Toward ‘digital twins’ of FA individuals

Studies on SCC prevention in FA are limited by the small number of individuals with the condition, who are spread around the world. In addition to regular histopathological diagnosis of oral cancer development and cytology-based screening methods (32), reliable molecular markers are limited. Moreover, the scarcity of genotype-phenotype associations in FA makes it highly likely that each patient will respond in an individual way to drug treatments and/or lifestyle changes. Since robust predictive in vitro and in vivo FA models are lacking, drug screening and testing cannot be generalized for all FA individuals. For example, in vitro analysis of radiation sensitivity of fibroblasts from FA individuals does not correlate with the clinical response of the same patient to radiotherapy (69, 70) and the amount of chromosomal breaks found in lymphocyte cultures does not correlate with the severity of the disease, e.g., bone marrow function of the individuum. These issues motivate the effort to create multi-level, dynamical computational models of FA that can aid clinicians in tailoring therapies to each specific FA patient. Models of this type have been termed “medical digital twins” (71, 72).
Although a consensus definition of a medical digital twin does not yet exist, the concept of a digital twin is common in engineering disciplines (73). Often referred to as “industrial digital twins,” these models are computational replicas of complex devices or processes, such as jet engines or wind turbines, that are used to diagnose technical problems and guide interventions. Industrial digital twins are typically composed of multiple, interconnected computational models of the constituent components of an engineered system. Critically, this integrated “template” model of the base processes of the engineered system in question is subsequently tuned, or “calibrated,” to a specific instance of that system, e.g., a particular jet engine, using performance data collected from sensors in real time. It is this “twinning” process, involving consistent feedback from real-time data streams, that differentiates a multi-level, computational model of a dynamical system from a true digital twin (74). Construction of digital twins for medical and clinical applications has been receiving increased interest lately (71, 72). However, biological systems are far more complex than engineered systems, making their practical implementation much more challenging. Nevertheless, there have been successful applications of medical digital twins for the treatment of type 1 diabetes (75) and pediatric cardiac patients (76). Furthermore, it is important to note that medical digital twins differ from alternative approaches gaining popularity in biomedical sciences, such as statistical and machine-learning models (77), in that they are based on a mechanistic understanding of the underlying biological system. As such, they are not constrained by the confines of the experimental data on which they are constructed, which in FA is sparse.
The utility of a FA medical digital twin will be to aid clinicians in determining best courses of action for individual patients in both the prevention and treatment of malignant tumors. The template model for an FA medical digital twin will comprise the biological processes mentioned previously, including microbiome interactions, DNA damage sensing and repair, EMT, cell cycle progression, and cell death, among others (Fig. 3). Calibrating the template model to individual FA patients will be challenging and require collecting spatially resolved, single-cell resolution multi-omics data, epigenome profiling, and metagenomics of the oral microbiome, from patients at regular intervals,e.g., every three months in accordance with clinical care guidelines for FA individuals with potentially premalignant lesions in the epithelial tissue. Additional patient data, such as blood draws, genetic screens, and oral swabs, together with standard data from electronic health records, can also be integrated into the calibration data stream. Once the model is personalized in this way, it will be possible to test, in silico, numerous preventive and/or therapeutic options before applying them clinically to FA patients. Furthermore, medical digital twins should be flexible and extensible, able to grow in precision and predictive power as new knowledge is accrued and experimental data sets are generated. In this way, the FA digital twin will develop together with the patient and their clinician, ultimately forecasting, with high accuracy and precision, responses to novel personalized interventions and therapies.