Digital Twins in Healthcare: Can SIMCO Model Preterm Birth Prevention?

Digital twins in healthcare can connect an untreated immune state, simulated drug perturbations, and time-to-labor predictions. This is a conceptual illustration, not an original figure or efficacy result from the paper. LuTra Studio

Digital twins in healthcare are bringing an engineering idea into biomedical research: when large exploratory drug trials are difficult or ethically constrained, can we first create a computational counterpart of an individual's biological system and compare how different interventions might change its trajectory?

On October 21, 2026, *Science Advances* published “Single-cell–level digital twins for preterm birth prevention strategies.” The researchers introduced SIMCO—Simulated Immunome Modeling of Clinical Outcomes—a framework that connects ex vivo whole-blood drug screening, mass cytometry, single-cell optimal transport, and clinical outcome modeling.

The study is compelling, but its central message is not that “AI has found a drug that prevents preterm birth.” A more accurate interpretation is that SIMCO offers a way to rank candidate interventions and generate mechanistic hypotheses. It predicted that some interventions could shift modeled time to labor in existing datasets; it did not demonstrate treatment efficacy, safety, or preterm-birth prevention in pregnant patients.

Why build a digital twin for preterm birth prevention?

Preterm birth is generally defined as birth before 37 completed weeks of pregnancy. The World Health Organization estimates that 13.4 million babies were born preterm in 2020—roughly 1 in 10 live births—and complications of prematurity remain a major cause of death among children under five.

Spontaneous preterm birth is not a single disease. Infection, cervical structure, placental dysfunction, chronic maternal conditions, inflammation, and other mechanisms may contribute, and many cases have no clearly identified cause. Pregnancy also requires a dynamic immune balance: maternal–fetal tolerance must be maintained early, while inflammatory programs rise as labor approaches. Premature or dysregulated activation of this “immune clock” may be associated with spontaneous preterm birth.

Pregnancy trials face unusual safety, ethical, enrollment, and cost constraints. A blood-based model that helps researchers reject weak candidates, identify potential responder groups, and prioritize interventions for animal and prospective clinical studies could reduce development risk. That is the problem SIMCO is designed to address.

What are digital twins in healthcare?

Digital twins originated in engineering. A virtual model of an aircraft engine, factory machine, or building receives data from its physical counterpart, updates its state, and helps predict failure or compare operating strategies.

In healthcare, the physical twin is a patient—or a defined biological system within that patient—and the digital twin is a virtual representation built from individual data. Under a strict definition, a clinically meaningful digital twin should be:

1. Personalized: It represents a specific individual, not only a population average.

2. Dynamically updated: It can be recalibrated when laboratory, imaging, omics, wearable, or treatment-response data change.

3. Predictive and decision-oriented: It compares future risks or interventions in a way that supports a testable decision.

That means not every patient avatar, 3D organ model, or machine-learning risk score qualifies as a digital twin.

Model: Electronic health record · Primary function: Stores clinical history · Key difference from a full digital twin: Records the past but may not simulate interventions

Model: Static virtual model · Primary function: Represents an organ, cell, or disease structure · Key difference from a full digital twin: May not update with new individual data

Model: AI prediction model · Primary function: Predicts risk or class from population data · Key difference from a full digital twin: May be a one-way input–output system with no synchronization

Model: Digital twin · Primary function: Maintains a personalized state, updates it, and compares “what if A versus B?” · Key difference from a full digital twin: Must be validated for prediction, updating, and its intended decision

A healthcare digital twin is not a static record. It connects personal data, computational models, intervention scenarios, and future updates in a feedback loop. Concept illustration, not an original figure from the paper. LuTra Studio

This is a demanding standard. A 2025 scoping review of 149 human digital-twin studies found that only 18 met all the authors' criteria for personalization, dynamic updating, and clinical decision support. Many systems called digital twins today are better described as digital models, digital shadows, or one-time in silico simulations.

