Oncology remains one of the highest-attrition areas in drug development, with only a small minority of candidates entering clinical testing going on to reach approval. The reasons are familiar, including lack of efficacy, unexpected toxicity or insufficient differentiation, but the deeper cause is often a persistent disconnect between what is measured in the laboratory and what ultimately matters in patients.

The recent discontinuation of Roche’s tiragolumab programme, an anti-TIGIT antibody, is instructive. An encouraging randomised Phase II signal supported development across several tumour types and thousands of patients. The results proved inconsistent. The pivotal NSCLC study missed its overall survival endpoint, while a parallel study in oesophageal squamous cell carcinoma met both of its primary endpoints. TIGIT is a well-characterised inhibitory receptor and the rationale for blocking it was biologically coherent. What the programme lacked was a way to anticipate where that rationale would hold. Trials enriched for PD-L1 expression, but no marker specific to TIGIT biology distinguished the settings in which the combination helped from those in which it did not.

Failures of this kind are rarely failures of data volume. The mechanistic evidence existed and the trials were large. What was absent was a chain connecting laboratory measurement to the decision it was meant to inform, so that the biology could be tested in the population where it applied. As data generation accelerates and computational methods mature, the central challenge is therefore no longer access to data. It is whether that data is captured, connected and interpreted well enough to guide confident decisions at every stage of a programme.