[Paper Review] MolLIBRA: Genetic Molecular Optimization with Multi-Fingerprint Surrogates and Text-Molecule Aligned Critic

Masahi OKADA, Kazuki Sakai+3 more|arXiv (Cornell University)|Jan 29, 2026

Machine Learning in Materials Science0 citations

TL;DR

Mol LIBRA introduces a multi-fingerprint GP surrogate ensemble plus a zero-shot CLAMP-based critic to guide a GA for sample-efficient molecular optimization under tight oracle budgets, achieving strong Top-10 AUC on PMO-1K tasks.

ABSTRACT

We study sample-efficient molecular optimization under a limited budget of oracle evaluations. We propose MolLIBRA (MultimOdaLity and Language Integrated Bayesian and evolutionaRy optimizAtion), a genetic algorithm based framework that pre-ranks candidate molecules using multiple critics before oracle calls: (i) an ensemble of Gaussian process (GP) surrogates defined over multiple molecular fingerprints and (ii) a pretrained text-molecule aligned encoder CLAMP. The GP ensemble enables adaptive selection of task-appropriate fingerprints, while CLAMP provides a zero-shot scoring signal from task descriptions by measuring the similarity between molecular and text embeddings. On the Practical Molecular Optimization (PMO) benchmark with a budget of 1,000 evaluations (PMO-1K), MolLIBRA-L, our variant with a language-model-based candidate generator, attains the best Top-10 AUC on 14/22 tasks and the highest overall sum of Top-10 AUC across tasks among prior methods.

Motivation & Objective

• Address sample-efficient molecular optimization under a limited oracle-evaluation budget.
• Reduce sensitivity to fingerprint choice by using a multi-fingerprint GP surrogate ensemble.
• Leverage a text-molecule aligned zero-shot critic (CLAMP) to warm-start and guide candidate ranking without scored data.
• Demonstrate effectiveness on the Practical Molecular Optimization (PMO) benchmark under 1,000 evaluations (PMO-1K).

Proposed method

• Use a GA-based molecular optimizer that pre-ranks candidates with multiple critics before oracle evaluation.
• Construct an ensemble of Gaussian process surrogates defined over six fingerprint types (ECFP, FCFP, Avalon, Pharmacophore, MAP, BoC) with a Tanimoto kernel.
• Incorporate a zero-shot CLAMP critic that scores molecules via text–molecule embedding similarity to the task description.
• Select critics probabilistically and update selection weights online as new oracle data arrive.
• Pre-evaluate and rank candidates using either CLAMP or one of the GP surrogates, then perform oracle evaluations on the top batch.
• Candidates are generated via a BiG (Graph GA) or LLM-guided editing (Mol LIBRA-L) components.

Experimental results

Research questions

• RQ1 Can a multi-fingerprint GP surrogate ensemble reduce performance sensitivity to fingerprint choice in low-budget molecular optimization?
• RQ2 Does integrating a zero-shot text–molecule critic (CLAMP) improve early ranking and sample efficiency before sufficient labeled data are available?
• RQ3 How does Mol LIBRA perform on the PMO-1K benchmark compared to state-of-the-art baselines?
• RQ4 What is the contribution of model weighting and critic selection to overall optimization performance?
• RQ5 Does combining GA-based generation with language-model-assisted editing (Mol LIBRA-L) yield performance gains?

Key findings

• Mol LIBRA variants achieve strong performance on PMO-1K, with Mol LIBRA-L obtaining the best overall Top-10 AUC sum across tasks.
• The multi-fingerprint GP surrogate ensemble reduces sensitivity to fingerprint choice and enhances robustness under budget constraints.
• Incorporating CLAMP as a zero-shot critic provides valuable early ranking signal, improving pre-evaluation decisions when scored data are scarce.
• Mol LIBRA-L (language-model-based candidate generation) outperforms several baselines on a majority of PMO-1K tasks, ranking top on 14/22 tasks in Table 1.
• Ablation studies show the contributions of multi-fingerprint surrogates and CLAMP in the overall performance gains (Mol LIBRA-G and Mol LIBRA-L).
• Compared to Tripp’s GP BO and LLM-based baselines, Mol LIBRA variants consistently demonstrate competitive or superior performance under the 1,000-evaluation budget.