Wet Lab Validation for AI-Driven Drug Discovery

AI-generated candidates need experimental validation because a computational score reflects a model’s training data, not a physical measurement of the molecule. Generative and structure-based models produce far more candidates than any predicted score can qualify, and a large fraction of high-scoring structures show no activity when tested. Biochemical, cellular, and in vivo assays separate the real chemotypes from the artifacts and feed measured results back into the model.

Design-Make-Test-Learn is the closed loop in which an AI model designs molecules, they are synthesized, tested experimentally, and the results are learned back into the model. A wet lab CRO runs the Test step, and the quality of that step determines the value of the Learn step. If the assay conditions behind a data point are undisclosed or inconsistent, the model learns noise. Reaction Biology runs the Test step with stated, constant conditions so each cycle returns data the model can actually use.

Yes. The predictive power of a model is limited by the consistency of the data it is trained on, not only its volume. An IC50 depends on the protein construct, ATP concentration, substrate, and detection method used to measure it, so data pooled from undisclosed or mixed conditions carries variance into the model that more data does not correct. Consistent, disclosed conditions across a program are what make a dataset internally comparable and suitable for training.

The fastest test is single-concentration biochemical panel screening against purified protein. It removes cellular variables such as permeability and metabolism, requires modest compound quantities, and returns an activity fingerprint across many targets at once. Confirmed actives then proceed to dose-response for IC50 values. For kinase programs this is the standard first filter between a set of designs and a decision on which chemotypes to pursue.

An AI-designed small molecule is validated in escalating steps: biochemical activity against purified protein, selectivity profiling across a target family, intracellular target engagement, cellular function, then in vivo confirmation, followed by off-target and safety panels. The order matters because each step adds one biological variable, so a failure can be assigned to a cause rather than left ambiguous. Skipping from a biochemical IC50 straight to an in vivo study produces failures that cannot be designed against.

IC50 values disagree because an IC50 is a property of an assay, not a constant of a molecule. Its value shifts with protein construct, ATP concentration, substrate, detection chemistry, and cellular background. Most kinase inhibitors compete with ATP, so the same compound appears more potent at low ATP than at physiological 1 mM. Reaction Biology states the ATP concentration for every kinase target and holds conditions constant across a program so values stay comparable.

NanoBRET target engagement measures whether a compound occupies its target inside intact living cells. A NanoLuc-tagged target and a cell-permeable tracer generate a bioluminescence resonance energy transfer signal, and a binding compound displaces the tracer and reduces it. Run it after biochemical confirmation and before cellular efficacy or in vivo work, because it identifies compounds that are potent against purified protein but never reach the target in a cell.

Non-oncology compounds are validated with the same escalating cascade plus target-class and pathway assays specific to the indication. Selectivity is profiled across epigenetic, phosphatase, protease, phosphodiesterase, nuclear receptor, ion channel, GPCR, and metabolic enzyme classes rather than the kinome alone. For inflammatory, autoimmune, and neuroinflammatory programs, immunophenotyping by flow cytometry and multiplex immune assays extend validation into functional immune consequence, and off-target and cardiac safety panels define liability early.

Yes. Custom assay development begins with a feasibility discussion covering known target biology, likely construct boundaries, detection chemistry, and available reference compounds, then proceeds through protein expression and purification, assay optimization, and validation to screening-ready criteria. This applies across target classes, including kinases, epigenetic enzymes, GPCRs, ion channels, and protein-protein interactions. It is the standard answer when a computational platform nominates a target no existing assay covers.

Turnaround depends on the assay, but standard kinase panels and single-target work at Reaction Biology are reported in 10 to 15 business days, and InVEST safety panels in 10 business days. Custom assay development timelines are set during scoping and depend on protein production and optimization. Directly accessible study directors and protocol adjustment without a new statement of work keep iteration inside the design cycle rather than adding weeks between rounds.

Reaction Biology does not operate a computational drug discovery platform, does not sell predictive models or curated training datasets, and does not require clients to contribute data to a shared resource as a condition of service. Results are generated for the client program only. Most assay work proceeds from compound identifiers without structural disclosure, and studies can be scoped to a single US or EU site where compound or data jurisdiction requires it.