concept · intermediate · partly written
Chemistry-trained reaction models applied to enzymes and metabolism: what was measured
Models pretrained on patent chemistry help on enzymatic and metabolite prediction when mixed with enzyme data, and most of the evidence is the authors' own benchmarks.
Also called: ECREACT, EnzymeMap, BioNavi-NP, enzymatic transformer, SyGMa, GLORYx, MetaTrans
Does chemistry-trained prediction transfer to enzymes?
Yes, partly. Patent-only models score about 10% on enzyme tasks, mixed training reaches about 50%, and enzyme text lifts it to 62.2%. A natural-product pathway model reached top-10 of 60.6% with general-chemistry pretraining and 0 without it on one subset. Nearly every benchmark was run by its own authors. 12
Reported, not yet replicated
How was the enzymatic transformer tested? more
Kreutter, Schwaller and Reymond trained a transformer on reactions from Reaxys with the enzyme described in words. Without enzyme information, patent-only training scored about 10%, enzyme-only about 20%, and mixed training about 50%. With the full enzyme description top-1 was 62.2%, with only the "-ase" words 54.0%, and with scrambled descriptions about 15%. On 43 cytochrome P450 test reactions it scored 40%. The enzyme data come from a paid database and cannot be shared, and the model carries no rate information. 1
Reported, not yet replicated
What does the open ECREACT benchmark show? more
ECREACT combines 62,222 reaction-EC pairs from public sources (Rhea, BRENDA, PathBank, MetaNetX) with about one million USPTO reactions. Forward top-1/5/10 were 49.6/63.5/68.8%, and randomising the EC tokens gave 41.3% within class and 38.3% across classes, so the enzyme class adds about 8 to 11 points. An EC number is the four-part Enzyme Commission code for the reaction an enzyme catalyses and carries no sequence. Ligases scored 8.1% when EC tokens were scrambled across classes. 3
Reported, not yet replicated
What does the repaired BRENDA data give? more
EnzymeMap atom-mapped 63,137 BRENDA reactions and found 8,599 unbalanced entries. Retrosynthesis top-1 was 0.46 against 0.28 for MetAMDB and 0.18 for Rhea on the authors' own test sets. Adding it lifted forward top-1 from 0.49 to 0.54 and lowered backward top-1 from 0.60 to 0.55. The same paper reports BRENDA recording a ketone as an aldehyde for EC 1.1.1.103 (L-threonine 3-dehydrogenase) hundreds of times. 4
Reported, not yet replicated
How good are drug-metabolite predictors? more
On 37 parent drugs, GLORYx had recall 0.77 and precision 0.061, and the rule-based SyGMa had recall 0.68 and precision 0.12. MetaTrans, pretrained on about 900,000 patent reactions, reached top-10 recall 57.5 against 58.7 for SyGMa. In each case the tool's own authors ran the test, and the GLORYx paper says neither tool has precision high enough to use without ranking. 56
Reported, not yet replicated
Where does the transfer break? more
Enzymatic models see an EC code, not a sequence, so they cannot tell two enzymes of the same EC apart. They predict the same product for a slow engineered enzyme because the training data hold no rates. Turnover-number predictors fail on distant enzymes: below 60% sequence identity to training, DLKcat's R2 is negative, and 67.9% of its test enzymes are in its training set. That critique came from a group that built a rival model. 137
Reported, not yet replicated
Which chemistry method has transferred best? more
Component contribution, which estimates reaction Gibbs energy from molecular fragments, reached RMSE 2.7 kJ/mol against an E. coli dataset and 3.1 for the human Recon 1 network. The fully covered reactions are 9.4% of E. coli and 8.3% of Recon 1 reactions, but they carry about half of E. coli flux and about a third of human flux. This is a chemistry method validated on biological data, not an industry dataset. 8
Reported, not yet replicated
How well do rule-based pathway planners perform? more
In the Galaxy-SynBioCAD benchmark, 83% of benchmark targets had a validated pathway in the top 10, against 58% for a simple baseline score. In the wet lab 30 and 33 of 88 constructs worked at one condition, with 12 in common. The benchmark and the lab work were run by the groups that built the tool. 9
Reported, not yet replicated
Plain answer
A model trained only on patent chemistry scores about 10% on enzyme reactions, but mixed with enzyme data it reaches about 50% without enzyme information and 62.2% with enzyme text. Scrambling the enzyme text drops it to about 15%, so the enzyme text carries real signal, while most of the lift comes from the chemistry. 1
Reported, not yet replicated
Why it matters
Still open here
- Does a chemistry-pretrained model beat an enzyme-only model on a sequence-clustered split? Not found.
- Has any route proposed by a learned model been run in a lab? None was found in the sources read; abstracts for the rule-based RetroBioCat tool leave an external lab use unconfirmed.
Not yet written: Equation, Real numbers.
Linked from
Sources
- David Kreutter, Philippe Schwaller, Jean-Louis Reymond. Predicting enzymatic reactions with a molecular transformer. Chemical Science (2021). paper; opened 2026-11-17 all records citing this
- Daniel Probst, et al.. Biocatalysed synthesis planning using data-driven learning. Nature Communications (2022). paper; opened 2026-11-17 all records citing this
- Esther Heid, et al.. EnzymeMap: curation, validation and data-driven prediction of enzymatic reactions. Chemical Science (2023). paper; opened 2026-11-17 all records citing this
- Christina de Bruyn Kops, et al.. GLORYx: Prediction of the Metabolites Resulting from Phase 1 and Phase 2 Biotransformations of Xenobiotics. Chemical Research in Toxicology (2021). paper; opened 2026-11-17 all records citing this
- Litsa, Das, Kavraki. MetaTrans: transformer-based prediction of drug metabolites [title not copied from the page]. journal not copied (2021). paper; opened 2026-11-17 all records citing this
- Alexander Kroll, Martin J. Lercher. Critique of a deep-learning kcat predictor, with a rival model [title not copied from the page]. journal not copied (2024). paper; opened 2026-11-17 all records citing this
- Elad Noor, Arren Bar-Even, Ron Milo, et al.. Consistent estimation of Gibbs energy using component contributions. PLoS Computational Biology (2013). paper; opened 2026-11-17 all records citing this
- Faulon and Carbonell groups. Galaxy-SynBioCAD: automated pathway design workflows [title not copied from the page]. Nature Communications (2022). paper; opened 2026-11-17 all records citing this
Where it sits on the map
History of this record (1)
This record has been changed 1 time, on the dates below. The note kept against each change is written for the people who maintain the atlas, in their own shorthand, so it is not shown here; it is in the record itself.
- Status
- partly written
- Last updated
- 2026-11-17
- Record
- content/concepts/enzymatic-reaction-prediction-and-transfer-from-chemistry.json