MoleculeQA: A Dataset to Evaluate Factual Accuracy in Molecular Comprehension

TL;DR

MoleculeQA introduces a taxonomy-guided, large-scale QA benchmark to evaluate factual understanding in molecular language models and to quantify factual bias in generated molecular content. By grounding questions in a three-level domain taxonomy built from authoritative sources and crafting high-quality QA pairs with careful negative sampling and scaffold-based splits, the dataset reveals substantial gaps in current models, particularly for property- and application-related knowledge. The study evaluates a wide spectrum of molecular and general LLMs, showing that multi-modal architectures and larger models offer gains but that factual accuracy remains challenging, even for state-of-the-art systems like GPT-4 in few-shot settings. These findings highlight the need for domain-specific corpora, better multi-modal fusion strategies, and scaling-aware training to improve molecular understanding with real-world impact in chemistry and drug discovery.

Abstract

Large language models are playing an increasingly significant role in molecular research, yet existing models often generate erroneous information, posing challenges to accurate molecular comprehension. Traditional evaluation metrics for generated content fail to assess a model's accuracy in molecular understanding. To rectify the absence of factual evaluation, we present MoleculeQA, a novel question answering (QA) dataset which possesses 62K QA pairs over 23K molecules. Each QA pair, composed of a manual question, a positive option and three negative options, has consistent semantics with a molecular description from authoritative molecular corpus. MoleculeQA is not only the first benchmark for molecular factual bias evaluation but also the largest QA dataset for molecular research. A comprehensive evaluation on MoleculeQA for existing molecular LLMs exposes their deficiencies in specific areas and pinpoints several particularly crucial factors for molecular understanding.

Figures (8)

• Figure 1: The performance of three representative models on the traditional metrics for the molecule caption task (e.g. BLEU etc.) and the factual accuracy metric we defined.
• Figure 2: The process of constructing a molecular domain taxonomy. The procedures involve the selection of the information source, extraction of topics, normalization and structuralization of topics, and hierarchical clustering by domain experts.
• Figure 3: The process of constructing a molecular domain taxonomy. The procedures involve the selection of the information source, extraction of topics, normalization and structuralization of topics, and hierarchical clustering by domain experts.
• Figure 4: An overview of MoleculeQA topics distribution. Four coarse-grained aspects occupy the inner circle, and in the outer circle we list finer-grained non-leaf topics.
• Figure 5: Accuracy of different finer topics under 4 coarse-grained aspects on the MoleculeQA test set. We select BioT5- and T5-base as representatives of Molecular LLM and General LLM, respectively, represented by solid and dashed bars.