AI Digital Twins Now Guide Multiple Myeloma Treatment

Multiple myeloma is the second most common blood cancer, and it remains stubbornly heterogeneous. Two patients who look identical on paper — same stage, same genetics, same treatment — can have wildly different outcomes. One walks into a durable remission; the other relapses inside a year. Oncologists have spent two decades trying to predict which is which, and they have mostly failed.

In 2026, a new generation of AI digital twins is starting to change that. These models do not just classify patients into risk buckets. They build a running simulation of each patient’s disease — complete with clonal evolution, drug resistance mechanisms, and projected response curves — and let doctors stress-test therapy choices before committing. A paper published this month in Frontiers in Digital Health calls this shift “the move from prediction to simulation,” and it is already reshaping how the biggest myeloma centers pick treatment.

Why myeloma was a prediction nightmare

Myeloma sits in a hard spot for traditional AI. The disease evolves. A patient’s malignant plasma cells undergo clonal shifts during treatment — minor subclones that were invisible at diagnosis become dominant after three cycles of therapy and drive relapse. Static risk scores like R-ISS and mSMART assume the disease stays roughly the same from month one to month thirty-six. It does not.

Older machine learning tools inherited the same flaw. They took a snapshot of clinical variables, ran a classifier, and produced a risk category. That works for triaging patients at diagnosis. It fails the moment treatment starts modifying the biology.

The second problem is data sparsity. Most myeloma patients see their oncologist every four to eight weeks. Between visits, the disease is a black box. Risk models trained on sparse longitudinal data have to average across long windows, which smooths out exactly the signal that matters — the early kinetics of minimal residual disease (MRD) and the pace at which a clone rebounds.

What a digital twin actually does

An AI digital twin for myeloma is a generative model of a specific patient’s disease trajectory. Instead of outputting a single probability, it outputs a simulated timeline. Run it once, you get one possible future. Run it a thousand times under different treatment assumptions, you get a distribution of outcomes you can compare.

The architecture typically combines three components. A transformer-based longitudinal model ingests multimodal inputs — bone marrow biopsies, flow cytometry, FISH cytogenetics, serum paraprotein, free light chains, PET-CT imaging — and encodes the patient’s state over time. A biological simulator models clonal dynamics: which subclones exist, how they respond to each drug class, how resistance mutations emerge under selection pressure. A decision layer then scores candidate regimens against the simulated trajectories.

The result is a practical question an oncologist can actually ask: if I give this patient daratumumab-lenalidomide-dexamethasone versus daratumumab-bortezomib-lenalidomide-dexamethasone (a quadruplet), how likely is a deep MRD-negative response at month 12, and what does the projected progression-free survival look like under each choice? The twin gives you a probability distribution, not a single point estimate.

The CORAL-histomorph breakthrough

One of the most striking 2026 results is a model called CORAL-histomorph, trained to predict genomic abnormalities directly from bone marrow biopsy images. This matters because cytogenetic testing is expensive, slow, and often incomplete — many smaller hospitals cannot run a full FISH panel, and patients wait two weeks for results.

CORAL-histomorph looks at standard H&E-stained bone marrow slides and predicts high-risk features like t(4;14), del(17p), and 1q gain with accuracy approaching direct genomic testing. The model was trained on roughly 3,000 annotated cases and validated externally. Critically, it integrates its genomic predictions into a digital twin architecture rather than outputting them as standalone results. That means the risk prediction and the treatment simulation share a single representation of the patient.

In practice, a pathologist scans the biopsy, the model runs in under a minute, and the oncologist receives a risk-stratified treatment simulation before the FISH results even come back. The FISH results, when they arrive, get folded into the twin to refine the projection.

From MRD kinetics to early intervention

The most clinically useful output of myeloma digital twins may be MRD trajectory modeling. MRD — detecting one myeloma cell among 100,000 or a million normal cells — has emerged as the strongest surrogate endpoint for long-term outcomes. Patients who reach sustained MRD negativity at 10^-6 sensitivity have dramatically better progression-free survival.

But MRD is measured in snapshots. A patient might test negative at month 6, negative at month 12, then positive at month 18. That rebound signal is what matters, and traditional statistical models cannot tell you much about when it will happen or how fast it will grow once it does.

