ML on Phage-Assisted Continuous Evolution

Machine Learning to Infer Fitnesses from Continuous Evolution. (a) PRANCE (phage- and robotic-assisted near-continuous evolution) links phage reproduction to the desired fitness of a given protein and then optimizes fitness through a cycle of infection, reproduction, and mutation. (b) A probabilistic graphical model of a continuous evolution process that represents PRANCE, to be optimized with machine learning. (c) The process used to generate PRANCE simulations for training and validation of the machine learning.

My first PhD project focused on machine learning on the outputs of continuous evolution experiments, like phage-assisted continuous evolution (PACE) and phage- and robotic-assisted near-continuous evolution (PRANCE). I developed a simple latent-variable machine learning model for inferring fitness landscapes and generated large numbers of PRANCE simulations to validate my method. Using careful hyperparameter tuning and extensive validation, I demonstrated that my method was generally highly accurate under a wide variety of simulation conditions.

Unfortunately, very few variants are observed in a typical PACE or PRANCE experiment, so very few fitnesses can be inferred by this method. To augment the quantity of data available, I developed a negative data heuristic that generated upper bounds on fitnesses that were not observed in a given continuous evolution experiment. However, even with this upper bound, I found that the quantity of fitness data generated by a single continuous evolution experiment was generally quite small compared to the amount of sequencing required, so I switched to more promising methods like DHARMA and FLIGHTED.

The results of this work are not and will not be published, but they are available in my thesis (Sundar, 2025). If you would like to use my code, please contact me directly.

References

2025