The real world is made of a complex web of interconnected events. The greatest understanding occurs when these events are studied as part of a multitude of systems across multiple scales. Multiscale modeling presents the promise of synthesis, yet the synthesis only holds if we are willing to confront the methodological, social, and philosophical discontinuities head-on. My goal has been to build models that are accountable—to data, to structure, to history—as opposed to building models that are merely complex. This accountability, in my view, is what multiscale modeling must do. Below are some of my thoughts in this area.

Multiscale modeling must consider the epistemological frictions that emerge when integrating data across biological, behavioral, and social domains. These frictions are not theoretical nuisances—they are structural realities that, if unaddressed, compromise model integrity and policy relevance. In my experience working with multi-omic data on malaria infection through the NIH-funded MaHPIC consortium, we developed within-host models that were constrained by, and responsive to, empirical ‘omics datasets collected longitudinally in nonhuman primates. This effort required bridging physiological time with immunological dynamics. Multiscale modeling cannot be an exercise in aesthetic complexity—it must maintain fidelity to the epistemic structure of the contributing sciences.

Multiscale modeling must include sociocultural structure as a constitutive variable, not as a contextual modifier. I argue this not from abstraction, but from practice. In my work at the intersection of epidemiology and quantitative modeling, I have developed frameworks that incorporate the impact of socioeconomic heterogeneity on infectious disease dynamics. My approach is what I describe as Asimovian, from a psychohistorian point of view: Models that are attentive to human structures and not merely biological mechanisms. This line of work has been supported by field collaborations and funding from NIH’s International Centers of Excellence in Malaria Research (CLAIM Project), where our models were used to interpret malaria dynamics in heterogeneous regions of Latin America.

Multiscale modeling must be transparent about its assumptions, especially when integrating social data. I have expressed this concern in multiple forums. Social data are often drawn from platforms that are not representative and that encode structural biases in usage, reporting, and interpretation. In my presentations on COVID-19 modeling and behavioral inference, I highlighted how model misidentification can emerge not just from technical error but from uncritical use of data sources that were not designed for population inference. This issue has been discussed in public health circles but remains under-theorized in much of computational epidemiology.

Multiscale modeling must include a rigorous strategy for causal identification if it aims to inform equity-aware interventions. In many proposals I have reviewed, there is an increasing push to derive policy from models without accounting for the assumptions necessary for causal inference—ignorability, non-interference, and measurement validity. These cannot be wished away with jargon or deferred to machine learning pipelines. My NIH-funded work on training biomedical research teams in data science emphasizes the methodological discipline needed for reproducibility and interpretability across scales (NIH award 1R25GM151182).

Multiscale modeling must evaluate interventions not only for efficacy but also for their embedded ethical and structural consequences. My participation in DARPA’s THoR’s HAMMER project on resilience in infectious disease was grounded in this principle. We explicitly modeled host responses under stress to examine how therapeutic strategies might interact with underlying immunological and social variance. These are not merely questions of optimization—they are questions of justification, requiring sensitivity to who benefits, who is burdened, and how knowledge is produced.

Finally, multiscale modeling must acknowledge its computational and empirical limits. In many high-dimensional models, sensitivity analysis is miscast as a matter of brute-force computation. The computational goal cannot be simply to eliminate uncertainty but to quantify it meaningfully and constrain it through real data.

Two Approaches to Multiscale Modeling

Multiscale models can be constructed in two ways, primarily. First, a single integrated model can be developed that encompasses one or more scales directly, from the molecular interactions of the parasite with its human and mosquito hosts, up to the population dynamics of disease spread. This approach offers the advantage of a unified framework that can capture direct interactions between different scales. However, it often faces challenges related to computational complexity and the integration of disparate data types. Alternatively, one can build a collection of models, each dedicated to a specific scale, and then link these models to address overarching questions that span multiple scales. This modular approach facilitates focused refinement at each scale and can be computationally more manageable. However, ensuring consistent and meaningful communication between the individual scale models can be intricate.

I have been studying both methodologies, with their respective advantages and drawbacks, and how they have provided valuable insights into malaria dynamics, highlighting the versatility and adaptability of multi-scale modeling techniques. In my opinion here is the ABC of multiscale modeling:

• Asimovian: «In the year 12,067 of the Galactic Era (G.E.), Hari Seldon, mathematician and psychologist, developed psychohistory, a new field of science that allows the prediction of future events through accurate modeling based on sociology, psychology, mathematics and statistics.» Premise behind Isaac Asimov’s “Foundation”, 1951 (-18 G.E.) The Asimovian approach is to include sociocultural factors into modeling, in addition to the traditional approaches.
• Boxian: «Since all models are wrong… simple but evocative models are the signature of the great scientist.» Box, George E. P. (1976), “Science and statistics”, Journal of the American Statistical Association, 71 (356): 791–799. The Boxian approach is to make models as simple as possible, yet insightful.
• Coxian: «The idea that complex physical, biological or sociological systems can be exactly described by a few formulae is patently absurd.» Cox, David. R. (1995), “Comment on “Model uncertainty, data mining and statistical inference””, Journal of the Royal Statistical Society, Series A, 158: 455–456. The Coxian approach is to include more complexity into models.

I am personally inclined to follow the Asimovian approach. That is, one must absolutely include sociocultural factors into multiscale modeling. Making this statement is easy, but following through… that is where the rubber meets the road. I have developed a few approaches to do this, but it is the subject of another post.