Digital twins in medicine, with a leukemia example

In engineering, a digital twin is a simulation of a physical object kept in step with the object's own measurements, so that a change can be tried on the model before it is made to the real thing. In medicine the object is a patient, or one organ or cell type of a patient, and the change is a treatment.

Digital twins in medicine

Several fields have working examples. In cardiology, models of a patient's heart built from imaging and electrical recordings are used to plan ablation and device therapy ^[1]. In type 1 diabetes, a simulator of 300 virtual patients spanning the measured variability of glucose metabolism was accepted by the FDA in 2008 as a substitute for animal trials when testing closed-loop insulin controllers ^[2]. In immunology, models of the immune response to viral infection are being built to test interventions per patient ^[3]. In systems medicine, network models of individual patients are treated computationally with thousands of drugs to rank candidates ^[4]; a 2024 scoping review maps the field ^[5].

The examples share four parts: a mechanistic model of the biology; measurements from the individual patient that set its parameters; an intervention that can be simulated; and a measured outcome to compare against. Without the fourth part a twin is a hypothesis, not a tool.

A twin of a leukemia cell

The Beat AML study of a BET inhibitor has all four parts ^[6]. The model is a network of the signaling, metabolic, epigenetic, and transcriptional pathways of a cancer cell, written as ordinary differential equations from published experimental data: about 112 pathways, over 75,000 reactions, and 3,300 cancer-relevant genes. The patient measurements are cytogenetics and whole-exome sequencing from 100 patients in the Beat AML cohort. Each finding enters the network as its functional consequence: a variant that more than five prediction algorithms call deleterious becomes a gain of function in an oncogene or a loss in a tumor suppressor, and genes in deleted or amplified chromosome segments are knocked down or over-expressed. Solving the network with these changes gives the patient's steady state, a map of which pathways are more or less active than in a non-malignant baseline.

The intervention is JQ1, entered as inhibition of its targets BRD2, BRD3, and BRD4, with downstream effects from the literature, chiefly reduced transcription of MYC. It is applied across a dose range and its effect is summarized as a disease inhibition score (DIS): the percentage reduction in a proliferation index built from the four active CDK-cyclin complexes, combined with the reduction in a viability index of survival markers (AKT1, BCL2, MCL1, and others) relative to apoptosis markers. The measured outcome is the ex vivo sensitivity of the same patient's cells to JQ1, an IC50 from a three-day assay.

Reading the result as four quadrants

Each patient is one point on a plot of predicted DIS against measured IC50. Two thresholds divide it. On the measured axis, an IC50 below 2.7 micromolar counts as sensitive; that is the peak plasma concentration of JQ1, so it separates concentrations a patient could reach from those a patient could not. On the predicted axis, a DIS of 30 percent or more counts as a predicted responder.

Of the 100 patients, 89 were sensitive in the assay and 11 resistant. The model placed 84 of the 89 and 9 of the 11 correctly, for 93 matches; the mismatches were five sensitive samples predicted resistant and two resistant samples predicted sensitive. Sensitivity was 94.4 percent and specificity 81.8 percent; the positive predictive value was 97.7 percent and the negative predictive value 64.3 percent ^[6]. Two cautions apply. The cohort was heavily weighted toward sensitivity, so the small resistant group carries most of the information about whether the model can tell the two apart. And the negative predictive value is the weakest figure: when the model called a sample resistant it was right about two times in three.

What the twin showed about the drug

The agreement rate validates a twin; the mechanism is what makes it useful, because the network can be read back to see why each patient landed where it did. JQ1 acts on the BET proteins that hold chromatin open at oncogene promoters, and blocking BRD4 lowers MYC transcription. The responder profiles show routes into that mechanism: one carried gains in BRD2 and BRD4 themselves and a loss of DUSP6 that leaves ERK active and MYC high; another carried a loss of NPM1, which normally restrains BRD4, so the BRD4-dependent program that JQ1 shuts off was running freely. The resistant profiles show routes around it: one had lost EP300, the acetylase that recruits BRD4 to chromatin, leaving less BRD4 activity for JQ1 to remove, together with an amplified AMPK subunit that promotes a survival response; another carried a gain in FGFR4, a receptor kinase that keeps a parallel growth pathway running regardless of BET inhibition ^[6].

Across the cohort the same picture held. Mutations that deregulate the ERK pathway (NRAS in 23 patients, DUSP6 in 9, KRAS in 7, NF1 in 6) sat predominantly among responders, and every patient with del(7q) or monosomy 7, trisomy 8, or del(5q) responded in both the model and the assay; trisomy 8 fits the mechanism directly, since MYC is on chromosome 8 ^[6]. These are candidate inclusion criteria for a trial, arrived at by reading the model rather than by fitting the outcome.

Where the twin stops

A twin of this kind contains only the biology curated into it; a variant with no known functional effect contributes nothing, and the paper attributes its false predictions to that and to incomplete genomic data. The input was DNA alone, so two patients with the same mutation list receive the same twin. The outcome was a laboratory assay, not clinical response; the same platform was later tested prospectively in 50 patients with AML or myelodysplastic syndrome, predicting the response to the prescribed treatment correctly in 55 of 61 cases ^[7]. Within those limits the example meets the engineering definition: a model set from one patient's measurements, an intervention tried in it, and a measured outcome it was checked against.