Read on Twitter
1) A thread on our new paper with Lance Munn, Triantafyllos Stylianopoulos, @ankitbpatel7 Rakesh Jain and @MGHSteeleLabs “In Silico Dynamics of COVID-19 Phenotypes for Optimizing Clinical Management” ( https://www.pnas.org/content/118/3/e2021642118 ) which is out in PNAS today.
2) Our paper describes a comprehensive mathematical model of COVID-19 viral dynamics and the host immune response which is based on the Jain group’s previous published work.
3) Let me explain why I think modeling, which is obviously very far from clinical research, is essential to understanding not just COVID-19 but critical illness in general.
4) Almost no attempts to develop specific therapies for critical illness syndromes such as ARDS and septic shock have been successful. In part, this is due to two related features of critical illness – complexity and heterogeneity.
5) Heterogeneity means critical illness syndromes are not single diseases but syndromes in which patients who are clinically similar (presenting with b/l infiltrates and low P:F, for instance) may nonetheless be suffering from qualitatively different pathophysiology.
6) Complexity refers to the fact that critical illness is time-varying and multi-dimensional. Multiple cytokines and diverse cell types interact with pathogen kinetics, clotting cascades, therapies and patient co-morbidities to produce a clinical presentation.
7) Heterogeneity is a problem for clinical trials because if too diverse a group of patients are enrolled an overall negative result can be masking a sub-group of patients who benefit, or an overall positive result can be masking a sub-group of patients who are harmed.
8) Complexity is a problem because treatments may have unexpected effects (increasing mortality in one famous example ( https://pubmed.ncbi.nlm.nih.gov/14707556/ ) or only work at certain times in the disease course.
9) We can combat heterogeneity by enriching trial populations in similar patients – the Calfee group ( https://pubmed.ncbi.nlm.nih.gov/30291376/ ) has done pioneering work using biomarkers to do this – but the identification of makers may be limited by available clinical data.
10) Complexity can be sorted out over time by repeating trials in different subgroups (immunosuppression in ICU patients, but not patients not on oxygen) but trial and error is a lengthy and expensive way to go about developing new treatments.
11) Modeling can help with both complexity and heterogeneity. In a model, you can study an unlimited number of pathways and follow cytokines, cell populations, viral load etc. continuously in time – not just at the discrete time points which may have been clinically measured.
12) You can also incorporate a theoretically unlimited number of clinical features and biomarkers and study their association with outcome or particular biologic events (worsening oxygenation, balance of adaptive vs innate immune response).
13) Now the obvious (and correct) objection is that computers are not patients. But models can be a rigorous method of hypothesis generation. Humans are notoriously bad at determining causality in complex environments and that can lead to expensive and negative trials.
14) We now know for sure that just because elevated IL-6 is associated with poor outcome in COVID-19 it does not follow that inhibiting IL-6 will always improve outcome. However, we are easily convinced that these things must be true because it just ‘makes sense’.
15) Sensible seeming conclusions are a bad way of developing therapeutic hypothesis because we implicitly assume linearity and simplicity. In reality, there are numerous time-varying pathways simultaneously active in a critically ill patient.
16) It such a situation it is not at all clear how changing one variable will affect the overall evolution of the system, but we can write down equations describing the dependencies of various factors and follow the evolution of all of them after a simulated intervention.
17) In this way we make predictions about which biomarkers define distinct biological states (endotypes) and which interventions at which times are most likely to help.
18) The predictions of the model must then validated against clinical data, of course, but hypotheses guided by a mathematical theory are likely to be more accurate and sophisticated than the simple cause and effect stories we tell ourselves without the aid of the equations.
19) As an example, our COVID-19 model incorporates the renin−angiotensin system, ACE2, key elements of the innate and adaptive immune responses, the role of inflammatory cytokines, and the coagulation cascade for thrombus formation.
20) The model predicts the evolution of viral load, immune cells, cytokines, thrombosis, and oxygen saturation based on patient baseline condition and the presence of co-morbidities. The model was developed and validated against clinical data.
21) We found that the outcome of any treatment depends on the sustained response rate of activated CD8+ T cells and sufficient control of the innate immune response.
22) We further found that the optimal treatment strategy varies by time point – immune stimulation may actually be helpful early, but immunosuppression is helpful late in severe cases.
23) These insights may seem intuitive, but the model also enables us to make detailed predictions about which biomarkers correspond to which optimal treatment and which patients are more or less likely to benefit from a given treatment at which time.
24) There is much refinement to be done – the model does not yet incorporate antibody response for example. But we feel that this paper is a strong argument for the role of theoretical models in hypothesis generation for clinical research.
