Response to “Caution: Fundamental data quality issues underlying intelligent prediction of enzyme optimal pH”

Abstract

We respond to concerns regarding fundamental data quality issues in datasets used for machine learning prediction of enzyme optimal pH (pH_opt). We re-examined six questioned protein entries by tracing their annotated pH values to the original literature. Our analysis confirmed substantial discrepancies for several entries, including incorrect experimental values, misassigned protein identities, and proteins lacking the reported catalytic activity, whereas one entry was consistent with the literature. These findings support the need for rigorous verification of biochemical data underlying predictive models. We emphasize that careful literature based curation is essential for improving dataset reliability and model validity, and future iterations of pH_opt predictors will be developed using more rigorously validated datasets.

To the Editor,

We would like to thank the rigorous dataset quality assessment conducted by the author of ‘Caution: Fundamental Data Quality Issues Underlying Intelligent Prediction of Enzyme Optimal pH’. The dataset of enzyme optimal pH (pH_opt) used to develop EpHod [1], Seq2pHopt [2], and CatOpt [3] was obtained from ‘Machine learning prediction of enzyme optimum pH’, the first enzyme pH_opt predictor, available at https://doi.org/10.5281/zenodo.14252615. In ‘Caution: Fundamental Data Quality Issues…Enzyme Optimal pH’, the author stated that B5LXD7, A0A2Z5VGL1, A0A2Z5VGM9, A0A2Z5VGN0, Q6HA05, and Q97VS7 are erroneously annotated in the dataset of EpHod. Here, we re-examined the annotated pH_opt values and associated references of these data entries (Table 1). For B5LXD7, activity assays in Luo et al. [4] and Xu et al. [5] reported pH_opt values of 1.0–1.5 and 4.0, respectively, both measurements were deviant from the pH_opt value in the EpHod dataset, 11.5. For A0A2Z5VGL1, A0A2Z5VGM9, and A0A2Z5VGN0 that are uniformly annotated with a pH_opt of 2.4, the follow-up experiments showed that these three proteins lacked the catalytic activity of endo-beta-N-acetylglucosaminidase [6], and they are no longer in the reference proteome of Sphingobacterium sp. (species). Therefore, these three data entries should not be used in machine learning model development. Brier et al. reported a pH_opt value of 2.0 for pepsinogen A2 from Trematomus bernacchii [7], however, this value was mistakenly assigned in the EpHod dataset to Q6HA05, which corresponds to pepsinogen A1 from T. bernacchii. Regarding Q97VS7, Huang et al. reported that its pH_opt is ~1.8 at ~80°C [8], consistent with the value in the EpHod dataset.

Table 1.

Selected nine entries from Seq2Topt dataset.

UniProt accession number	pHopt in EpHod dataset	pHopt in literature	Reference
B5LXD7	11.5	1.0–1.5/4.0	[4, 5]
A0A2Z5VGL1	2.4	–	[6]
A0A2Z5VGM9	2.4	–	[6]
A0A2Z5VGN0	2.4	–	[6]
Q6HA05	2	2.5	[7]
Q97VS7	1.8	1.8	[8]

In essence, we share the view that the dataset underpinning enzyme pH_opt predictive models, and more broadly, machine learning models for other protein properties [2, 9, 10], warrants careful scrutiny and refinement to ensure its validity. Rather than relying on the straightforward extraction of data from established repositories such as UniProt [11] and BRENDA [12], we advocate for a more deliberate curation process. Although such curation entails considerable cost, it remains imperative until these biochemical databases can autonomously guarantee a high standard of data integrity [13, 14]. In line with this principle, future iterations of pH_opt predictors (e.g. CatOpt) will be built upon a more rigorously validated dataset.

Key Points

• Literature verification confirms substantial errors in several pH_opt annotations.

• Incorrect protein identities and activities can compromise model development.

• Rigorous biochemical data curation is essential for reliable prediction.

• Future enzyme property datasets should undergo more rigorous validation.

Conflict of interest

The author declares that there is no conflict of interest.

Funding

None declared.

Data availability

No new data was generated in this study. This response letter refers to the data provided in the EpHod study [1].