Skip to main content
2024 BMES Annual Meeting Main banner
Icon Legend
This session is not in your schedule.
This session is in your schedule. Click again to remove it.
Back

Bioinformatics, Computational and Systems Biology

Bioinformatics, Computational and Systems Biology - Poster Session A

Poster L14 - A foundation model for microbial growth dynamics

Thursday, October 24, 2024
10:00 AM – 11:00 AM EST
Location: Exhibit Hall E, F & G

Introduction: Microbes express diverse phenotypes and behaviors depending on environmental conditions and interactions with other species. We can learn about all of these behaviors by studying the way microbes grow in these different conditions. However, existing growth models fail to capture key details that can be observed from an experiment. Using data from over 100k individual growth curves from clonal populations and microbial communities, we apply machine learning to capture key growth dynamics, which allows us to understand and predict underlying behavior. We have designed a foundation model for microbial growth which can represent diverse microbial growth conditions, including clonal populations grown in different experimental conditions as well as microbiome communities from human guts and environmental samples. We provide a method to capture key growth parameters from a broad array of growth data, and these parameters can help other scientists to study their own systems.

Materials and

Methods: Our project involves data collection, model training, and application of the model. We gather three primary types of growth curves. First, we use plate readers to generate growth curves from primarily clonal populations of E. coli or other common lab bacteria. These data are generated in a large variety of experimental conditions, laboratory equipment, and genetic backgrounds. Second, we use data from robotic systems to control bacterial growth in continuous processes. Finally, we use microbiome time series from human guts and saliva as well as water and soil. Each microbiome dataset has hundreds or thousands of individual species, and they were collected over the course of 100 to 300 days. Finally, we augment the experimental data with simulation data. To generate the simulation data, we use a variety of mechanistic models to represent microbial growth, including variants of the logistic function and the generalized Lotka-Volterra equations. Our foundation model is a variational autoencoder, which is capable of representing the time series with a minimal set of latent parameters. This model was designed and trained using PyTorch. For our applications of the model, we use the encoded growth curves to predict growth parameters as well as predict trajectories of consortia over time.

Results, Conclusions, and Discussions: Our foundation model represents microbial growth dynamics by encoding microbial growth curves into latent parameters and then reconstructing the original curves. The model takes a time series with a predetermined length, such as 128. The time series is processed through multiple layers of convolutions resulting in an array with a predetermined number of latent dimensions, such as 8. By sampling the latent space with a normal distribution we add a variational component to our autoencoder. We use this latent vector to reconstruct the original time series by using multiple layers of transposed convolutions. Our reconstructions have a coefficient of determination of at least 0.97 when using 8 latent dimensions.
We explore two downstream applications. First, we encode time series data to predict growth parameters. Using a regression model, we map the latent vector of a curve to the growth parameters of the experiment or simulation. This trained regression model accurately predicts growth rate and lag time of unseen samples with a coefficient of determination of 0.87. Second, we encode community dynamics of microbial consortia. This approach provides benefits over traditional approaches, because our model represents community dynamics with a set of parameters that scale linearly with the number of members in a consortia, as opposed to traditional models, such as the generalized Lotka-Volterra model, which will scale exponentially.
Across the biomedical engineering world, scientists are applying machine learning and artificial intelligence to diverse applications ranging from image analysis to protein design. Our contribution supplies the first generally applicable microbial growth dynamics foundation model, which has been sourced from a large array of experimental data. Our model can be used in a range of downstream applications, such as determining growth parameters, analyzing experimental conditions, and predicting consortia behavior. These applications are instrumental in aiding engineers to design synthetic biology systems ranging from metabolic engineering of chemicals to engineering microbiomes for use in the gut or soil.

Acknowledgements (Optional):