Zero-shot generation is not data-free

A proposed molecule can emerge from an algorithm without training a new generator for that particular request. That does not mean the algorithm has no learned chemical history. SiMGen, a 2025 molecular-generation method, is an instructive example because its local similarity machinery and pretrained components are described separately.¹

A reference shapes the search

SiMGen constructs a time-dependent similarity landscape using local atomic representations, then updates candidate structures using that guidance.¹ The published implementation uses representations from a pretrained MACE model.¹ The distinction we want a reader to retain is between learning a new generative model and using previously learned information to guide a new structure.

For a practical comparison, we would ask a researcher to show the reference set alongside the outputs. What examples were deliberately included, and which were left out? What aspect of the desired result did the chosen references express? These questions make a generation request inspectable without assuming that a visually appealing output must satisfy an unstated chemical objective.

Shapes and fragments are different requests

The authors describe point-cloud priors for shape control and using the method to guide an existing generator towards a reference fragment.¹ These are computational demonstrations of conditioning. They are not measurements showing that the generated molecules perform a useful biological or chemical function.

Our interpretation is that a generation report should keep the requested constraint separate from the properties evaluated afterwards. It should say whether the output met the request, how that was checked and what remained outside the check. Otherwise a reader may mistake success at one narrow geometric task for success at the eventual discovery goal.

This separation would also make failures informative. If a generated structure misses the requested shape but satisfies another test, the discrepancy should stay visible. The record should not become a gallery containing only whichever successes are easiest to illustrate.

Account for refinement

The method adds hydrogens and refines geometry using trained components and a pretrained force field; the paper describes the hydrogenation model as a design choice rather than a requirement of the general approach.¹ Zero-shot therefore describes the absence of further generator training in this use, not the absence of models, reference molecules or computational preparation.¹

When assessing a reported cost advantage, we would list those dependencies explicitly. Which computation is reused? Which is repeated for every candidate? Which assumptions entered before generation began? A clear answer would let another group decide whether the same advantage carries over to its own task.

A candidate needs an evidence record

For this launch analysis, the useful advance is a different, inspectable way to guide generation. We would keep a candidate record containing its references, requested constraint, refinement history and later evaluations. That is our proposed reporting practice, not a claim that we have tested an integrated discovery workflow.

AiChemEx has not run SiMGen, reproduced its benchmarks or synthesised its outputs. A generated structure remains a candidate for further investigation. The method's title becomes more informative when the reader can see exactly which training step is avoided and which inherited chemical knowledge remains essential.