Active learning-guided optimization of large gene insertion effectors in mammalian cells​

I’m Gavin Ellis and I’ll be sharing how ElevateBio uses active learning, machine learning, and mammalian and cell-based testing to optimize large serine recombinase for targeted gene insertion, helping partners improve protein function for their genomic medicine programs.

So at Elevate Bio, we’re a genomic medicine company and we develop, discover, and manufacture advanced therapies as part of our genome editing toolbox, which has five different modalities. Large gene insertion is one of the modalities that we’re going to talk about today, namely the identification and protein engineering of LSR proteins for large genome insertion. So our computational biology team identified some novel LSR systems by examining our large curated bioinformatic database. And after laboratory validation of the top approaches, we moved into protein engineering to enhance function of our proteins.

So in order to increase the potency of our large serine recombinases, we’ve turned to machine learning in order to probe the protein activity landscape. Rather by training our machine learning model on evolutionary fitness, which compares our protein to other proteins that exist in nature, we decided to use active learning to train our regression model on protein activity, or how well the protein inserts a gene into a genome. So the way we do this is by using a protein language model to generate embeddings of our single amino acid mutants of our protein. And then, in an iterative approach, use automation to quickly evaluate 40 single amino acid variants at a time.

So in round one, we do an unbiased approach where we probe the entire protein activity landscape. So we have 40 different amino acids that accurately represent this mountain that we’re trying to climb in order to get the best potency. So the results of round one seen here show a wide swath of protein activity. Some are completely nonfunctional, while some are already better than wild type. We then take these data and train our regression model to predict the activity of all of the potential variants that we can test.

Then we take our top 40 variants and test them in the same system to see how potent they are at inserting a gene into a genome. And as you can see in round two, some of our variants are up to 24.4-fold better than wildtype. We then took all of the data and folded that into round three, and found that our predictions were then worse, which told us that we’ve reached the limit of our single amino acid variants.

We then decided to see how well these single amino acid mutations work in concert with each other by making double amino acid variants in pairwise screening. So what we did was take the top ten variants in round two and made pairwise mutations to make 45 new double mutants that are seen here. When we evaluated these double mutants in our assay, we saw that with some of them improved activity up to 3.4-fold higher than our any of our single mutants could.

So not only do we now have a new top performing variant that we were able to generate quickly and cost effectively, we also are now holding a epistasis aware data set that allows us to model higher order mutations into, so we can make mutants that have seven, eight, nine, many, many different mutations at the same time. And we know which of these single mutations does not pair well with another mutation.

So this poster illustrates just one example of our active learning pipeline that exists in our parallel protein engineering strategy. And this enables us to do systematic development of the advanced therapies that we have here at ElevateBio including not only the LRS I showed you here, but also our R2 retrotransposases and all of the other proteins we have in our genome editing toolbox.

So in conclusion, we were able to use active learning guided machine learning models to quickly and cost effectively enhance the potency of large gene insertion by our novel LSRs.  We found that three rounds of single mutant screening improved potency greater than two-fold versus wildtype protein, and then pairwise screening of our top single mutants as double mutants further increased potency and captured episodic interactions needed to make higher order mutations. And active learning can be applied to other engineering campaigns here at ElevateBio, where protein structure is unknown, including those with multiple objectives.

This framework gives partners a faster, data driven way to improve gene insertion effectors and advance more potent fit-for-purpose candidates for genomic medicine development.