AI for drug discovery: Thomas Kaiser proposes testing predictions in each chemical series

AI for drug discovery: Thomas Kaiser proposes testing predictions in each chemical series

In a September 8 BioSpace opinion column, physician, medicinal chemist, and machine learning specialist Thomas Kaiser examines the conditions under which foundation models can help drug development. A foundation model is trained on a broad range of publicly available data and then adapted to a specific task. Kaiser proposes checking both the quality of the scientific findings used in training and whether the relationship between molecular structure and a property holds in a new chemical series.

An engineer can configure a particular version of a model to give the same answer to the same prompt. This consistency reflects how the computation is configured. A scientific publication may report a relationship that subsequent studies cannot reproduce, and the model may consistently repeat it. Kaiser describes the risk this way: “A reproducible relationship based on irreproducible science is, unfortunately, no progress.”

Making an answer repeatable and validating a scientific relationship are two separate tasks.

Kaiser illustrates the second risk with small molecules, which are relatively small organic compounds. A chemical series is a family of molecules with similar structures. The same small structural change can affect a molecular property differently in two series. Kaiser describes encountering this problem at Avicenna Biosciences: a rule that helps predict a property in one series may mislead the model in another because the causal relationship between the structural change and the molecular property changes.

Kaiser therefore proposes asking a foundation model a narrowly defined scientific question with an established chemical context. Potential tasks include identifying biological networks, meaning connected genes and proteins that could serve as drug targets, and analyzing clinical data to select patients. The model could also look for relationships among measurable treatment outcomes, patient selection, and preclinical models that reproduce features of human disease. Kaiser leaves experimental validation of these relationships and careful assessment of the data to the researcher.