SIMCO should be viewed along this spectrum. It produces individualized untreated and drug-treated immune twins and supports counterfactual simulation. But it does not continuously ingest each participant's post-treatment data or recalibrate inside a clinical workflow. It is therefore most accurately described as a proof-of-concept single-cell immune digital-twin framework, not a complete virtual replica of a pregnant patient.

What does AI add to a digital twin?

Traditional digital twins often rely on physical equations, mechanistic models, or predefined biological rules. These models make their assumptions explicit, but the human body spans molecular, cellular, organ, and behavioral scales, with nonlinear interactions and many unknown parameters.

AI can expand what a digital twin can do without replacing biological reasoning:

1. Integrate heterogeneous data. Machine learning can combine imaging, laboratory values, electronic health records, genomics, single-cell data, and wearable signals to estimate a patient's current state.

2. Learn individual variation. Representation learning and foundation models can learn shared structure from large populations and adapt it using smaller amounts of individual data.

3. Simulate counterfactuals. Generative models, causal inference, and perturbation models can compare no treatment, treatment A, treatment B, or different doses—not merely predict a single risk score.

4. Recalibrate over time. New measurements or observed treatment responses can update parameters and reveal model drift.

5. Translate complex outputs into decisions. Knowledge graphs and language models can organize records, literature, and model results. A language model, however, is not itself a validated physiological simulator.

The most credible direction is a hybrid digital twin: physiology, pharmacology, and biological constraints define what inferences are plausible, while AI helps with high-dimensional patterns, missing parameters, and computational speed. Both components need uncertainty estimates and validation for a clearly stated context of use.

Where are AI-enabled digital twins being applied today?

The practical goal is not an all-knowing replica of the entire human body. It is a task-specific twin built around a measurable, testable decision.

Application: Drug discovery and repurposing · Question a twin may help answer: Which cells, pathways, or organ states may change after an intervention? · Reasonable current role: Candidate ranking, mechanism hypotheses, and off-target screening

Application: Dose and regimen exploration · Question a twin may help answer: How might exposure–response change across dose, schedule, or combination? · Reasonable current role: Trial-design support, not a replacement for pharmacokinetics, toxicology, or human dose escalation

Application: Clinical-trial design · Question a twin may help answer: Which patients are likely to progress, and what placebo outcome is expected? · Reasonable current role: Prognostic covariates, enrichment, or synthetic/external-control research, subject to regulatory acceptance

Application: Organ-specific planning · Question a twin may help answer: How might cardiac flow, tumor growth, or a surgical path change under different interventions? · Reasonable current role: Focused organ or procedural simulations with testable outputs

Application: Chronic-disease monitoring · Question a twin may help answer: Do new glucose, activity, sleep, or laboratory data indicate a changed trajectory? · Reasonable current role: Dynamic risk prediction and personalized management requiring prospective validation

Application: Precision obstetrics · Question a twin may help answer: How might a patient's immune state shift under candidate interventions? · Reasonable current role: Early candidate ranking and biomarker-enriched trial design, as illustrated by SIMCO

FDA recognizes modeling and simulation as tools that can complement in vitro, animal, and clinical evidence. A 2024 FDA–CTTI public-workshop presentation also discussed a review scenario in which AI-generated digital twins estimated participant-specific placebo outcomes and were combined with prognostic covariate adjustment. This was a workshop example—not an FDA approval of a specific digital-twin method—and it does not mean that a digital twin can simply replace a placebo arm. Data provenance, representativeness, endpoints, prespecified analyses, bias, uncertainty, and the consequences of model failure must all be assessed for the intended use.

The same restraint applies to generative AI. It may help parse unstructured records, connect knowledge, propose testable scenarios, or explain model outputs. Without real biological data, mechanistic constraints, and prospective validation, a fluent output remains a plausible narrative—not a patient-ready digital twin.

How did SIMCO build single-cell digital twins?

The study can be understood in four layers.

