Mechanistic Modeling
Immunetrics prides itself on careful integration of mathematics and software with biology. Our disease state models are built with the strictest adherence to state-of-the-art biological knowledge.
Mechanistic modeling of biological systems is rooted in two basic premises: Every observed phenomenon is the sum of multiple inter-connected biological processes, and when the most significant processes are represented mathematically, the simulated output resembles the actual observations.
Alternative modeling approaches certainly exist. Statistical models can identify patterns and correlations between observed variables in large data sets. However, they are completely dependent on the data to which they are trained, and will not reveal fundamental causality between variables or any dynamic aspect of biological processes. Phenomenological models are a subset of statistical models. Although they can be more dynamic in nature, with outputs that vary over time, phenomenological models are still data-dependent and cannot predict behaviors independent of the original training data.
Because of the limitations of alternative modeling strategies, Immunetrics believes that mechanistic modeling is the best choice for building biological models. Mechanistic models are based upon the most comprehensive set of available biological knowledge, which empowers the model, but not solely the data used to train it. When certain aspects of the biology are not fully understood and therefore cannot be mechanistically modeled, we complement our mechanistic approach with statistical and phenomenological models. As the unknown science is elucidated and accepted in the published literature, we incorporate it into our models and the phenomenological sections are replaced with hard-coded biology.
Mechanistic models are certainly more difficult to build, and generally larger in scale. However, when executed properly, this type of modeling provides unprecedented power for extrapolation and prediction in domains in which all other techniques fail. Immunetrics takes pride in setting the bar high by validating models against data/scenarios that are absolutely dissimilar to the original set of training data.
A sample output from the Immunetrics model interface.
The Immunetrics approach
Other modeling companies apply math and software to your needs, but at Immunetrics, we also bring biology to the problem. Our approach is to simulate healthy physiology and then layer in the disease physiology. Why is a healthy state foundation so important? Because healthy physiology is well-documented, consistent, and robust, and helps us understand the changes that occur under disease conditions.
Less is more
We maximize our efficiency and productivity by setting sensible limits on the complexity incorporated in our models. This gives us an edge over conventional systems models, often mired in excessive detail and clinically unimportant, unquantifiable outputs. We carefully consider the relevant biological functions and incorporate enough detail to accurately predict clinically relevant outputs, always mindful of simulation time and the dimensionality curse. If additional information does not lend added accuracy to our predictions, then we choose not to add it. We have the expertise and understanding necessary to make informed judgments about where to add detail and where to compromise in the name of efficiency and improved predictability.
Both healthy and disease state physiology involve very complex interactions between the organ systems. Our models represent these relationships in a mechanistic manner.
How do we make decisions about the biology incorporated into our models? At Immunetrics, we believe that the burden of mechanistic modeling is understanding the biology, not writing the equations. Immunetrics models are initially based on published data from peer-reviewed papers, painstakingly researched by our own biology experts. A trained biologist's expertise in reading papers and assessing their scientific rigor cannot be replaced by automated data mining technologies. We work as a team of modelers and biologists to carefully tease out the details of each biological process, and only then does the model's structure take shape.
As an example of our strict standards of mechanistically modeling biological processes, consider our approach to a routine intervention: the administration of fluids. We model the direct mechanistic result of giving intravenous fluids: an increase in blood volume. In addition, the most common secondary results of fluid therapy, improved blood pressure, increased cardiac output, altered drug distribution, and improved tissue perfusion are then modeled, each via its respective mechanism.