AI has demonstrated remarkable progress in binding prediction and other stages of drug discovery, but impressive molecules can still fail if not matched to the full translational context. The drug discovery lifecycle spans a chain of interdependent decisions from disease selection to clinical trial strategy. Weak decisions at any point can compromise the program.
As a result, clinical outcomes for AI-discovered drugs remain mixed: many candidates can reach Phase 1, but few have reported convincing human efficacy signals to date. Consider pancreatic ductal adenocarcinoma (PDAC). An AI system might design a potent molecule against a validated driver such as KRAS G12D or KRAS G12V, but potency alone may not suffice. PDAC tumors are often desmoplastic, hypovascular, and immunosuppressive, which can limit drug exposure and make single-agent pathway inhibition vulnerable to adaptive resistance.
Working backward from the patient, the molecule question broadens into a systems question. Beyond "can we inhibit KRAS?", an end-to-end AI platform must identify which KRAS-mutant PDAC subgroup to treat, whether the drug can reach the relevant tumor compartment at sufficient exposure, what resistance or microenvironmental biology must be addressed, and what biomarker-selected trial would actually test the mechanism.
As outlined in our lab philosophy, many critical problems in drug discovery may now be constrained by model judgment, not model intellect. The first plank of our platform builds on this hypothesis and focuses on advancing novel therapies and drug repurposing strategies for rare diseases. We are currently testing the platform against a rare cancer type.
