One of the purposes of the organic reaction dataset is in computer assisted retrosynthetic planning (CASP). Several ventures exist and some of them also open-source or open-access. In 2021 this blog had a look at RXN for chemistry hosted by company IBM, a website that is free to use. Askcos is also a website and maintained by MIT. Aizynthfinder, developed and maintained by company AstraZeneca is an open-source tool that can be run on the command-line. It would be nice to pit these ventures against the humans and these humans can be represented by a selection of those chemists who have contributed to the chemical reaction database in 2025. This 2025 batch is about 80K large and picking one target at random (from the USPTO dataset) and assuming the reaction is new and unique (and not already been used in a model training) is trivial. It is time for a game of humans against algorithms!

First up as retrosynthesis target is 6,8-dichloropyrido[2,3-b]pyrazin-2-ol (details here) that if it was up to the humans is best synthesised from the aromatic diamine and ethyl 2-oxoacetate in a single step. If this was a horse race then IBM would have tripped over de chalk of the starting line because in default mode it reports a prediction that is identical to the target with a confidence of zero In AI mode IBM fares better with top suggestion a methoxy group hydrolysis (0.37 confidence). The human choice comes in at suggestion number 4 with 0.091 confidence. Askcos gets it right the first time and confirms the pyrazine synthesis from ethyl 2-oxoacetate with an average plausibility score of 1 and an atom economy of 0.77. The tree explorer has a total of 200 trees and the site allows you to download the all of them in a json file. I must say the results after tree number 5 strike me as very implausible. The Aizynthfinder tool lives on the command line and yields 7 trees, the top suggestion of which with a score of 0.96 is included in the CRD database as synthetic data and very different from all other suggestions with a conversion first to an amine oxide and then benzyl alcohol transferring oxygen (not very likely?). The human synthesis comes in as number 2. Of course each of the tools comes with many settings en plenty opportunities to tweak results and in this test we try to stick to the default settings. Let's try another one.

For the second challenge we look at a Suzuki reaction here. In the human design the coupling reaction goes before any BOC-protection, N-methylation or fluorination. The IBM prediction however ranks BOC protection first, then N-methylation and then methyl group to carboxylic acid oxidation and only then a coupling. It seems decorating is valued more than deconstruction. Askcos does understand deconstruction but elects to start at the other end of the molecule with a Buchwald-Hartwig coupling. The proper coupling reaction is tree number two but only specifies the coupling of an aromatic chloride and an aromatic bromide whereas in the human design a boronic acid and a triflate are specified. Aizynthfinder is almost spot-on: it correctly predicts the boronic acid and only has the aryl triflate replaced by the aryl chloride.

Let's try another one. In the synthesis of tert-butyl 7-(5-(2-amino-4-fluorophenoxy)pyrimidin-4-yl)-2,7-diazaspiro[4.4]nonane-2-carboxylate the aromatic bromine is replaced by an amino group only as the last step (graph here), a decision that the humans involved have no doubt given plenty of consideration but we are not goingto ask them about it. Will the algoritms fall into the trap? IBM certainly does, of the max 7 trees the first 6 suggest reducing nitro or azido groups or several amine deprotections, Only at suggestion number 7, the aromatic bromide gets the spotlight. The reaction conditions yell Buchwald-Hartwig when the actual reaction conditions whisper sodium azide and sodium ascorbate but this is contest of starting material and not of chemical process. If you ask Kos, the first tree suggests a Buchwald-Hartwig deconstruction and the human preference does not make it to the top-20 trees too busy replacing a benzyl protecting group for a BOC protecting group. Aizynthfinder is also very focused on an initial Buchwald-Hartwig followed by nitro reductions.

Preliminary conclusions? Of course we assume that only the humans make the rational decisions and of course they have constraints that we do not know of. And in each instance the algorithms give enough food for thought to get to the perfect retrosynthesis. But one to top it off, in this example an alcohol oxidation to a ketone takes place after an indazole cyclization. Askcos is now unable to find any tree but the IBM prediction has this scenario as tree number 2. Interestingly tree number 1 is addition of a tosyl group. The final conclusion must be that the algorithms have no idea what to do with protective groups. They are added to enable certain transformations and removed once the transformations are done. They is no point adding protective groups after the fact as demonstrated in this example and also in the previous examples. Room for improvement!