Multiscale computational modeling of complex biological systems has been an active field of research, leading to notable advances in both mechanistic and agent-based disease models. The overarching goal is quantitative representation of interconnected biological processes that cannot be easily delineated experimentally. Such model should be able to capture the spatiotemporal interdependencies across all scales—from genomic, transcriptomic, proteomic, and metabolomic scales to tissue and organ level, and ultimately to the individual and population levels (Bhattacharya et al., 2019; Tiwary, 2020). Recently, the interleaving of traditional computational modeling and simulation with artificial intelligence approaches has emerged as a strong theme to complex systems modeling (Alber et al., 2019; Tolk et al., 2022), with several applications in drug design, structural biology, and neuroscience to name a few (e.g., (Stephan et al., 2015; Huys et al., 2016; Romero et al., 2019; Liu et al., 2021; Paul et al., 2021; Ramanathan et al., 2021).

Since system complexity increases when bridging different scales, there are technical challenges as the number of system parameters rapidly increases. Furthermore, the integration of computational representations across different organ systems remains an outstanding challenge (Beard et al., 2012). Developing efficient computational tools that can manage multi-modal, multi-scale data as well as demonstrate effective utilization of high-performance computing resources are critical for us to overcome current barriers.

One outstanding and ever-increasing challenge with complex multi-scale computational models and the latest artificial intelligence models known as language models or transformers (Vaswani et al., 2017; Wei et al., 2022) is the growing demand for computational power (Liu et al., 2021). Model training, hyper-parameter optimization, uncertainty quantification, and validation become exponentially more demanding as efforts move from single to multiple scales (Renardy et al., 2019). As energy-efficiency becomes a bottleneck for large scale computational science, efficient algorithmic development will be necessary for scalable computational medicine. It has been proposed that a modular approach to multi-scale computational medicine, with interoperable and reusable computational tools, mimicking the first principles computational chemistry and physics modeling approaches maybe appropriate as it has been very successful in materials science (Ford Versypt, 2021). Although biological systems differ from physical systems, there is potential in investing in such effort to derive important building blocks that bridge a few spatiotemporal scales. Still, standards for reproducible research in computational medicine are lacking. Creating and maintaining data and model repositories are critical for ensuring reproducibility and reliability. As many of these endeavors become computationally very intensive, the burden of reproducibility is immense for the average researcher. We need to invest substantial resources in compute-and-data infrastructures to scientific integrity, reproducibility, and reliability of data and models.

The breadth of technical and algorithmic challenges exemplifies the need for strong collaborations across disparate scientific domains, across different and also competing approaches, as well as across different stakeholders. For example, in 2016, the National Cancer Institute (NCI) and the Department of Energy (DOE) in the US partnered in a collaboration to accelerate advances in predictive oncology. The collaboration brought together multidisciplinary experts in the biological, computational, data, and physical sciences to develop, demonstrate, and disseminate advanced computational capabilities that help answer driving scientific questions across molecular, cellular and population scales (Bhattacharya et al., 2019). The community effort is growing by adding new scientific challenges that build upon strong collaborations (Buchsbaum et al., 2022).

The concept of the digital twin has gained a lot of traction within computational medicine (Barbiero et al., 2021; Hassani et al., 2022). The digital twin is a virtual representation of a patient as a multi-modal system which incorporates patient data to inform personalized medical decisions related to disease prediction, diagnosis, therapeutic interventions, and prognosis. A digital twin can be created at different levels of detail (e.g., organ, individual, population, healthcare system) using various data sources as they become available. Merging non-traditional data sources (e.g., environmental factors, socioeconomic conditions, lifestyle choices) with multi-scale patient data, digital twins offer in silico modeling of patient health trajectories by taking into account the complex interplay of all factors and processes that affect wellbeing. Such models can be regularly interrogated to explore different scenarios (e.g., different treatments, lifestyle choices) to predict future risks and outcomes and empower individuals to make decisions at critical times and from the earliest stages in life. Furthermore, such longitudinal models can be dynamically adaptive and updated as new data becomes available. The potential to execute “what-if” scenarios completely in silico can be very empowering for patients, physicians, researchers, and healthcare systems as each tries to optimize outcomes based on individual criteria and incentives. References (Mulder et al., 2022; Okegbile et al., 2022) provides an insightful discussion of the potential of digital twins in the future of medicine while disease-specific examples are emerging (Allen et al., 2021; Hernandez-Boussard et al., 2021).

