Featured image: A printed circuit board illustrates the computing behind generative models. Contextual photograph; not a MolGenBench model, generated molecule or benchmark result. Photo: Vishnu Mohanan / Unsplash. Unsplash licence.
Generative AI can propose molecular structures at astonishing speed. Drug discovery asks a harder question: do those structures have the right properties for a particular biological target, avoid chemical liabilities and improve on what scientists already know?
MolGenBench, a peer-reviewed benchmark published in Nature Communications on 1 October 2026, evaluated 17 leading molecular-generation methods under tasks designed to resemble real projects. The overall result was sobering: many models produced plausible-looking chemistry but struggled with target-specific activity, risky motifs and proteins they had not encountered before.
A generated molecule has to satisfy several constraints
A molecular generator creates candidate chemical structures, often represented as text strings or graphs. De-novo design starts with a target or set of desired properties and proposes new structures. Hit-to-lead optimisation begins with an experimentally active “hit” and tries to improve it into a more promising lead.
Neither task is solved by making a structure that is merely valid on paper. A useful candidate needs appropriate potency, selectivity, solubility, metabolic behaviour, synthetic accessibility and safety. These requirements can conflict: improving binding to one protein can make a compound harder to manufacture or increase an unwanted interaction elsewhere.
That is why a single score can flatter a model. A method may generate diverse structures while ignoring the target, or optimise a prediction made by another model without improving real experimental activity.
MolGenBench changes the test data
The authors assembled a structurally diverse dataset spanning 120 protein targets and 600 compound series, according to the open study manuscript. Compound series group related molecules that medicinal chemists modify while learning how structure changes affect behaviour.
The benchmark separates familiar from unseen proteins and includes both de-novo generation and hit-to-lead tasks. This matters because a model can appear successful by reproducing patterns close to its training data. Performance on a protein held out from training is a stronger test of whether it has learned transferable relationships.
Seventeen methods were compared through a common evaluation pipeline. The authors also released the MolGenBench code and data tools, giving other researchers a way to inspect the setup and test new models.
The weak points were scientifically consequential
Across the benchmark, virtual-screening performance was limited and models often showed weak awareness of the requested target. Generated sets frequently contained high-risk structural motifs—chemical patterns associated with reactivity, interference or other development problems.
Generalisation to unseen proteins was also poor. That is a critical limitation for claims that a model has learned general chemical or biological rules. If performance falls when the target changes, the system may be interpolating within familiar examples rather than reasoning reliably about a new protein.
Hit-to-lead optimisation posed another trap. A candidate can score better on a computational objective while moving away from the experimentally grounded chemistry that made the starting hit useful. The benchmark rewards improvement across realistic constraints rather than one model-friendly number.
A benchmark is not a wet-lab verdict
MolGenBench is a computational evaluation. It does not synthesise every generated compound, measure its activity in cells or establish that none of the methods can produce a useful medicine. Laboratory experiments remain the decisive test of a candidate’s properties, followed by animal studies, human trials and regulatory review.
The benchmark also depends on its datasets, prediction methods and definitions of success. New models may improve, and a specialised system used with expert review could behave differently from a general comparison. The value of the work lies in making those conditions visible and repeatable.
Our assessment is that the paper raises the standard for AI-drug-discovery claims. A compelling demonstration should identify the target, show performance on genuinely held-out chemistry, screen out liabilities and validate prioritised compounds experimentally. Counting molecules or displaying attractive structures is not enough.
AI can still help medicinal chemists search a vast design space. MolGenBench shows why the next advance will be measured less by how many structures a model can imagine and more by how many survive contact with chemistry and biology.
Reporting and source checks completed on 12 October 2026. By FutureTechDose Editor.
Leave a Reply