How the platform works, and how much of it to believe
The questions people actually ask, answered without rounding anything in our favour. Where a number is unflattering it is here too, because a tool you cannot check is a tool you cannot use.
21 questionsLast reviewed September 2026
Most of these come down to one question: how much of this should I believe? The answer differs by surface, which is why the platform labels every answer with the kind of thing it is. A record is evidence. A rule-derived outcome is a deduction. A mechanism is a teaching pathway. None of them is dressed up as another.
What this is
- What is ReactionLens?
- A chemistry workstation with four surfaces: look a structure up and get its properties with the origin of each value; search reactions that were actually run and read their conditions; ask what a set of reactants would give and see the evidence behind each candidate; and enumerate analogues of a lead compound. The common thread is that every number is labelled with where it came from and how it was obtained.
- Who is it for?
- Synthetic and medicinal chemists deciding what to run next, and students trying to understand why a reaction goes the way it does. If you need a number you can defend in a meeting or a viva, the provenance on every value is the point of the product.
- How is this different from a search engine or a reaction database?
- A database tells you what was recorded. A model tells you what it guesses. This does both and keeps them apart: records are labelled verified, rule-derived outcomes are labelled rule-derived, and a model's vote is reported as a vote rather than folded into a single score. You are never left unable to tell which kind of answer you are looking at.
- Do I need an account?
- Not to explore. Structures, reaction records, mechanisms, the calculators and the whole public library are open without signing in. Accounts are not open during the beta: prediction, analogue design and quantum-chemical jobs belong to a workspace, because they consume compute and are counted against an allowance, so they open when accounts do.
Prediction, and how good it actually is
The honest version, including the numbers that are less flattering than the ones we could have quoted.
- How accurate is the outcome prediction?
- On 394 reactions the engine was never trained or tuned on, the product that was actually run and published is our first candidate 56.5% of the time, and in our top five 59.3% of the time. The validation page sets out every figure and what it measures.
- Why is 56.5% not the same as being right 56.5% of the time?
- Because the test asks whether we rank first the one product that happened to be run and written down. Several proposals are often chemically reasonable, and only one was performed. A candidate we rank second may be perfectly correct chemistry that nobody happened to try. The number is a ranking measure, not a correctness rate.
- Then what should I use to judge it?
- The suites that ask whether an answer is wrong rather than whether it is first. A correctness audit of 54 reactions with a known right answer — regiochemistry, chemoselectivity, stoichiometry, and cases where the correct answer is that nothing happens — currently passes 54 of 54. A blind set of 32 reactions drawn from families the engine must not confuse passes 32 of 32. Those are the better guide.
- Does it predict yield?
- No. Nothing here predicts yield, and we would rather say so plainly than produce a number that looks authoritative. Where a yield appears it was measured by whoever ran the experiment and is shown as their measurement.
- What is the score next to a candidate?
- It measures how much verified precedent that transformation has, especially on substrates like yours. It is not a calibrated probability that the reaction works, and the interface says so wherever it appears. A high score means a lot of documented chemistry does this; it does not mean your reaction will.
- What chemistry is outside the validated domain?
- Photochemistry, electrochemistry, organometallic steps beyond the four catalytic cycles that are written out, solid-phase and flow chemistry, and anything where the solvent or the counter-ion decides the outcome. Stereochemistry is predicted only where a transformation states it on its own evidence; everywhere else the product is marked as having unspecified geometry rather than implying it was evaluated.
Reaction records and evidence
- Where do the reaction records come from?
- Public reaction datasets, principally the Open Reaction Database, with the source recorded on every record and a link back to it. The attribution and licences are listed on the sources page.
- Why do some records not show the written procedure?
- Because the licence of the source does not allow us to republish the text verbatim. The record says the procedure is withheld and links to the original rather than paraphrasing it into something that might be wrong.
- What does 'verified' mean on a record?
- That the record came from a source that published it as an experiment performed, not that we repeated it. It is a statement about provenance, not about reproducibility.
- A reaction I searched for shows a similar record instead. Why?
- Because no record matches your exact reactants, and rather than return nothing we show reactions that perform the same transformation on different molecules. Those are labelled as analogues, with a similarity figure, and where the figure is low the record is marked as a lead rather than as support.
- What is transformation similarity, as opposed to structural similarity?
- Structural similarity compares the molecules. Transformation similarity compares the bond changes — what breaks, what forms, what changes order — which is what tells you whether a record is evidence for the reaction you are asking about. A record can be 95% similar in transformation and involve molecules that look nothing like yours. That is usually the more useful of the two.
Mechanisms
- Are the mechanisms proven?
- No, and every mechanism screen says so. A mechanism here is the canonical pathway taught for its class, run on the structures you supply. It is rule-derived. It is kept separate from the reaction record, which is the part that is evidence.
- Why does my reaction show no mechanism?
- Either no pathway has been written for that class, or the written one does not reach your product from your structures. A pathway that does not arrive is not that reaction's pathway, so nothing is shown. Twenty-six classes are covered, and refusing the rest is deliberate.
- Can I see the curly arrows?
- Yes. Each step declares its arrows, and they are resolved to the actual atoms in your molecules and animated on them. Press play and the pathway runs through its steps in order. How long a step takes on screen is not a rate — every step is given the same time, and where relative speed matters it is stated in words.
Safety, data and access
- Is there safety screening?
- Every structure is checked against the safety policy before any engine is offered. A restricted match stops the request outright; a broader structural alert is reported in those words. This is a screen, not a risk assessment, and it does not replace reading the primary source or following your institution's rules.
- Who can see my work?
- Projects, saved routes and calculations belong to a workspace. Everyone you invite to that workspace sees them, and nobody outside it does. Molecules and reactions you enter are different: they join the shared record every user searches, as the privacy page explains.
- What does it cost?
- Nothing during the public beta. The plans page lists what each plan allows, and nothing is billed.
Something not answered here
The explainer pages go further into the chemistry, and the sources page names the databases behind the data and their licences. If you want to see the behaviour rather than read about it, the Reaction Lab needs no account.