Digital Twins As Control Groups to Accelerate Assessment of Safety and Efficacy of Novel Treatment Strategies in Ph-Negative Acute Lymphoblastic Leukemia

Introduction. Digital twins are defined as virtual representations of real patients, generated from multi-layer longitudinal data. In the medical setting, this approach can be leveraged to enhance diagnosis, prognosis, and prediction of optimal treatment within a specific patient population, enabling the simulation of virtual scenarios to improve the decision-making process. In B-lineage acute lymphoblastic leukemia (B-ALL), the prognosis of Philadelphia-negative (Ph-neg) patients is still suboptimal if compared with the results achieved in Ph-positive ALL (Foà et al, NEJM 2020 and JCO 2024). The frontline LAL1913 GIMEMA trial for adult Ph-neg ALL (Bassan et al, Blood Adv 2023) based on a pediatric-inspired and minimal residual disease (MRD)-oriented strategy provided promising results on 183 patients: complete remission (CR) was achieved in 91% of cases, a post-consolidation MRD negativity in 75.5%, and the 3-year overall survival (OS) was 67%. The protocol was later used in the real-life Campus ALL setting in 421 patients (Lazzarotto et al, P396 EHA 2024; Haematologica, in press) and the LAL1913-like scheme is currently considered in Italy as standard of care (SOC).

In the present study we aimed at generating digital twins of Ph-neg ALL to be used as a control group in in silico trials, as reported for acute myeloid leukemia (Piciocchi et al, EJHaem 2024).

Methods. Data from the GIMEMA LAL1913 study and real-life data of patients treated with a LAL1913-like therapeutic scheme were merged to form a unique cohort named realPh-neg cohort of 604 patients. To create the synthetic cohort, named synthPh-neg, a machine learning generative model was constructed from the real data of the realPh-neg cohort, capturing its patterns and statistical properties, without identifiable data from the original dataset. To develop the synthetic cohort, the R package “synthpop” (Nowok et al, J Stat Soft 2016) and different parametric methods were used to generate variables according to their types. Survival variables were simulated using classification and regression trees (CART). Next, we verified the adherence of the virtual cohort to the original one in terms of age, WBC count, gender, risk category, cytogenetic class, CR. Virtual and real cohorts were also compared in terms of survival outcomes and by stratifying according to lineage, age, risk category.

Results. Using the real-patient dataset, the synthPh-neg cohort of 3020 patients, corresponding to 5 times the realPh-neg cohort, was generated. The comparison between the original and synthetic cohorts showed that the clinico-biological characteristics and CR of the 2 cohorts did not differ. In detail, the CR rate was 92% in the realPh-neg cohort and 93% in the synthLAL1913 cohort (p=0.75). Moreover, OS and disease-free survival (DFS) were superimposable (p=0.69 and p=0.82, respectively). At 3 years the OS was 67.6% (63.4%-72.1%) in the original cohort and 68.7% (66.7%-70.6%) in the synthPh-neg cohort. DFS at 3 years was 59.5% (55.0%-64.3%) in the original and 58.7% (56.7%-60.8%) in the synthetic cohort. The similarities between the original and the synthetic dataset were maintained also when stratifying patients according to the biologic features at diagnosis.

Conclusions. We generated a large cohort of digital Ph-neg ALL patients which faithfully reproduces the features and outcome of the original dataset. Since the LAL1913 scheme is the current SOC in Italy, synthPh-neg may qualify as the optimal control group in in-silico trials when testing innovative treatments. In this case, enrolled patients would receive only the novel – potentially more active – treatment, thus allowing a quicker evaluation of its safety and efficacy. Indeed, though randomized trials represent the gold-standard, the rapid development of novel drugs requires the evolution of clinical trial design, especially in rare diseases/subsets for which the conduct of a randomized trial would be challenging.

In Ph-neg ALL, this is particularly relevant, given the biologic and clinical heterogeneity of ALL. To this end, GIMEMA has recently opened 2 phase II trials for the treatment of Ph-like ALL and very high-risk T-ALL respectively with ponatinib and daratumumab incorporated in the LAL1913 backbone (Eudra CT Number 2022-000633-17 and 2024-511627-34-00) . In the future, a subset of digital twins belonging to synthPh-neg ALL could be used as a control group in phase III virtual studies.