Artificial intelligence for small-molecule chemical probe design: why drug-centric models may not transfer cleanly

Artificial intelligence (AI) is increasingly used to design and optimize small molecules. However, most current workflows remain centered on drug-candidate objectives, such as potency, ADMET, and drug-likeness, rather than on the distinct requirements of small-molecule chemical probes. We argue that this distinction matters for drug discovery because chemical probes are essential for target validation and mechanistic interrogation. Their value, however, depends on stringent selectivity, demonstrable target engagement in relevant biological contexts, fit-for-purpose concentration windows, and appropriate control compounds that support causal interpretation. These requirements make probe discovery a fundamentally different optimization problem from conventional drug design. Here, we examine how contemporary AI methods (including property prediction, multi-target modeling, generative optimization, and synthesis-aware design) can contribute to chemical probe discovery while highlighting why drug-centric models may not transfer cleanly. We further discuss emerging applications in fluorescent and covalent probe design, where photophysical behavior, signal detectability under biologically relevant conditions, intrinsic reactivity, and site selectivity introduce modality-specific constraints that are poorly captured by conventional datasets and benchmarks. We propose that meaningful progress will require probe-native datasets, probe-centric evaluation frameworks, and closed-loop experimental validation strategies that treat probe success as a function of the molecule, evidence, and context of use.