The clinical implementation of the digital twin in computational medicine is still in its very early stages facing similar challenges as those outlined in the previous section, namely data quality, data integration, reproducibility, reliability, as well as continuous quality assurance due to its dynamic nature.

With computational medicine’s great promise comes an even greater responsibility. We must recognize the pitfalls and possible ill-intended uses of the computational models. Since these models rely heavily on patient data, there are many legal and ethical considerations related to data collection, sharing, and use. Access to large amounts of patient data is fundamental to understand individual and population-level health outcomes over time. However, liberating and providing access to patient data is both a technological and a policy challenge. Furthermore, to create a richer picture, medical data must be combined with other data points to provide context on a patient’s living conditions that has substantial implications in predicting patient health trajectories. Although we all recognize the scientific value of human data, the debate over data ownership is ongoing in terms of how best to balance the promise of transparent innovation with the risks of unethical data handling, intentional or unintentional privacy breaches, and adversarial data use by hostile or malicious actors (Kostkova et al., 2016). To maintain a strong ethical framework for computational medicine, we need to answer this fundamental question: Who owns the intellectual property of data-driven computational models in healthcare? The patient? The medical center collecting the data by providing the healthcare services? Or the model developer? Clearly, no single entity alone could deliver the breakthrough technology.

Next is the topic of data trustworthiness. There is plenty of evidence that low data quality and problematic data representativeness can compromise model validity (Winslow et al., 2012) by creating or exacerbate existing racial or societal biases in healthcare systems (Nordling, 2019). During the development phase, scientists should promote a rigorous statistical framework to monitor for potential biases in the collected data. During deployment phase, model developers should implement rigorous quality control, monitoring model performance across subgroups to confirm robust performance or identify performance gaps. We should work to communicate to patients and healthcare providers openly and clearly what they should expect from the technology so that they are informed consumers of the technology.

Other than data quality, data-driven models benefit from access to large volumes of representative medical data. With the explosive growth of large-scale deep learning models (LeCun et al., 2015), the need for data sharing is pressing. State of the art deep learning models have billions of parameters. Although these models can push boundaries in learning and generalizability (Sun et al., 2017; Wang et al., 2019), they require massive amounts of training data due to their large parameter space. Federated learning has emerged as a successful collaboration mechanism to address the privacy constraint with sensitive data sharing. Instead of sharing data, the collaborative entities share the model parameters after local training and fine tuning (Li et al., 2020). Several studies demonstrated that models trained using federated learning are as accurate as those trained using centrally hosted data sets and they are far more accurate than models trained with single institution data. Nevertheless, federated learning is in its early stages of technical development. There are still outstanding concerns with reverse engineering of the trained “super-model”, and legal implications if the model is broadly shared and used without proper authorizations in clinical care (Rieke et al., 2020; Zawad et al., 2022).

Another grand challenge is that the clinical translation of computational models is not straightforward. If the computational models are used for hypotheses generation or for knowledge discovery, they are easier to embrace as they are used as scientific instruments. However, if computational models are used to perform or assist clinical tasks, then acceptance expectations are much higher. Often these models must be regulated, and they need to clearly demonstrate efficacy (i.e., performance equivalent to medical experts) and safety. Previous experience with AI for clinical decision support demonstrated that AI is capable of performing narrowly defined, repetitive tasks exceptionally well. Still though, if physicians over-rely on such decision support technologies, they may lose critical skills needed for performing more difficult tasks (Darcy et al., 2016). Another challenge with clinical integration is ensuring that the computational model is capable of assessing its confidence (i.e., uncertainty quantification) and providing a justification (i.e., explainability) for its prediction (Begoli et al., 2019; Benrimoh et al., 2021). Since one of the key drivers of the ‘digital twin” is to empower the patients as they try to manage their disease and gain better understanding of the short-term and long-term implications of the decisions they need to make, how to convey the model’s reasoning and prediction confidence to the patient vs the healthcare provider is an understudied topic.

Ultimately, humans and computational models will have to work well together. But this synergy will not happen organically, as past health AI experiences have demonstrated. It is important to train both healthcare providers and patients in how to use computational models responsibly, and how to remain vigilant avoiding mistakes of over-reliance when supported by the models. Objective benchmarking of datasets and models against community consensus metrics to detect, monitor, and possibly correct dataset biases or inconsistent model performance must become part of the practice of computational medicine.