AI Is Learning to Find Cancer Before It Comes Back

After a cancer patient finishes chemotherapy or immunotherapy and their scans come back clean, the question that haunts both patient and physician is the same: is it really gone? Traditional imaging has limits. A CT scan can't reliably detect a few thousand cancer cells hiding in bone marrow or circulating in the bloodstream. But those cells, if left unchecked, are often what leads to relapse months or years later. That's the problem minimal residual disease detection was built to solve, and AI is making it dramatically more powerful.

This week at the American Society of Clinical Oncology (ASCO) annual meeting, MRD was one of the hottest topics on the floor. Multiple sessions highlighted new clinical data on AI-assisted MRD assays, and several oncology analytics companies walked away with industry recognition for their work in this space. The momentum is real, and the implications for how cancer is monitored and treated are significant.

What MRD Actually Means

Minimal residual disease refers to the small number of cancer cells that can remain in a patient's body after treatment ends, even when standard tests show no detectable disease. In leukemia, for example, patients can achieve what looks like complete remission, yet still harbor one cancer cell for every million healthy cells. That's enough to eventually cause relapse in many cases.

MRD testing has existed for years in hematologic cancers like acute lymphoblastic leukemia (ALL) and multiple myeloma. Techniques like flow cytometry and PCR-based assays can detect one cancer cell per 100,000 or even one per million. But these methods have historically been expensive, technically demanding, and not broadly available across cancer types. They also generate enormous amounts of complex data that require expert interpretation.

Enter AI.

How AI Changes the Detection Game

The core challenge with MRD detection isn't just sensitivity, it's signal versus noise. When you're looking for trace amounts of circulating tumor DNA (ctDNA) in a blood sample, you're sifting through millions of DNA fragments. Many apparent "signals" are sequencing errors, normal genetic variants, or artifacts from sample handling. False positives are a serious problem because they can lead to unnecessary treatment intensification. False negatives are equally dangerous because they give false reassurance.

AI models, particularly those trained on deep sequencing data from large patient cohorts, can learn to distinguish genuine tumor-derived mutations from background noise with a level of precision that rule-based algorithms simply can't match. Companies like Foundation Medicine, Guardant Health, and newer entrants like Haystack Oncology (acquired by Quest Diagnostics in 2023) have been building machine learning pipelines specifically designed to push the limit of ctDNA detection further down.

Haystack's approach, for instance, uses personalized tumor-informed assays that sequence a patient's primary tumor first, identify hundreds of somatic mutations unique to that tumor, and then train a model to look for that specific fingerprint in blood draws over time. By tracking dozens or hundreds of mutations simultaneously rather than just one or two, the statistical confidence of detecting even a single cancer cell equivalent's worth of DNA increases substantially.

The Numbers Behind the Sensitivity

Some of the newer AI-assisted platforms are reporting limits of detection at variant allele frequencies below 0.001%, meaning they can detect one mutant DNA molecule among 100,000 normal ones. That's not just incrementally better than older methods; it's a qualitative shift in what's clinically possible. Clinical data presented at ASCO this year showed that in colorectal cancer, ctDNA positivity detected by these ultrasensitive assays after surgery predicted recurrence with a lead time of six to twelve months ahead of standard imaging in a significant portion of patients.

Six to twelve months of advance warning is enormous. It's the difference between catching a recurrence when it's still potentially resectable and catching it when it's already spread.

Shifting Clinical Practice in Real Time

The way oncologists make decisions is already starting to change because of MRD data. Traditionally, treatment decisions after first-line therapy were based on imaging every few months, sometimes combined with tumor marker blood tests like PSA for prostate cancer or CA-125 for ovarian cancer. These are blunt instruments. MRD assays offer something more granular: a continuous read on whether residual disease is present and, in serial testing, whether it's growing or stable.

In multiple myeloma, MRD negativity has essentially become the gold standard endpoint in clinical trials. Achieving MRD negativity, confirmed by next-generation sequencing at a sensitivity of one cancer cell per million, is now considered a stronger predictor of long-term outcomes than older response criteria. The FDA has been actively engaging with the oncology community about potentially using MRD as a surrogate endpoint to accelerate drug approvals, which would compress drug development timelines significantly.

