henhaobio
Sign in

bioRxivSystems BiologyPreprint

Accessing Enzyme Kinetic Data and Prediction Methods at Scale

Saleh AlweriD, Hugues EscoffieriD, Karim Taha, Veda Sheersh BoorlaiD, Han YuiD, Somtirtha Santra, Zechen WangiD, Chidi Egwu, Abraham OsinugaiD, Supantha DeyiD, Vaishnavey Srinivasan Raghunath, Farid ZareiD, Jack McGoldrickiD, Jan-Niklas WederiD, Eduard Kerkhoven, Xiaozhou LuoiD, Costas D. MaranasiD, Liangzhen ZhengiD, Ulrike WittigiD, Ratul ChowdhuryiD, Rajib SahaiD, Nadine TöpferiD, Thomas SauteriD, Ronan M. T. FlemingiD

PDFDOIData
In-depth full-text digest available · Jump to digest ↓

Abstract

Enzyme kinetic parameters inform metabolic models, yet experimental measurements are sparse. A growing body of work predicts them from protein and substrate features, but software fragmentation hinders adoption, so downstream tools lock into the most accessible method. We present OpenKinetics Predictor (at predictor.openkinetics.org), an open-source platform integrating thirteen methods in isolated environments behind one interface. The platform optionally reports similarity between query proteins and each method's training data to contextualise reliability. A common featurisation-prediction abstraction keeps it extensible, and independent parties, including original authors, contributed many methods. We pair it with a data portal (at data.openkinetics.org) that exposes CatLog, a curated kinetic dataset, with precomputed embeddings, predicted binding sites, and standardised splits. Both offer a web interface and an API, and the GECKO modelling toolbox already calls the predictor API. As a case study, we predict across an E. coli model and find inter-predictor agreement varies with metabolic context and data availability.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (28 Sept 2026) · CC BY 4.0

OpenKinetics Predictor unifies thirteen enzyme kinetic predictors behind one interface and API; on E. coli iML1515, eight kcat predictors correlate pairwise at r 0.08 to 0.67.

Each kinetic parameter predictor comes with its own inputs, dependencies and outputs, so modelling workflows tend to adopt one accessible method and rarely compare it with others or upgrade. This work adds a standardised multi-method platform, a curated data portal with shared features and splits, and a genome-scale analysis of agreement that shows where the choice of predictor most changes model parameters.

Key results

  1. ShownA single interface wraps thirteen published kinetic predictors (kcat, KM, kcat/KM) with input validation, per-method isolated virtual environments, cached protein embeddings, optional measured-value substitution, and per-query similarity to each method's training set; GECKO now calls its API.

    Benchmark on 1,000 reactions with 100 unique protein sequences: runtime spans two orders of magnitude across methods; embedding caching cut repeat-query runtimes from tens of minutes to below one minute in several cases; GPU embeddings ran faster than CPU. MMseqs2 reports max and mean identity per training set. GECKO users can now pick a predictor instead of the hard-coded DLKcat default. · Platform architecture and unified interface; DISCUSSION; Sequence similarity analysis; Tables

  2. ShownAcross iML1515, kcat predictions from eight method configurations differ widely in correlation, centre and spread, so the parameters assigned to the E. coli model depend heavily on which predictor is chosen.

    Pairwise Pearson r on log10 kcat ranged from 0.08 (DLKcat vs TurNuP) to 0.67 (KinForm-L vs KinForm-H, which share architecture and training data); each configuration pair was compared over 3,577 to 4,567 jointly predicted enzyme-reaction pairs. · Predictors disagree substantially across the…; iML1515 case study; QUANTIFICATION AND STATISTICAL ANALYSIS; Figure legends

  3. SuggestedWithin E. coli iML1515, agreement between predictors increases where close experimental data exist and varies by pathway: strong in amino acid metabolism and energy production/conversion, weak or negative for transport and membrane-associated reactions.

    Each configuration's mean r with the other seven peaked at low WILDkCAT penalty scores (0 = exact BRENDA/SABIO-RK match) and fell as scores approached 16 (no measured value); several transport and membrane method pairs had negative r. Category-level r values appear only in figures, not in the text. · Agreement is structured by data availability and…; iML1515 case study; QUANTIFICATION AND STATISTICAL ANALYSIS; Figure legends

  4. Claimeddata.openkinetics.org releases a curated kinetic dataset (CatLog) with precomputed protein features and versioned splits, giving predictor developers a shared basis for training and evaluation without repeated preprocessing.

