Research map / Probabilistic and scientific modeling / Bayesian optimization

Multi-Objective Molecular Design Through Learning Latent Pareto Set

AAAI 2025 · Probabilistic and scientific modeling › Bayesian optimization · approach: Multi-objective Bayesian optimization · 3 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 MLPS, which learns a global preference-to-Pareto-set neural network over a latent molecular space combined with local GP-based trust-region Bayesian optimization, for multi-objective molecular design.

Problem

Multi-objective molecular design requires approximating the full Pareto set of expensive, conflicting molecular objectives in a high-dimensional discrete chemical space, while existing methods return only a limited set of trade-off molecules.

Approach

Learn a global neural network mapping user preference vectors to Pareto-optimal latent molecular embeddings, trained with surrogate losses from local GP surrogates in multiple trust regions (centers chosen by hypervolume contribution), so global and local search share information.

Key results

• MLPS achieves the highest hypervolume in all five multi-objective molecular design tasks, e.g., 0.922 on QED+SA and 0.714 on the four-objective task, versus 0.907 and 0.578 for the best baselines HN-GFN and MARS respectively.
• MLPS achieves the best product metric (SR x Nov x Div) in all five tasks, e.g., 0.789 on QED+SA versus 0.782 for HN-GFN.
• Ablation on the four-objective task shows that removing either the global or local model reduces hypervolume relative to full MLPS.

Measured effects reported in the paper

Surprising findings

• The PSL baseline, adapted to the molecular encoder-decoder, performs very poorly on the product metric (around 0.000-0.002), indicating that a pure global Pareto set learner lacks local search.

Limitations

• Scalability to a larger number of objectives is left for future work.
• Real-world case study uses surrogate predictors (Chemprop, ToxinPred3.0) rather than wet-lab validation.

Experimental setup

Datasets: 5000 antimicrobial peptides (antifungal case study)

Benchmarks: QED+SA GSK3β+JNK3 QED+SA+GSK3β QED+SA+JNK3 QED+SA+GSK3β+JNK3 antifungal peptide activity and toxicity

Metrics: hypervolume success rate novelty diversity product metric

Baselines: GA+D JT-VAE GCPN RationaleRL MARS MolSearch PSL RetMol HN-GFN

Builds on: Turbo PSL P-MOCO SELFIES

Closest papers by meaning

• Molecule Design by Latent Prompt Transformer (NeurIPS 2024 · similarity 0.88)
• Uncertainty-Aware Multi-Objective Reinforcement Learning-Guided Diffusion Models for 3D De Novo Molecular Design (NeurIPS 2025 · similarity 0.88)
• Enhancing Ligand Validity and Affinity in Structure-Based Drug Design with Multi-Reward Optimization (ICML 2025 · similarity 0.88)
• Searching for High-Value Molecules Using Reinforcement Learning and Transformers (ICLR 2024 · similarity 0.87)
• Efficient Evolutionary Search Over Chemical Space with Large Language Models (ICLR 2025 · similarity 0.86)
• ProteinOPD: Towards Effective and Efficient Preference Alignment for Protein Design (NeurIPS 2026 · similarity 0.86)