FDA Just Made MRD a Myeloma Drug Endpoint. AI Reads It.

On January 20, 2026, the FDA published draft guidance that quietly rewrites how multiple myeloma drugs reach the U.S. market. The document makes minimal residual disease — MRD — negativity an acceptable primary endpoint for accelerated approval. Drugmakers no longer have to wait six to eight years for progression-free survival data to mature. They can now seek approval based on whether their therapy clears the deepest layer of cancer cells from a patient’s bone marrow.

That sounds straightforward until you read what MRD actually requires. The FDA’s accepted threshold is one cancer cell hidden among 100,000 normal cells, with newer assays pushing toward one in a million. No pathologist on Earth can find that under a microscope. The shift to MRD-based approval is, in practice, a shift to AI-read endpoints. The data is too dense, too noisy, and too high-dimensional for anything else.

This is one of the largest oncology regulatory changes of the decade. It also commits the system to algorithms most patients have never heard of.

What MRD Means at One in a Million

MRD is what is left behind after standard treatment looks like it worked. After induction therapy, transplant, and maintenance, a multiple myeloma patient can show a complete response on every standard test — clean bone marrow biopsy, normal blood protein readings, no measurable monoclonal protein. Yet the disease relapses anyway, often within five years. Something was hiding.

That something is now measurable. The two accepted methods are next-generation sequencing of immunoglobulin gene rearrangements and next-generation flow cytometry. Both can detect a single myeloma cell among hundreds of thousands of healthy bone marrow cells. The gold-standard sensitivity now expected in clinical trials is 10⁻⁵, with 10⁻⁶ becoming the new ceiling.

At those depths, MRD-negative patients live dramatically longer. A 2024 meta-analysis covering more than 8,000 patients across 44 studies showed that achieving MRD negativity at 10⁻⁵ roughly doubled progression-free survival regardless of treatment, transplant status, or risk category. For older patients ineligible for transplant, MRD negativity is now the strongest single predictor of long-term outcome the field has.

The FDA recognized this. The new guidance treats sustained MRD negativity at 12 months as an endpoint that “reasonably predicts” clinical benefit, which is the legal language that triggers accelerated approval.

Why a Pathologist Cannot Read This

Multiple myeloma is rarely a single clone. It is usually three to ten genetically distinct clones living in the same marrow, each with different mutations, different treatment sensitivity, and different ability to relapse. When a patient achieves MRD negativity by one assay, sometimes the dominant clone is gone but a minor clone is hibernating. When the patient relapses, that minor clone is what comes back.

This is where machine learning becomes load-bearing. NGS-based MRD assays generate millions of read sequences per sample. Each read has to be matched against a baseline tumor signature, sorted from background noise, and interpreted in the context of the patient’s clonal architecture at diagnosis. A 2024 paper in Blood Cancer Journal trained machine-learning models on 482 patients at UCSF and showed that combining MRD depth with AI-derived clonal diversity scores predicted relapse better than MRD alone. Patients who looked equally MRD-negative on a single readout had wildly different outcomes once clonal complexity was factored in.

Flow cytometry is no easier. Next-generation flow analyzes ten or more surface markers across a million cells per sample. Identifying the rare aberrant population takes pattern recognition that humans can do for one or two markers but not ten. AI-based gating tools — many already cleared as research-use-only software — now do this in minutes with reproducibility human cytometrists cannot match between sites.

The FDA guidance does not name vendors. But it makes clear that any key trial submitting MRD as a primary endpoint must use a validated assay with documented analytical performance. In practice, that means software-defined readouts.

Beyond Myeloma: Leukemia and Lymphoma Are Next

The myeloma guidance is the first to land, but the pattern is identical in chronic lymphocytic leukemia, acute lymphoblastic leukemia, acute myeloid leukemia, and follicular lymphoma. CLL trials have used MRD as a regulatory endpoint since 2018. AML is moving the same direction, with AI-based flow assays now central to deciding which patients in remission still need a bone marrow transplant and which do not.

The clinical impact is the inverse of what oncologists usually do. Most cancer treatment escalates when imaging looks worse. MRD-guided care lets clinicians de-escalate when blood and marrow look clean. A patient on lenalidomide maintenance for myeloma who hits sustained MRD negativity for two years can now plausibly stop therapy under physician supervision in clinical trials testing exactly that question. That conversation was impossible five years ago.

The AI piece quietly extends to MRD-guided treatment selection. AI digital twin platforms are already using MRD trajectories — the rate at which MRD falls during induction — to predict which patients will need consolidation and which will not. AI biomarker tools are mining the genomic data behind MRD assays to predict relapse before any cells are detectable at all.

The De-escalation Question

Patients want to know when they are done. Years on lenalidomide, dexamethasone, or daratumumab maintenance carry real costs — secondary cancers, financial toxicity, fatigue, infection risk. Clinicians want to give honest answers about whether a patient is still at risk. MRD makes that conversation possible. AI makes the MRD readout reliable enough to base the conversation on.

The MRD2STOP trial at the Levine Cancer Institute randomized myeloma patients in long-term MRD negativity to stop maintenance versus continue. Early signals suggest stopping is safe in a subset — and AI-based risk scores combining MRD depth, clonal diversity, and cytogenetics can identify that subset prospectively. The first-line decision becomes: induce hard, sequence to MRD, then de-escalate when the algorithm agrees.

This is the same logic CAR-T programs are racing to. Sustained MRD negativity after CAR-T manufacturing improvements pull the cost lever down from $400,000 means the therapy has done its job and the patient can come off bridging maintenance. Without an AI-trusted MRD readout, no one would dare make that call.

The Hard Limits

The FDA’s draft guidance is careful about scope. MRD is accepted in newly diagnosed and relapsed/refractory myeloma. It is not accepted in maintenance, smoldering myeloma, monoclonal gammopathy of undetermined significance, or extramedullary disease. Each of those represents a setting where the relationship between MRD and survival is not yet locked down.

There are also assay-level concerns. MRD readouts vary between centers, between assays, and between bone marrow draws taken on the same day. The FDA flagged “rigorous analytical performance and harmonized application across trial arms” as required. In English: every site has to read MRD the same way. AI models trained on one institution’s flow data will not always generalize to another’s. The validation problem here is genuinely unsolved.

There is also the question of what MRD does not see. Multiple myeloma can recur as extramedullary disease — soft-tissue plasmacytomas that bypass the bone marrow entirely. A patient can be deeply MRD-negative in the marrow and still relapse outside it. PET-CT imaging plus AI-based image analysis is now being added to MRD definitions to close that gap, but it is not yet standard.

Where This Goes

The FDA is accepting public comments on the draft guidance through March 23, 2026, with finalization expected later in the year. Once final, the first MRD-based accelerated approvals could land in 2027. Several myeloma drugs already in late-stage development — particularly bispecific antibodies and next-generation CAR-T constructs — were designed with MRD endpoints in mind.

The deeper change is that AI-read endpoints are now a regulatory reality, not a research idea. The same logic — a measurable, software-defined endpoint that beats clinical-progression endpoints on speed and statistical power — will spread to other cancers, then to other diseases. The MRD precedent is the camel’s nose. Once a regulatory agency accepts that an algorithm reads the answer, the question is no longer whether AI guides treatment. It is how soon, and for what.

For myeloma patients, the effect is closer to home. The drug that saves a life next decade is being approved on the basis of whether software can find a single cell in a bone marrow sample. That cell, found or unfound, is the new finish line.