Bioactivity data extraction: Accuracy of unit prefixes?

[flange-ipb]

Hello,

I recently read the paper Finding the dark matter: Large language model-based enzyme kinetic data extractor and its validation (doi:10.1002/pro.70251) where the authors discuss the detection of the unit prefixes "µ" and "m" in their EnzyExtract pipeline (page 7). I'm wondering whether BioChemInsight's bioactivity extraction pipeline involving PaddleOCR takes this into account or whether there are any specific problems with unit prefix detection. Maybe this is something you could discuss in the paper.

[dahuilangda]

Hi @flange-ipb

First of all, my apologies for the very late reply — my Gmail storage was full over the past few months, so I didn’t actually receive your email/notification and only saw your message recently.

Thanks a lot for the pointer to “Finding the dark matter: Large language model-based enzyme kinetic data extractor and its validation” and for raising the question about unit prefixes like “µ” vs “m”.

In BioChemInsight, PaddleOCR is only used as a front-end to recover the raw text from figures/tables. The actual parsing and normalization of bioactivity values (including units such as µM, nM, mM, etc.) is handled by a large language model. This means that the distinction between “µ” and “m” is mainly resolved at the LLM layer based on the local textual and numerical context (for example, typical value ranges for IC₅₀/Kᵢ/Kd, neighboring units, and surrounding wording), rather than by hand-crafted prefix rules.

So in principle the pipeline can distinguish “µM” from “mM”, and in our qualitative tests we haven’t observed a systematic collapse of these unit prefixes. That said, we haven’t yet done a dedicated, EnzyExtract-style benchmark focused specifically on µ vs m confusion, and we are still limited by the quality of the OCR output (e.g. badly embedded fonts where “µ” is dropped or misrendered).

One important caveat: BioChemInsight is currently optimized for small-molecule / new-drug patent data and chemical–bioactivity extraction, which is the main scenario we had in mind when designing and tuning the pipeline. For more general enzyme kinetic data across the broader literature, the performance may be weaker and we haven’t systematically validated those use cases yet.

Your comment is very helpful – in the next revision of the manuscript I plan to add a short note in the methods/limitations section to explicitly discuss unit-prefix handling and this potential failure mode.

Best regards,
Zhe