immune system, cancer, infectious disease, virus, human disease

Understanding cellular systems

to treat diseases

We build machine learning and AI models with concrete biomedical challenges in mind. Many of these applications are in cell biology which is central for the development of several types of preventions and therapeutics.

Experimental methods to measure molecular profiles (the "state") of cells are currently improving drastically, promising a much deeper understanding of molecular mechanisms that could be targeted with new therapeutics. Today, using these new assays, we can characterize a tissue with billions of data points in a single experiment.

However, much of the complexity measured is difficult to grasp. We deploy machine learning and AI models to break this complexity down to tangible insights, paving the way to biomedical impact.

Learning from big data in biomedicine

using prior knowledge

While parsing tangible patters from these large biomedical datasets with mathematical models is difficult, they can very often be interpreted by biological domain experts with respect to interesting mechanisms. Why do machine learning and AI models struggle in the same scenarios?

A key difference lies in how individual datasets are contextualized in the corpus of biomedical knowledge: while scientists are trained to contextualize experiments, it is not easy to give models this same contextual understanding. We address this challenge by equipping machine learning and AI models with access to the wealth of existing knowledge in biology, using large-scale priors and inductive priors structured in networks or graphs.

Inferring dynamic mechanisms that

underlie snapshots

One key challenge in understanding mechanisms that underlie complex human diseases is to infer dynamic mechanisms from experiments that capture snapshots of cellular systems — glimpses into the molecular state of a sample. Because these snapshots are destructive, we can measure tissues only once and, therefore, lack temporal resolution.

To address this limitation, we design machine learning and AI models to recover some aspects of homeostatic and dynamic mechanisms from snapshot experiments. By using these mechanisms as tangible representations of the aforementioned big datasets, we reduce billions of datapoints down to concrete regulatory patterns that allow us to understand complex disease.

Applying mechanistic insights to

engineer cellular systems

Once we understood these regulatory patterns that underlie how cells function in homeostasis (health) and dysfunction in disease, we can exploit them to intervene: we can design precision therapeutics for complex diseases. 

We are particularly excited about applications of these computational methods in immuno-oncology, infectious disease, autoimmunity and in the engineering of cellular therapeutics. In all of the cases, we can leverage knowledge about disease mechanisms to improve therapeutics that target disease patterns with high precision.

  • We work with single-cell and spatial omics data that provide high-throughput and high-resolution molecular profiles of cells, for example assaying gene expression through RNA content. Consider these reviews on single-cell omics and spatial omics as starting points for further reading. Many of these datasets have been recently generated in the context of the Human Cell Atlas.

  • It has been difficult to study disease mechanisms in many complex human diseases, including several autoimmune conditions, cancers, and infectious diseases, because these diseases affect many parts of the human body. These new genomics experiments enable us to characterize these disease-induced changes with much higher precision than previously possible, thus allowing us to study putative therapeutic targets and prevention strategies. For further reading, you might start with this review.

  • In contrast to primarily predictive machine learning models, we focus on designing and deploying models that are interpretable in terms of tangible biological mechanisms. This interpretability is typically achieved by embedding prior knowledge about mechanism into the model architecture, creating computational hybrids that combine structured knowledge typically used by humans with the flexibility of machine learning models.

Want to learn more?