Few-shot selectivity still needs a history

A model that needs only a handful of examples for a new task can sound as if it has learned chemistry from almost nothing. That is the wrong starting image for a 2025 study of asymmetric hydrogenation. The method learns from a larger reaction history before adapting its predictions to smaller sets of examples.¹

Identify what the model is asked to decide

Singh and Hernández-Lobato use prototypical networks, a meta-learning approach, with a literature-derived collection of 11,932 reactions involving iridium, rhodium and cobalt catalysts.¹ Their principal classification task separates reactions above and below an 80% enantiomeric-excess threshold.¹ That is a classification of selectivity, not a promise to predict every numerical experimental result exactly.

For a reader, the threshold is part of the scientific question. We would state it before presenting a performance score and ask why that decision boundary is useful for the proposed application. A different question may require a different evaluation, even if the underlying reaction records are unchanged.

Small support sets, substantial prior learning

The authors test support sets of 16, 32 and 64 examples and compare their meta-learning predictions with single-task methods.¹ A support set supplies examples for the current prediction task. The earlier meta-training supplies experience shared across tasks. Calling only the smaller set the training data would hide a substantial part of the information available to the system.

Our suggested comparison would give each method a complete information budget: prior reaction records, task-specific examples and any chemical descriptors computed beforehand. This would let a reader distinguish efficient reuse of experience from an unequal contest. It would also make the practical cost of adopting a pretrained model easier to discuss.

Ask how unfamiliar the test really is

The paper evaluates cluster-based splits and a separate set of 245 literature reactions outside the original dataset; it notes that some substrates or ligands in that later set remain similar to training examples.¹ These are useful, qualified tests of generalisation. They should not be retold as evidence that the model can handle any new reaction family.

We would therefore describe unfamiliarity in chemical terms: what changed between the examples used for learning and the examples used for judging the result? A reader can then decide whether that change resembles the problem they actually face. Merely calling data held out does not supply that explanation.

The next experiment remains an experiment

The practical value we would investigate is whether a prediction helps choose an informative early test in reaction development. That requires an explicit prospective record: what was recommended before the result was known, which alternatives were considered and what the subsequent measurement showed. This article does not claim that AiChemEx has performed such a comparison.

The study provides a reason to examine how experience can be shared across related selectivity tasks. Our interpretation keeps the prior learning, the classification boundary and the test chemistry visible together. The small-data result becomes more useful when readers can see precisely which larger history made it possible.