For trial design, AI-powered MRD also opens up a different kind of patient selection. If you can identify who still has residual disease after initial therapy, you can enrich clinical trial populations with patients most likely to benefit from an experimental consolidation treatment. That makes trials more efficient and more likely to produce a clear signal.

The Challenge of Solid Tumors

Most of the mature MRD data comes from blood cancers, where tumor cells shed DNA into circulation at relatively high rates. Solid tumors are harder. A localized pancreatic tumor or an early-stage lung tumor may shed very little ctDNA, and the signal can be intermittent. AI is making inroads here too, but the sensitivity requirements become even more extreme.

Multi-cancer early detection (MCED) companies like GRAIL, whose Galleri test screens for over 50 cancer types from a single blood draw, are grappling with exactly this problem. Their platform uses machine learning to analyze methylation patterns across thousands of genomic regions, looking for tissue-of-origin signatures that indicate cancer. The test is not an MRD assay per se, but the underlying AI architecture, learning to find faint cancer signals in a complex biological background, is directly applicable to MRD use cases in solid tumors.

What This Means for Patients

The practical implications for someone who has been treated for cancer are significant and, in many ways, psychologically profound. The current standard of care often leaves patients in a state of uncertainty between imaging appointments. You feel fine, but you don't know. MRD testing could provide more frequent, more sensitive reassurance, or earlier warning, depending on the result.

That said, more information isn't automatically better without clinical context. A patient who tests MRD positive after surgery but has no standard treatment options left in that setting faces a difficult situation. Knowing the cancer is still there without having a clear path to act on that knowledge creates its own kind of distress. This is why the oncology community has been careful about deploying MRD testing outside of contexts where results can actually guide decisions.

Access is also a real barrier. These assays are expensive, often running between $2,000 and $5,000 per test, and insurance coverage is inconsistent. Medicare has begun covering certain MRD tests in specific cancer types and clinical scenarios, but coverage gaps remain wide, particularly for solid tumors where the clinical evidence is still accumulating.

The AI Infrastructure Behind It All

Building these models isn't a weekend project. Training an AI to reliably detect ctDNA at one-in-a-million sensitivity requires access to massive annotated datasets, samples from thousands of patients with known outcomes, paired tumor and blood data, longitudinal draws over time, and carefully controlled laboratory processes to minimize pre-analytical noise. The companies doing this well have invested years and hundreds of millions of dollars in both the wet lab infrastructure and the computational side.

One area generating particular interest is the use of large language model-style architectures applied to genomic sequences. Rather than treating DNA as a series of individual variants to look up in a database, newer approaches treat long stretches of the genome the way a language model treats text, learning contextual patterns across thousands of base pairs simultaneously. Early results suggest this can improve both sensitivity and specificity beyond what traditional variant-calling pipelines achieve.

The convergence of AI capability, cheaper sequencing costs (whole genome sequencing that cost thousands of dollars a decade ago now runs under $500 in many settings), and growing clinical datasets is creating a real inflection point. The models are getting better faster than the clinical frameworks for using them, which is both exciting and a bit uncomfortable for a medical establishment that rightly prioritizes evidence before adoption.

Where Things Go From Here

ASCO 2024 made clear that MRD is no longer a niche research curiosity. It's becoming a core component of oncology practice in blood cancers and is rapidly advancing in colorectal, lung, and breast cancer. The AI layer is what makes the ultra-high sensitivity possible, and it's what will ultimately determine how broadly this can be applied across cancer types and clinical settings.

The next few years will likely bring FDA approvals of AI-powered MRD assays as companion diagnostics tied to specific therapies, meaning the test and the drug are developed and approved together with predefined clinical decision rules. That's the clearest path to both regulatory clarity and insurance coverage, and several companies are already in late-stage discussions with the FDA along these lines.

For patients, the promise is a future where "your scans are clear" is replaced by something more precise, more honest, and ultimately more actionable. That future is closer than most people realize.