How substitution screening actually works

“Latent space” and “similarity graph” get thrown around a lot in this space. Here’s what they mean in practice, and where the honest gaps still are.

1. Turning a molecule into numbers

Before anything can be compared, a source ingredient — a name, a CAS number, a SMILES string — needs to become a numerical representation a computer can reason about. Depending on the approach, that can mean molecular fingerprints (bit vectors marking the presence of structural fragments), learned embeddings (a neural model trained to place chemically similar molecules near each other), or explicit similarity graphs built from shared substructures. {{TODO: confirm which representation(s) SUBSTØR uses in production}}

2. Finding structurally similar candidates

Once every indexed molecule has a numerical representation, finding candidates is a nearest-neighbor search: which indexed molecules sit closest to the source ingredient in that representation space? This is the “latent space navigation” part — it’s a search problem, not a lookup table.

3. Filtering by what actually matters

Structural similarity alone isn’t enough — a structurally close molecule can still be unusable because it’s unstable, harder to source, or already flagged by a regulator. This is where constraint filtering comes in: candidates get checked against the requirements a formulator actually cares about before they’re ranked, not after.

4. Validating that it works

The honest way to test a system like this is a leave-one-out evaluation: take a substitution that’s already known and documented, remove it from the reference data, and check whether the model recovers it using only what’s left. It’s not proof that every future recommendation will be right — nothing offers that — but it’s a concrete, repeatable way to measure whether the approach is finding real chemistry or noise. {{TODO: publish current leave-one-out results once available}}

Where this doesn’t replace a chemist

None of the above replaces lab validation. What it does is narrow a search space that would otherwise depend on one researcher’s memory and literature search time, down to a short, explainable, constraint-checked list worth testing.