PREMISE

Recently, plant science has seen transformative advances in scalable data collection for sequence and chemical data. These large datasets, combined with machine learning, have demonstrated that conducting plant metabolic research on large scales yields remarkable insights. A key next step in increasing scale has been revealed with the advent of accessible large language models, which, even in their early stages, can distill structured data from the literature. This brings us closer to creating specialized databases that consolidate virtually all published knowledge on a topic.

METHODS

Here, we first test different combinations of prompt engineering techniques and language models in the identification of validated enzyme-product pairs. Next, we evaluate the application of automated prompt engineering and retrieval-augmented generation to identify compound-species associations. Finally, we build and determine the accuracy of a multimodal language model-based pipeline that transcribes images of tables into machine-readable formats.

RESULTS

When tuned for each specific task, these methods perform with high (80-90%) or modest (50%) accuracies for enzyme-product pair identification and table image transcription, but with lower false-negative rates than previous methods (decreasing from 55% to 40%) for compound-species pair identification.

DISCUSSION

We enumerate several suggestions for researchers working with language models, among which is the importance of the user's domain-specific expertise and knowledge.