Skip to the content
Browse the library

Every reaction the engine can perform

Each of these is a rule the platform applies to your structures: what bonds have to be there, what they become, and what has to be in the flask for it to run at all. They are the vocabulary behind every predicted outcome.

69 transformations68 with a written mechanism23 classes

A transformation is written as a pattern, not as an example. "Fischer esterification" is not a particular acid and a particular alcohol — it is the statement that a carboxylic acid and an alcohol under acid give an ester and water, and it applies to whichever acid and alcohol you bring.

What makes these usable rather than decorative is the conditions attached to each. A rule that fires on anything shaped correctly would propose oxidations with no oxidant and couplings with no catalyst. These say what they need, and withhold themselves when it is not there.

How these are kept honest

Every transformation is checked against reactions with a known answer before it ships: a set that asks whether the engine gets the right regiochemistry, the right chemoselectivity and the right answer when the answer is "nothing happens". It currently passes 54 of 54, alongside a blind set of 32 reactions drawn from families the engine must not confuse with one another.

Those are the numbers worth quoting. Out-of-sample top-1 accuracy — how often the single recorded product of a real experiment is our first candidate — is 55.8%, and it is a weaker guide, because several proposals are often chemically reasonable and only one was run. The questions page goes through what each number does and does not mean.

Questions people ask

What is a transformation, as opposed to a reaction record?

A transformation is a rule: a pattern of bonds that becomes another pattern of bonds. It is written once and applies to any substrate that matches it. A record is one experiment somebody ran. The rule tells you what would happen; the record tells you what did.

Why does a transformation sometimes refuse to fire?

Usually because what it needs is not present. An oxidation needs an oxidant, a Suzuki needs a transition metal, and an alcohol drawn on its own does not oxidise itself. Rather than quietly producing the product anyway, the engine reports the requirement by name.

Are the typical conditions a recommendation?

No. They are what the literature usually does, written down so you can see whether your conditions are in the ordinary range. They are not a procedure, they are not scaled to your flask, and they do not replace the primary source or a risk assessment.

Does a transformation predict yield?

No. Nothing on this platform predicts yield. The engine ranks candidate outcomes by how much documented precedent each has on substrates like yours, and reports that count rather than converting it into a number that looks like a probability.