Reaction-rate model trained only on simple molecules gets all six complex test cases right
Yokohama National University chemists trained a reaction-rate model only on simple molecules and got the right site and side in all six complex test cases. Drug candidates and natural products are the team's stated targets for the method.
The Scientist · Science desk
Key figures: about 1.6% chance that a random guess picks the reacting site in all six complex test molecules. Each test molecule has 2 ketone groups. The model correctly predicted the main site and side in all 6 test molecules.
| Figure | Value | Note | Claim |
|---|---|---|---|
| Chance a random site guess gets all six right, about | 1.6% | 14 | |
| Ketones per test molecule | 2 groups | 6 | |
| Test molecules predicted | 6 all correct | 5 |
What happened
- Yield is the usual measure of reactivity, and it cannot show which ketone group in a molecule reacted or from which side.
- The team mixed complex molecules with varying amounts of sodium borohydride in methanol and recorded which ketone group reacted first, building a reaction-rate map.
- They combined that rate map with three-dimensional electronic information measured around each reacting site to build the prediction model.
Why it matters
- What doesn't add up contradiction Phys.org's headline calls the approach transferable, but the paper's title names sodium borohydride in methanol, so the transfer on show runs from simple to complex molecules.
- What it limits constraint With two ketone groups in every test molecule, molecules with three or more reactive sites, where the site call has more options, sit outside anything reported.
- What's now possible capability The curated kinetic database and descriptor workflow are reported for reuse, so other groups can build their own models and test mechanistic hypotheses on the team's rate data.
In our view, the competition design is the stronger half of the paper. The ketone groups in one molecule face the same reagent in the same mixture, so each molecule is its own control [3]. Yields are a weaker comparison: they differ for the same molecule across synthesis scenarios, and purification and competing side reactions feed into them [13].
Six is the denominator [5]. Each test molecule had two ketone groups [6], so the site call is a two-way choice, and a coin flip would get all six right about once in 64 tries, or 1.6% [14]. The thing this doesn't tell you is how much the model adds over a simple baseline: Phys.org's account does not mention one, or the number of simple molecules used for training [11].
The study is in JACS Au [1]. Daimon Sakaguchi, a doctoral student and first author, said: "We found that simple competition experiments can produce reaction-rate data that connect laboratory measurements, quantum-chemical calculations and machine learning," [9]
Hiroaki Gotoh, an associate professor in the Faculty of Engineering and the corresponding author, set out the aim [12]. He said: "Our ultimate goal is to predict where and from which side drug candidates, natural products and functional molecules will react before synthesis. This will help reduce trial and error and accelerate the discovery of new molecules and synthetic routes." [8]
What to watch
- Whether the team or another lab tests the model with a reagent or solvent other than sodium borohydride in methanol.
- Whether other groups build models on the released kinetic database and report how they compare with this one.
- Whether a test set includes molecules with more than two ketone groups, or actual drug candidates or natural products.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence40
- Adoption
- Not scored
- Hype gap+20
- Incentives
- Not scored
- Confidence45
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
A study describing the modeling work was published in JACS Au.
- [2]
Reactivity is typically measured by product yields after multistep synthesis processes, and product yield measurements do not show which ketone group reacted or from which side.
- [3]
The researchers mixed the complex molecules with varying amounts of sodium borohydride in methanol and measured which ketone group reacted first, which let them build a kinetic scale (reaction-rate map) of the ketone groups.
- [4]
The researchers combined the kinetic scale with three-dimensional electronic information measured from around each reacting site to develop a model that predicts chemical reaction rates.
- [5]
The model was trained only on simple molecules, but it correctly predicted both the main reaction site and the side from which the reaction occurred in all six more complex molecules.
- [6]
Each of the six more complex test molecules contained two ketone groups.
- [7]
The paper is titled 'Competition-Derived Relative Reactivity and 3D Electronic-State Analysis of Site- and Facial Selectivity in NaBH4/MeOH Ketone Reductions'. Phys.org's headline for its story calls the work a 'Transferable chemical prediction approach'.
- [8]
"Our ultimate goal is to predict where and from which side drug candidates, natural products and functional molecules will react before synthesis. This will help reduce trial and error and accelerate the discovery of new molecules and synthetic routes."
- [9]
"We found that simple competition experiments can produce reaction-rate data that connect laboratory measurements, quantum-chemical calculations and machine learning,"
- [10]
The authors report a curated kinetic database and descriptor workflow that, in Sakaguchi's words, provide a foundation for researchers to perform further analyses, construct new predictive models and test mechanistic hypotheses for complex reduction models.
- [11]
Phys.org's account of the study does not state how many simple molecules were used for training or report a comparison of the model with any simpler baseline.
- [12]
Hiroaki Gotoh is the corresponding author and an associate professor in the Faculty of Engineering at Yokohama National University; Daimon Sakaguchi is the first author and a doctoral student in the Gotoh laboratory.
- [13]
Multistep synthesis processes frequently involve purification and competing side reactions that affect product yield, and yield differs across the same molecules used in different synthesis scenarios.
- [14]
With two ketone groups per test molecule, a random guess of the reacting site has a 1-in-2 chance per molecule, so guessing all six correctly has probability (1/2)^6 = 1/64, about 1.6%.
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
1 independent publisher whose own reporting we read for this story.
- phys.orgTransferable chemical prediction approach may help reduce trial and error
1 article · October 16, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.