We demonstrate molecular similarity to be a surprisingly effective metric for proposing and ranking onestep retrosynthetic disconnections based on analogy to precedent reactions. The developed approach mimics the retrosynthetic strategy defined implicitly by a corpus of known reactions without the need to encode any chemical knowledge. Using 40 000 reactions from the patent literature as a knowledge base, the recorded reactants are among the top 10 proposed precursors in 74.1% of 5000 test reactions, providing strong quantitative support for our methodology. Extension of the one-step strategy to multistep pathway planning is demonstrated and discussed for two exemplary drug products.
Alex: Welcome to another episode of ResearchPod. Today we're discussing a study on automating retrosynthesis — the process of working backwards from a complex molecule to figure out how to build it from simpler, commercially available ingredients.
Sam: So the paper is asking whether a computer can plan chemical synthesis by looking at what worked before, rather than being programmed with every single rule of chemistry?
Alex: Exactly. To understand why that matters, think about how traditional software works. It relies on rigid, hand-coded rules — thousands of them — each one telling the computer exactly what to do in a very specific situation. This study proposes something more flexible: using molecular analogy, where the computer learns from past examples rather than following a fixed rulebook.
Sam: That sounds like the difference between memorising every grammar rule versus just reading enough books that you develop a feel for the language.
Alex: That's a useful way to think about it. And the key tool that makes this possible is a way of measuring how similar two molecules are to each other. Imagine you could take any molecule and convert it into something like a fingerprint — a numerical summary of its structure. The system then compares fingerprints to find the closest match in its database of known reactions. The researchers call this measure Tanimoto similarity.
Sam: So if the computer encounters a new, unfamiliar drug molecule, it searches its database for the molecule with the most similar fingerprint, and then borrows the reaction used to make that one?
Alex: Precisely. And it doesn't just copy blindly. It identifies what chemists call the "reaction center" — the specific part of the molecule that actually changes during the reaction. Everything else stays the same. So the system lifts that transformation and applies it to the new target.
Sam: That's a tidy trick. But I can see a potential problem — what if the new molecule has some extra chemical group hanging off it that the original reaction never had to deal with? Couldn't that cause unexpected side reactions?
Alex: The model handles this implicitly, which is one of its more elegant features. If the target molecule has extra parts that weren't present in the precedent, the similarity score between the two automatically drops. A lower score signals to the system that this reaction might not transfer cleanly. So the "distance" between molecules acts as a built-in safety check, without anyone having to manually list every possible conflict.
Sam: So the scoring system is doing double duty — it's both finding good matches and flagging risky ones.
Alex: Right. And it's actually a two-step scoring process. First, the system ranks candidate reactions by how similar the known product is to the new target. Then it runs a second check: how similar are the proposed starting materials to the starting materials used in that original, successful experiment? Both scores are combined to produce a final ranking.
Sam: So it's not just asking "does the finished product look right?" It's also asking "do the ingredients make sense for this goal?" That feels more rigorous than I expected.
Alex: It is more rigorous than a naive approach would be. And the fingerprints that make all this comparison possible are worth understanding. The researchers use what are called circular fingerprints. The computer looks at the local chemical environment around every single atom in a molecule — what atoms are nearby, how they're connected — and assigns a numerical label to each neighbourhood. Then it zooms out, looks at slightly larger neighbourhoods, and labels those too. By the time it's done, it has built a detailed numerical map of the entire molecule's structure.
Sam: Like mapping a city by first cataloguing every street corner, then every block, then every district, until you have the whole layout?
Alex: That's exactly the right picture. And once you have that map as a string of numbers, you can compare any two molecules mathematically, the same way you'd measure the distance between two points.
Sam: I want to go back to the "template explosion" problem you mentioned earlier. Is that what the old rule-based systems ran into?
Alex: Yes. In a rule-based system, you need a separate rule for every structural variation. Change one small part of a molecule — add a single atom in a slightly different position — and suddenly your existing rule no longer applies, and you need a new one. The number of rules required grows very quickly. By working from analogy rather than rules, this system sidesteps that problem entirely. It doesn't need to be told the "why" of chemistry; it just finds the closest known precedent and interpolates from there.
Sam: And how well does it actually work in practice?
Alex: The researchers tested it on a database of around forty thousand reactions drawn from patent literature. In roughly three-quarters of test cases, the correct starting materials appeared somewhere in the system's top ten suggestions. For a system that doesn't explicitly "know" any chemistry — just structural similarity — that's a meaningful result.
Sam: Can it handle more than one step? Building a real drug molecule usually takes many reactions in sequence.
Alex: The study demonstrates that the one-step approach can be chained together to plan multi-step pathways. They applied it to two drug molecules as worked examples to show it holds up across a longer synthesis.
Sam: So the system is essentially acting like an experienced chemist who doesn't memorise every rule, but has seen so many past experiments that they can recognise what a new problem resembles and reason from there.
Alex: That's a fair characterisation. And one practical advantage worth noting is interpretability. When the system suggests a reaction, it can point to the specific precedent it drew from. A chemist can look at that precedent, evaluate whether the analogy makes sense, and decide whether to trust the suggestion. That transparency is harder to achieve with some other computational approaches.
Sam: So it's not a black box — it shows its work.
Alex: Exactly. The researchers position it as a practical, accessible tool that lowers the barrier to chemical planning by drawing on the accumulated record of past experiments, rather than requiring a team to hand-code expertise into software.
Sam: It's a thoughtful approach. Using the history of chemistry as the knowledge base, rather than trying to distil all of chemistry into a set of rules.
Alex: And that distinction matters more than it might first appear. Chemistry is enormously complex, and the edge cases are endless. A system that learns from precedent can, in principle, improve simply by being exposed to more data — which is a more scalable path forward than trying to write rules for every situation.
Sam: That's a genuinely useful framing for where this kind of research is heading. Thanks for walking through it.
Alex: Thanks for listening to ResearchPod.