SIMCO connects ex vivo whole-blood perturbation, a single-cell immune atlas, CellOT transformations, and time-to-labor modeling to rank candidate strategies. Concept illustration, not an original figure from the paper. LuTra Studio

1. Build a pregnancy immune-perturbation atlas

The researchers collected whole blood from 8 nonpregnant and 11 pregnant participants at 20 to 27 weeks' gestation. Each sample entered a 96-well ex vivo design combining eight immune-assay states—seven stimuli plus one unstimulated condition—with nine candidate interventions and experimental controls. The candidates comprised eight individual drugs (chlorthalidone, maprotiline, metformin, lansoprazole, pravastatin, rifabutin, salicylic acid [SA, the major metabolite of aspirin], and 5-methyltetrahydrofolate [THF, an active form of folate]) plus the SA–lansoprazole combination (SALPZ).

Mass cytometry measured 28 immune-cell populations, 12 phosphoproteins, and 8 activation markers. The resulting atlas covered approximately 218 million leukocytes and 80 perturbation scenarios.

2. Learn how untreated cells might respond

Single-cell assays cannot usually measure the same cell before and after treatment because the measurement consumes the cell. CellOT uses neural optimal transport to learn a probable transformation between untreated and treated cell distributions, then applies that transformation to an unseen baseline sample.

The resulting twin is not a simulation of the entire patient, placenta, cervix, and fetus. It is a computational counterpart of peripheral immune state at one point in time. The same baseline sample can generate an untreated twin and several candidate drug-treated twins.

3. Project the model into an independent longitudinal cohort

SIMCO then used 177 blood samples from 63 pregnant individuals: the term group contributed 153 samples from 58 participants, whereas the preterm group contained only 14 samples from 5 participants. Samples were collected within 120 days before labor, with one to three time points per person; the study's primary outcome was time to spontaneous labor onset.

The model simulated stimulation and candidate-drug responses from each unstimulated sample. Stabl feature selection identified stable, interpretable immune features for a model predicting time to labor.

4. Compare treated and untreated twins

For each participant, the predicted drug effect was the difference between the treated twin and untreated twin. This paired design may reduce errors shared by both predictions, but it remains an in silico comparison—not a randomized clinical-trial treatment effect.

What were the main results?

In untreated digital twins, predicted time to spontaneous labor onset reached Pearson R = 0.64 with an RMSE of 24.1 days. Combining sampling time with predicted time to labor yielded an AUROC of 0.75 (95% CI, 0.59–0.88) for distinguishing term from preterm samples. The AUPRC was only 0.29 (95% CI, 0.12–0.54), however, and the independent dataset included just 14 preterm samples from 5 participants. The result supports feasibility; it is not sufficient for a clinical risk tool.

The three highest modeled gestational-length proxy effects were:

Candidate strategy: THF · Median predicted delay proxy: 24.7 days (IQR 11.4–45.9) · Participants predicted to benefit: 90.5% · Interpretation: Highest modeled value, not an observed extension of pregnancy

Candidate strategy: Maprotiline · Median predicted delay proxy: 18.8 days (IQR 8.6–27.5) · Participants predicted to benefit: 82.5% · Interpretation: High model ranking, but discontinuation and adverse effects reduce translational appeal

Candidate strategy: SA + lansoprazole (SALPZ) · Median predicted delay proxy: 17.3 days (IQR 3.5–33.8) · Participants predicted to benefit: 87.3% · Interpretation: Repurposing rationale, but no efficacy evidence in pregnant patients

SA alone was predicted to benefit about 46% of participants. Across the full perturbation profile for SALPZ, 79% of features were classified as additive, 3% synergistic, and 18% antagonistic. It would therefore be inaccurate to describe the combination as broadly synergistic.

The abstract emphasizes THF and SALPZ, while the numerical top three in the main text also include maprotiline. This is not a contradiction; it reflects translational judgment. Fatigue, weight gain, anticholinergic effects, and debate around antidepressant use during pregnancy make maprotiline a less attractive preventive strategy despite its modeled ranking.

