Biology Atlas

Map / Open register

question · intermediate · partly written

Does chemistry pretraining still help enzyme-reaction prediction when the test enzymes are unlike the training enzymes?

Whether the measured gain from patent-chemistry pretraining survives a split by enzyme family or sequence cluster is untested.

What do the measured gains show?

Pretraining on patent chemistry measurably helped on random splits: about 10% patent-only, about 20% enzyme-only and about 50% mixed in one study. Whether it helps on enzymes unlike the training set was not found. 12

Reported, not yet replicated

Why is the clustered split the missing test? more

Near-duplicate leakage inflated chemistry benchmarks (Molecular Transformer top-1 90.4% fell to 74.6% on a similarity split), and turnover-number prediction collapses below 60% enzyme sequence identity. The same leakage may be inflating the enzymatic numbers, which is an inference, not a measurement. 34

Our inference

Observation

Models trained on patent chemistry plus enzyme data beat enzyme-only models on random or product-disjoint splits. 125

Reported, not yet replicated

What is known

Mixed training reached about 50% against about 20% for enzyme-only data in one study, and BioNavi-NP reached top-10 of 60.6% against 27.8% for biochemical data alone. 15

Reported, not yet replicated

What is not

Nobody opened here reports the comparison on a split by enzyme sequence cluster, and the enzymatic models read use EC codes, not sequences. 24

Our inference

What would settle it

Re-split ECREACT by product and enzyme family, and report accuracy against the 38.3% obtained with scrambled EC tokens and against a nearest-precedent lookup. This is our own suggestion, using public data only. 26

Our inference

Who works on it

Still open here

Not yet written: Hypotheses, What is missing.

Sources

  1. 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
  2. Daniel Probst, et al.. Biocatalysed synthesis planning using data-driven learning. Nature Communications (2022). paper; opened 2026-11-17 all records citing this
  3. Kovacs et al.. Quantitative interpretation explains machine learning models for chemical reaction prediction and uncovers bias. Nature Communications (2021). paper; opened 2026-11-17 all records citing this
  4. 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
  5. Shuangjia Zheng, et al.. Deep learning driven biosynthetic pathways navigation for natural products with BioNavi-NP. Nature Communications (2022). paper; opened 2026-11-17 all records citing this
  6. Connor W. Coley, et al.. Computer-assisted retrosynthesis based on molecular similarity. ACS Central Science (2017). 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.

Kind
big-open
Register
open-question
Triage
needs-research
Status
partly written
Last updated
2026-11-17
Record
content/questions/does-chemistry-pretraining-beat-enzyme-only-models-on-clustered-splits.json