Digital twins simulate MRD kinetics continuously. They can flag patients with high projected rebound velocity and recommend earlier intervention — typically lenalidomide maintenance intensification, switching to a different immunomodulator, or early bispecific antibody referral — before the clinical relapse is obvious. Centers running these models are reporting meaningful gains in time-to-next-treatment in early real-world data.

This connects to broader progress in digital twins replacing control groups in clinical trials, where simulated patient cohorts reduce the ethical and logistical cost of randomization.

Quadruplet versus triplet: letting the twin decide

One of the hottest myeloma debates in 2026 is whether to use a four-drug induction (daratumumab plus bortezomib, lenalidomide, and dexamethasone — “Dara-VRd”) versus the older three-drug standard (VRd). The quadruplet produces deeper responses but adds toxicity, cost, and peripheral neuropathy risk.

Digital twins are now being used to individualize this choice. For a patient with high-risk cytogenetics and good performance status, the twin might project a 22 percentage-point PFS advantage for the quadruplet that clearly outweighs added toxicity. For an older patient with frailty markers, the same simulation might show the marginal benefit shrinks to 6 points while toxicity doubles — a case for the triplet.

This is the kind of question clinical trials cannot answer at the individual level. Trials report averages across enrolled populations. Digital twins let you ask the same question for this patient, with these cytogenetics, this renal function, and this social support profile.

Leukemia applications and where the models cross over

The same approach is being adapted for acute and chronic leukemias. In acute myeloid leukemia (AML), digital twins are being trained on multicenter cohorts of roughly 1,400 patients to predict complete remission and two-year overall survival with clinical, lab, cytogenetic, and molecular data combined. AML moves faster than myeloma — induction decisions happen in days, not weeks — so twins for AML focus on earlier, more aggressive triage questions: is this patient fit for intensive induction, or should they go to venetoclax-azacitidine?

Chronic myeloid leukemia (CML) is an interesting contrast. CML is one of oncology’s great success stories — tyrosine kinase inhibitors turned it into a chronic disease for most patients. But about 15-20% of patients develop resistance, and identifying them early is hard. Neural networks trained on frequent BCR-ABL transcript measurements are now showing strong relapse prediction performance, and some centers are testing twin-guided TKI switching before clinical resistance emerges.

These use cases share a common thread with other oncology AI work: the most valuable predictions are not “will this patient relapse?” but “when, and what do I do about it?” The same shift is visible in AI that predicts cancer drug resistance before treatment starts.

The integration problem

The technical wins are real, but deployment is the hard part. A myeloma digital twin needs access to EHR data, pathology systems, imaging archives, genomic databases, and specialty lab results — all of which live in different silos at most hospitals. The integration work is brutal and unglamorous. One major academic center spent 14 months building the data pipeline before a single twin prediction was generated.

Regulatory classification is also unclear. The FDA has guidance for AI/ML-enabled medical devices, but digital twins sit awkwardly between decision support (looser oversight) and diagnostic classification (tighter). A twin that recommends a specific regimen is doing something more than flagging abnormal lab values, and regulators are still figuring out the right framework — a question that connects to the FDA’s broader redefinition of “breakthrough” for AI medical devices.

Reimbursement is the quietest but possibly biggest barrier. CMS has not established clear billing codes for digital twin-guided therapy selection. Academic centers can absorb the cost through research budgets; community oncology practices — where most myeloma patients actually get treated — cannot.

What the next 18 months look like

Three developments are worth watching. First, the major myeloma cooperative groups are building prospective trials that randomize between twin-guided and standard-of-care treatment selection, with MRD and PFS as coprimary endpoints. Results from the first of these are expected in late 2027.

Second, CAR-T and bispecific antibody sequencing is becoming the hardest treatment question in myeloma, and digital twins are being positioned as the tool to resolve it. With four approved bispecifics, two CAR-Ts, and growing evidence that sequence matters enormously, the combinatorial question is exactly what twins are designed to handle.

Third, federated learning consortia are emerging to pool twin-training data across institutions without moving protected patient data. The International Myeloma Working Group has a federated pilot running across 11 centers. If it works, it solves the data-scale problem that has held back every previous attempt to build generalizable myeloma AI.

None of this replaces the oncologist. Myeloma patients still need someone to sit across from them and explain the plan, adjust for toxicity in real time, and make the judgment calls a model cannot. But the decisions those oncologists are making are becoming measurably better when a well-calibrated twin is running in the background. For a disease that has resisted personalization for decades, that is a real shift.