Integrating biophysical models with machine learning for biological discovery

Recent advances in omics technologies enable a detailed and high dimensional characterization of biological systems. However, building predictive models by integrating these data with mechanistic information remains challenging. I will talk about two recent developments in our lab to show how machine learning facilitates this integration leading to deeper biological insight.

In the first part, I will explore functional heterogeneity in cell populations. While we can collect high quality single cell data, predicting the extent of functional heterogeneity and to identify its mechanistic underpinnings is difficult. To that end, I will focus on one functional phenotype: cells’ ability to sense their environment. Current estimates of cellular sensing abilities neglect heterogeneity in cell populations and suggest that mammalian cells can barely detect the presence/absence of environmental signals. I will present a new information-theoretic framework to quantify the distribution of cellular sensing abilities using single cell data. I will show that cells in a population have differing sensing abilities and are significantly better at sensing the environment than current estimates. Using live cell imaging data on growth factor pathways, I will show that our predictions are closely reproduced in experimental cell populations. Importantly, our approach can identify biochemical variables that are predictive of cells’ functional readouts.

In the second part, I will focus on host associated microbiomes. Microbiome engineering is an attractive therapeutic avenue. Yet, predictive and mechanistic models that are essential for designing rational probiotic interventions have remained elusive. To that end, I will present a new latent variable model for host-associated microbiomes. Using the bovine rumen microbiome, I will show that the model accurately and simultaneously captures the compositions of and hosts’ phenotypes and the associated microbiome using a reduced dimensional description. The model can predict host phenotypes from the microbiome and vice versa. The model also identifies context-specificity of associations between host phenotypes and microbial organisms, a hallmark of host-associated microbial ecosystems. This approach promises to be a significant step in quantitative modeling of host-associated microbiomes.