Strengthening Individual-Patient Treatment Prediction in Precision Oncology: The Case for a National Evidence Infrastructure

Precision oncology has transformed cancer treatment by linking molecular features to therapies that can produce large and durable benefits.Imatinib in BCR::ABL1-positive chronic myeloid leukemia, trastuzumab in human epidermal growth factor receptor 2-positive breast cancer, osimertinib in EGFR-mutant non-small cell lung cancer, and larotrectinib in NTRK fusionpositive cancers are canonical examples.1-5 Across a broader evidence base, a meta-analysis of 570 phase II studies involving 32,149 patients found that biomarker-selected strategies were associated with higher median response rates (31% v 10.5%) and longer median progressionfree survival (5.9 v 2.7 months) than nonselected strategies.6 These achievements establish molecular enrichment as a powerful approach to drug development and treatment selection.However, population-level activity is not individual patient-level prediction.A predictive biomarker requires a treatment-by-biomarker interaction, not merely an association with outcome.7 Enrichment designs can establish activity in biomarker-positive patients without determining whether biomarker-negative patients benefit or which biomarker-positive patients respond.Among 35 (US) Food and Drug Administration (FDA) oncology approvals requiring pharmacogenomic testing, 24 (69%) relied entirely on biomarker-positive trials, and few directly showed lack of benefit in biomarker-negative patients.8 Thus, many companion biomarkers function primarily as eligibility gates rather than calibrated predictors of individual benefit (Table 1).The clearest exceptions are oncogene-addicted settings in which a dominant driver is tightly coupled to the target.Outside those cases, precision oncology still has a major predictive gap to close.The magnitude of this gap remains substantial.Between 2006 and 2020, the estimated proportion of US patients with cancer eligible for a genome-targeted therapy rose from 5.1% to 13.6%, while the proportion estimated to respond rose from 2.7% to only 7.0%.10 Separately, Marquart et al 11 reported a median overall response rate of 54% across genomic-aberration categories for FDA-approved genome-informed drugs through January 2018, using the highest response rate when multiple drugs were approved for the same genomic abnormality.Although this summary statistic is not a pooled patient-level response rate, it indicates that substantial nonresponse persisted across many biomarker-selected settings.Matching a detected alteration to a drug is not the same as predicting benefit for that patientexcept in the narrow set of cancers where the alteration is the tumor's dominant dependency.Recent examples reinforce this pattern: RAS(ON)-directed therapy targeting KRAS G12-mutant pancreatic cancer 12 and selective RET inhibitors in RET fusion-positive lung cancer 13 represent further additions to this narrow, oncogene-addicted category.Outside those oncogeneaddicted, single-driver settings, this bar is rarely cleared.Advanced melanoma is illustrative: nearly a decade into frontline dual PD-1/CTLA-4 blockade, no pretreatment biomarker has been validated to guide the choice between anti-PD-1 monotherapy and combination therapy-the pivotal trial itself cautioned that the optimal PD-L1 assay and cutoff for this purpose remain undetermined.14 PD-L1, tumor mutational burden, and interferon-g signatures can enrich a population for response, but they do not reliably tell a clinician which specific patient needs the more toxic combination.15,16