What immune mechanisms did the model suggest?

The time-to-labor model used 21 immune features. Five of the top 10 involved CD4-positive T-cell signaling, including increased pSTAT3 responses under some stimulation conditions as labor approached. Granulocyte pS6 signaling showed the opposite gestational trend.

In simulation, THF and SALPZ more consistently reduced pSTAT3 in naïve CD4-positive T cells and increased pS6 in granulocytes. The authors proposed two possible pathways:

1. Reduced JAK–STAT3 activity might decrease proinflammatory T-cell differentiation and prolong maternal–fetal tolerance.

2. Increased mTOR–pS6 signaling might preserve a more anti-inflammatory state in innate immune cells.

These are biologically plausible hypotheses, not proof that pSTAT3 or pS6 is a causal switch for preterm birth. Peripheral blood changes may reflect downstream signals from the placenta or maternal–fetal interface rather than upstream disease drivers.

How does this connect to existing evidence?

The nine strategies did not come from an unrestricted AI search. They were selected from the team's 2020 *JCI Insight* computational drug-repurposing pipeline, which looked for drugs that might reverse a preterm-birth gene-expression signature. Lansoprazole improved fetal viability in an LPS-induced fetal-wastage mouse model. That result came from an inflammatory animal model, not a prevention trial in pregnant patients.

Low-dose aspirin has established pregnancy use, but the current U.S. Preventive Services Task Force recommendation applies to patients at high risk for preeclampsia, beginning after 12 weeks of gestation at 81 mg/day. It should not be reframed as a universal treatment for spontaneous preterm-birth prevention.

The standard public-health purpose of folic acid is prevention of neural-tube defects. USPSTF recommends 0.4 to 0.8 mg daily for people planning or capable of pregnancy. In the SIMCO paper, Table 2 lists an 800-μg reference dose for folic acid, while the Discussion describes a 4-mg high-dose assumption for THF. That discrepancy requires clarification from the authors. The model does not justify self-directed high-dose folate use for preterm-birth prevention.

The paper also refers to historical FDA pregnancy categories A and B. FDA began replacing the A, B, C, D, and X letter categories under the Pregnancy and Lactation Labeling Rule in 2015. Current assessment should use each product's risk summary, clinical considerations, and supporting data rather than treating a historical letter as a present-day declaration of safety.

Digital twins may narrow the candidate search space, but modeled outputs still require preclinical and prospective clinical validation before they can support efficacy claims. Concept illustration, not an original figure from the paper. LuTra Studio

What has the study not proven?

1. The sample was small—especially the preterm group

The perturbation atlas included only 11 pregnant participants. The clinical dataset included 63 participants, but only 5 had spontaneous preterm birth. Generalizability across populations, infections, placental disease, gestational windows, and multiple pregnancies remains unknown.

2. Acute ex vivo signaling is not a long-term in vivo response

Blood was exposed to a drug for one hour and then stimulated for two hours. This captures acute phosphosignaling and some secondary signaling, but not hepatic metabolism, placental transport, tissue distribution, chronic dosing, infection defense, or fetal exposure.

3. “24.7 days” is not a clinical efficacy estimate

The study summed predicted effects across up to three sampling time points as an approximation of repeated dosing. The authors describe SIMCO as more suitable for comparing strategies and generating mechanisms than for exact quantitative forecasts. These proxy values cannot be reported as the number of extra days a medication would add to a pregnancy.

4. Peripheral blood is not the maternal–fetal interface

Critical events may occur in the decidua, placenta, fetal membranes, cervix, or amniotic compartment. Blood is scalable and minimally invasive, but it may miss tissue-specific mechanisms.

5. There was no prospective treatment-response validation

Participants did not receive THF, SALPZ, or the other candidate treatments so that predicted and observed responses could be compared. Future work requires external cohorts across pregnancy, appropriate animal models, prospective clinical studies, and assessment of infection risk, fetal safety, and long-term outcomes.