    CatLog: 50,312 kinetic parameters re-curated from BRENDA and SABIO-RK against primary literature through manual and agentic AI curation; ESMC, ESM2 and ProtT5 embeddings plus Pseq2Sites binding sites for each sequence; random, sequence-exclusive and substrate-exclusive splits in each release. · An open data portal for kinetic data and…; Data portal

  5. ShownAgainst measured kcat for 142 iML1515 pairs, the eight configurations fit in-distribution data moderately; the authors caution that this is not a general accuracy ranking, because the reference values come from the databases most methods trained on.

    On a log10 scale, RMSE ranged from 0.93 to 1.40 and Pearson r from 0.44 to 0.71 across configurations (MAE also reported); references are wild-type kcat values retrieved by WILDkCAT from BRENDA and SABIO-RK; methods built for out-of-distribution prediction are disadvantaged. · Agreement with in-distribution measured values; iML1515 case study; QUANTIFICATION AND STATISTICAL ANALYSIS; Figure legends

Methods, data and reproducibility

A Django/Celery web service and API validates protein-substrate inputs, runs each predictor in its own virtual environment through a two-stage featurisation-prediction abstraction with cached embeddings, and annotates MMseqs2 training-set identity. In the case study, WILDkCAT-derived iML1515 queries are submitted and log10 kcat is compared by Pearson r, RMSE and MAE.

  • Enzyme-reaction pairs from the iML1515 genome-scale metabolic model, each submitted as one kcat query to eight predictor configurations · Escherichia coli · n = 3,577 to 4,567 jointly predicted pairs per configuration pair
  • Measured wild-type kcat values matched to iML1515 pairs, retrieved by WILDkCAT from BRENDA and SABIO-RK · Escherichia coli · n = 142 enzyme-reaction pairs
  • CatLog curated kinetic parameter dataset (CatRange team), hosted on data.openkinetics.org with embeddings, binding sites and splits · n = 50,312 kinetic parameters
  • Runtime benchmark input set behind Table 1 (GPU vs CPU embeddings, with and without cache) · n = 1,000 reactions involving 100 unique protein sequences
Methods & tools
OpenKinetics Predictor (Django, Celery, Redis, SQLite), KinForm-L and KinForm-H, DLKcat, EITLEM, UniKP, CataPro, CatPred, TurNuP, WILDkCAT, MMseqs2, ESMC, ESM2, ProtT5 embeddings; Pseq2Sites binding sites, BRENDA, SABIO-RK, UniProt; GECKO; COBRApy and NumPy
Reagents & environment
Python Django backend with SQLite job store and Celery queue using a Redis broker; each method in its own virtual environment; the open-source release includes container definitions for every component; analysis used NumPy and COBRApy. No versions stated.
Assay conditions & compute
An optional remote GPU service computes protein embeddings over HTTP and falls back to local computation; Table 1 runtimes were measured on 1,000 reactions with 100 unique sequences (GPU or CPU, cached or not). Hardware specifications not stated.
Reproducibility gaps
The GitHub and Zenodo DOIs for code and iML1515 outputs are listed only in a key resources table that is not in the text; CatLog will be described separately; there is no installable local package yet, only the hosted web service and API.
Limitations (6)
  • The case study covers one well-characterised organism (E. coli iML1515); agreement, accuracy and pathway-level patterns may differ in organisms with sparser experimental coverage. · stated by the authors
  • Only eight of the thirteen integrated configurations were analysed, because the rest lack kcat output or were not available when the analysis was run. · stated by the authors
  • Benchmarking against measured kcat only tests fit to familiar data, since those references share a database origin with most training sets; retraining every method on common held-out splits was left outside this paper's scope. · stated by the authors
  • Correlating log-scale predictions tracks shared trends across reactions yet misses consistent scale shifts; predictors built on shared training sources may concur and still be jointly inaccurate. · stated by the authors
  • Runtimes reflect each method's released code rather than its inherent speed; predictions were not validated experimentally, and the effect of predictor disagreement on fluxes, growth or protein allocation was not tested. · stated by the authors
  • The agreement analysis reports no significance tests or confidence intervals, and pair counts differ (3,577 to 4,567) across configuration pairs, so the correlations are computed over non-identical reaction sets. · noted in this summary

Read this if you build enzyme-constrained metabolic models or develop kcat/KM predictors and want one API across many methods, curated training splits, and evidence on where predictors disagree.

Citations
0

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISBoorla '24Gollub '24Kroll '21Zielinski '24Sajeevan '26Morabito '21
72% semantic matchBioinformatics
Veda Sheersh Boorla, Costas D. Maranas·bioRxiv·Mar 10, 2024· 4 citations