Foundation models trained on massive datasets are paving the way for incredible improvements in AI-driven modeling using transfer learning. We are building the foundation models we need for environmental microbiology and microbial ecology, and demonstrating their applications in the wild. This approach helps us overcome the high-dimensionality, low-sample-size challenge of the life sciences.

We have an automated laboratory to develop and execute high-throughput experiments to generate large-scale biological data to test our hypotheses and create deep learning-scale datasets to feed into our models. We are developing active learning approaches to optimize dataset generation, model training, and increase the pace of scientific discovery in these ‘self-driving labs’.

In any given environmental sample, the function of 70% of sequences simply can’t be characterized with traditional bioinformatics approaches - meaning we are omitting the significant majority of microbial information from our analysis of marine ecosystems, and ignoring a treasure trove of potential resources for sustainable development and biotech. We are building models and running experiments in the lab to shed light on this ‘microbial dark matter’.

From the vast network of genes, organisms, and compounds that make up the marine microbial ecosystem, how can we monitor processes of interest cheaply and in real time? We are identifying techniques to identify the minimal key biomarkers of the marine carbon cycle; developing molecular probes for real time monitoring; and exploring autonomous technologies for environmental deployment.