 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.

Applies GFlowNets with trajectory balance and a GNN policy to fragment-based lead optimization, rewarding Caco2 permeability improvement and similarity to the starting molecule.

Problem

Lead optimization in drug discovery requires exploring large chemical space to improve properties such as cell permeability while retaining the scaffold of a lead compound.

Approach

Formulates lead optimization as sequential fragment-based modification of a complete input molecule, learned by a GFlowNet (GNN policy trained with trajectory balance) that samples diverse modifications proportional to a multi-objective reward combining predicted permeability and similarity to the starting molecule.

Key results

• Pre-trained random forest proxy reached R2 0.72, r 0.83, MAE 0.5, MSE 0.25 on the TDC Caco2 split.
• GFN agent achieved up to 95% molecular validity and 100% uniqueness after 500 epochs on an 8-core CPU.
• Generated modifications improved average predicted logPapp by -1.5 relative to the starting molecule (reported as improvement).

Measured effects reported in the paper

Surprising findings

• Generated molecules often contained charges and complex functional groups, requiring modified reward functions.

Limitations

• Diversity of modifications limited to the fixed fragment library; expanding it (e.g., Enamine) may increase training time.
• Generated molecules contain charges and complex functional groups; no toxicity objective yet.
• Only a single property (Caco2 permeability) is optimized; no baselines compared.

Experimental setup

Datasets: TDC Caco2 permeability (Wang et al.) DeepDelta Caco2 MMP dataset (455 pairs)

Benchmarks: TDC Caco2 permeability

Metrics: R2 Pearson r MAE MSE molecular validity uniqueness average logPapp improvement

Builds on: GFlowNet trajectory balance

Closest papers by meaning

• Pretraining Generative Flow Networks with Inexpensive Rewards for Molecular Graph Generation (ICML 2025 · similarity 0.91)
• Generative Flows on Synthetic Pathway for Drug Design (ICLR 2025 · similarity 0.90)
• SynFlowNet: Design of Diverse and Novel Molecules with Synthesis Constraints (ICLR 2025 · similarity 0.89)
• RGFN: Synthesizable Molecular Generation Using GFlowNets (NeurIPS 2024 · similarity 0.88)
• Multi-domain Distribution Learning for De Novo Drug Design (ICLR 2025 · similarity 0.87)
• Compositional Flows for 3D Molecule and Synthesis Pathway Co-design (ICML 2025 · similarity 0.87)