Finding the Signal Before It’s Too Late: Where AI Can Change the Course of an Oncology Clinical Trial

Anticipating Biology Before Treatment Failure

Cancer treatment has traditionally forced us to learn from failure. A therapy is designed against the biology we can see, then resistance emerges under treatment pressure. Only then do we begin to understand what changed and design the next response. That sequence costs time, capital, and, most importantly, treatment options for patients.

AI may allow development teams to investigate the tumor's next state earlier. Violet Zahedi, MD, of Synamics Therapeutics described using computational models to predict mutations and assess how they could change a target protein. Her team's platform reproduced resistance patterns already observed with EGFR inhibitors. In work involving an Aurora A kinase inhibitor, the model also identified four mutations later cross-checked against clinical and preclinical observations.

The immediate application is compound design. If researchers can anticipate the protein structures most likely to emerge under treatment pressure, they can screen molecules against those predicted states before selecting a lead. The same insight may later inform biomarker strategy and patient stratification. Instead of designing only for the cancer visible at baseline, a program can begin to account for plausible routes of escape.

Response prediction presents the same problem at the patient level. One actionable mutation rarely captures the full relationship among a tumor, its host and a treatment. Ezra Cohen, MD, of Tempus described models that combine molecular information, co-mutations, longitudinal clinical data, and pathology to estimate treatment benefit. One model evaluates the likelihood that a patient with EGFR-mutated non-small cell lung cancer will benefit from a tyrosine kinase inhibitor. Another uses real-world cohorts to improve prediction of immunotherapy response.

For trial developers, the value lies in sharper hypotheses. Predictive models may help define a more biologically coherent cohort, identify patients who are unlikely to benefit, and clarify which resistance mechanisms the protocol should monitor. Knowing where a drug may fail can be as important as knowing where it may work, particularly in early development when teams are deciding whether to expand a cohort, modify a combination, or continue investing in the asset.

Treating Eligibility as a Moving Clinical State

Trial protocols define eligibility as a set of fixed criteria. A patient's condition is anything but fixed. Disease status changes, scans arrive, laboratory values move, and one line of therapy gives way to the next. A patient may qualify for a study only briefly, which makes timing part of eligibility.

The current matching process depends on a physician or study coordinator recognizing that moment, reviewing a longitudinal record, and checking every criterion during an already compressed care pathway. That model helps explain why trial participation in the US has remained at roughly 4% to 5% for more than two decades, said Kent Thoelke of Paradigm Health.

AI can make screening continuous. A model embedded in the electronic health record can follow a patient's course and identify what would need to change for that person to qualify. In a second-line metastatic breast cancer study, for example, the system could track patients receiving first-line therapy and alert the physician when all criteria except progression have been met. If an upcoming scan confirms progression, the care team already knows that the trial may be an option.

This approach preserves the physician's judgment while removing chance from the first step. It also expands the number of records that can be reviewed. Thoelke reported that one rural North Dakota health system increased trial participation from 4% to 11% within a year of deploying this model.

This matters for access. Most people receive cancer care in community settings, while many oncology trials remain concentrated in academic centers. Community research teams often lack the staff required for continuous manual screening. AI cannot create local trial infrastructure on its own, but it may reduce one of the labor barriers that keeps more sites and patients from participating.

Over time, patient matching could also become more precise. Eligibility criteria establish whether someone can enter a trial. Multimodal models may help investigate whether that person's biology suggests a meaningful chance of benefit. Those remain distinct clinical and research questions, but development strategies will increasingly need to consider both.