6. Patent and competing-interest disclosures matter

Several authors reported advisory or executive roles at SurgeCare, and Stanford University filed a patent application related to the work. These disclosures do not invalidate the study, but they belong in the evidence context when evaluating translation or commercialization claims.

The real advance is prioritization—not prescribing

The paper's most interesting contribution is not that a familiar compound suddenly became a preterm-birth treatment. It is the proposed order of research:

1. Build a high-dimensional single-cell perturbation atlas from small blood samples.

2. Learn heterogeneous responses across cells and signaling pathways.

3. Project those rules into an independent cohort with clinical outcomes.

4. Compare each person's treated and untreated immune twins.

5. Rank system-level effects before selecting candidates for preclinical and human studies.

This is closer to the drug-development question than a conventional biomarker alone. It asks not only who may deliver preterm, but which immune intervention might move an individual's modeled system toward an earlier gestational state.

That promise makes evidence boundaries more important, not less. SIMCO is currently a proof-of-concept digital-twin framework, not a clinical decision-support device. THF and SALPZ are candidates for validation, not proven prevention therapies.

Conclusion: narrow the search space, but do not skip clinical validation

The challenge in pregnancy drug research is not a shortage of candidate molecules. It is deciding, safely, which candidates deserve the next experiment. By connecting single-cell immune profiling, optimal transport, and clinical outcome modeling, SIMCO provides a more granular prioritization strategy than average immune responses alone.

If replicated in larger, more diverse cohorts across pregnancy—and validated against prospective treatment-response data—it could become an early research tool for precision obstetrics. It might support biomarker-enriched trials, identify off-target immune effects, and reduce the number of weak candidates entering pregnancy studies.

For now, the responsible conclusion is narrower: this study shows that single-cell immune digital twins can generate testable hypotheses for preterm-birth prevention. It does not show that any simulated candidate safely or effectively prevents preterm birth in clinical practice.

References

1. Einhaus J, et al. Single-cell–level digital twins for preterm birth prevention strategies. *Science Advances*. 2026;12:eaed2741. DOI: 10.1126/sciadv.aed2741.

2. World Health Organization. Preterm birth. 2023.

3. Bunne C, et al. Learning single-cell perturbation responses using neural optimal transport. *Nature Methods*. 2023;20:1759–1768.

4. Le BL, et al. Computational discovery of therapeutic candidates for preventing preterm birth. *JCI Insight*. 2020;5:e133761.

5. U.S. Food and Drug Administration. Pregnancy and Lactation Labeling Resources.

6. U.S. Preventive Services Task Force. Aspirin Use to Prevent Preeclampsia and Related Morbidity and Mortality. 2021.

7. U.S. Preventive Services Task Force. Folic Acid Supplementation to Prevent Neural Tube Defects. 2023.

8. SIMCO authors. SIMCO source code. GitHub; accessed October 22, 2026.

9. Tudor BH, et al. A scoping review of human digital twins in healthcare applications and usage patterns. *npj Digital Medicine*. 2025;8:587.

10. De Domenico M, et al. Challenges and opportunities for digital twins in precision medicine from a complex systems perspective. *npj Digital Medicine*. 2025;8:37.

11. Coveney P, et al. Digital twins and Big AI: the future of truly individualised healthcare. *npj Digital Medicine*. 2025;8:494.

12. Venkatesh KP, et al. Health digital twins as tools for precision medicine: Considerations for computation, implementation, and regulation. *npj Digital Medicine*. 2022;5:150.

13. U.S. Food and Drug Administration. Modeling & Simulation at FDA. Accessed October 22, 2026.

14. U.S. Food and Drug Administration. Artificial Intelligence/Machine Learning: The New Frontier of Drug Development & Regulation. FDA–CTTI public-workshop material; 2024.

Fact-checked through October 22, 2026 (America/New_York). This article is a scientific-research interpretation, not medical advice. Decisions about medication, folate dose, or aspirin use during pregnancy should be made with an obstetric or maternal–fetal medicine professional based on individual risk.