Discovery

AI-directed protein engineering can optimize proteins for genomic medicines when machine learning models are trained on high-quality experimental activity data and paired with rapid design-build-test cycles.

At ElevateBio, we combine automation, protein language models, and active learning to identify protein variants with improved activity for genomic medicine applications.

Gene insertion by large serine recombinases (LSRs) or retrotransposases can make safer, more potent CAR T cell therapies and can be the solution to broadly curing monogenic diseases caused by a variety of DNA changes within the patient population.

Better Starting Proteins Create a Stronger Engineering Foundation

Our computational biologists mined ElevateBio’s large, curated bioinformatic databases to identify a diverse set of novel large serine recombinases and their corresponding attachment sites. Most of these wildtype proteins mediated large gene insertion in human cells, with several outperforming the literature standard Bxb1 molecule. More protein diversity means we can select the optimal LSRs to engineer for each therapeutic challenge.

Engineering in Mammalian Cells Improves Human-Relevant Performance

Engineering in mammalian cells avoids the common trap of finding proteins that work well in bacteria, but not in humans. At ElevateBio, we use experimentally derived activity data from mammalian cells to intelligently discover more potent molecules than those selected by nature, which translate into superior therapies.

Protein Language Models Focus the Search on Higher-Value Design Space

The limited throughput of mammalian cell wet lab experiments makes probing the vast protein design space difficult. Our use of protein language models combined with experimental data enables exploration of the vast and otherwise unmanageable protein fitness activity landscape. By identifying high fitness protein domains near activity maxima, we can rapidly focus our search on protein modifications that improve function.

Automation Accelerates the Design-Build-Test Cycle

With ElevateBio’s automation platform, we can synthesize protein variants in-house, reducing cost and speeding up the design-build-test cycle. After two rounds of machine-learning guided protein optimization completed in just 4 calendar weeks, we found many single mutations that improved large gene insertion, including one that was 2.4-fold better than wildtype protein from nature. By generating pairwise combinations of the most beneficial mutations, we found a new double-mutant variant 3.4-fold more active than wildtype (view and download poster on this topic). These data additionally measured interactions between pairs of individual mutations to facilitate design of higher order mutants which will further improve activity.

The takeaway is clear: protein engineering is most powerful when combining protein language models, machine learning, automation, and mammalian cell-based activity profiling as one integrated system. At ElevateBio, that integrated system helps accelerate protein optimization for genomic medicine development and deliver more effective solutions for our partners.

View the Related Poster and Video Presentation

Targeted large gene insertion is becoming one of the most important frontiers in genomic medicine. Many therapeutic applications benefit from or require inserting a full gene into a defined genomic location.

In many oncology or autoimmune therapies, this technology enables delivery of CARs. It also matters in monogenic disease, where inserting a functional gene may address multiple disease-causing variants with one therapeutic strategy.

The Field Needs Fit-for-Purpose Insertion Systems

Large gene insertion is not one problem. Payload size, delivery, genomic target access, insertion efficiency, specificity, and fit for either in vivo or ex vivo use are important considerations for the correct approach. No single enzyme class can address all potential applications.

Large serine recombinases (LSRs), CRISPR-LSR systems, and retrotransposons each bring different strengths. LSRs can insert large cargo without exposed double-strand breaks and are not limited by payload size. Retrotransposons offer an all-RNA delivery path. CRISPR-LSR approaches can make insertion more programmable, but with added complexity that may be better suited for ex vivo engineering.

The right question is not which modality is best. It is which modality best fits the biology, delivery route, target site, and development path.

Enzyme Diversity Creates More Paths to Success

Large insertion becomes more powerful when developers are not limited to a small set of naturally known enzymes. ElevateBio has built a discovery and engineering pipeline designed to expand that starting set.

Using a catalog of more than 10 billion proteins, ElevateBio mined more than 30,000 LSR candidates, and identified more than 100 active LSRs. Many demonstrated higher activity than the literature benchmark Bxb1 (view presentation and data).

This breadth creates optionality. Different LSRs have different recognition sites, activity levels, and integration profiles. A larger enzyme set gives developers more chances to find an insertion system that fits the therapeutic program instead of forcing the program to fit the tool.

Programmability Is the Next Barrier

LSRs are limited by the sites where they naturally integrate their cargo. Most natural or un-engineered LSRs do not have the desired insertion profiles into human genomes needed for therapeutic applications, which means the LSR technology cannot become broadly useful until engineered. ElevateBio is applying active learning, rational design, directed evolution, and generative AI to improve enzyme performance. The long-term goal is straightforward: choose the desired insertion location, then engineer or design an enzyme system matched to that site.

By leveraging our proprietary CRISPR technology to install landing pads, ElevateBio can make LSR-mediated large insertion programmable at defined genomic sites.  We developed this CRISPR LSR approach in primary T cells and have shown installation of LSR landing pad with greater than 90% efficiency at the TRAC locus. Addition of our LSR components achieved 67% installation of a CD19 CAR under the endogenous TRAC promoter. This approach is an attractive avenue for ex vivo manufacturing applications that require superb programmability.

The Future Is Fit-for-Purpose Insertion

A potent and diverse targeted insertion platform gives developers multiple paths to the same goal: inserting the right cargo, at the right site, in the right cell type, using a system matched to the therapeutic need.

Large gene insertion is not just another editing modality. It is a way to expand what genomic medicines can be designed to do.

View the Related Poster and Video Presentation

Base editing will not reach its full therapeutic potential if the field is limited to enzymes nature has already evolved. To expand what adenine base editors can do, we need to design deaminases for the job, not just search for better natural starting points. 

The Bottleneck Is No Longer Imagination, It Is Experimental Throughput 

Protein sequence space is too large to explore exhaustively. Even ambitious directed evolution campaigns test only a fraction of what is theoretically possible, and many enzymes that show early activity still carry unwanted attributes, including bystander editing profiles that limit targetability. Better enzymes exist. The challenge is finding them quickly enough, and with enough precision, to matter. 

Generative AI Breaks Evolutionary Constraints 

We wanted to explore beyond evolution. We trained a generative model on 370,000 putative adenine deaminases curated from ElevateBio’s 10+ billion protein catalog. Every training sequence was validated computationally, including protein folding with ESMFold, then filtered for quality before inclusion in the training body. The model learned the design grammar of this enzyme family, then generated diverse sequences while maintaining the active site, binding sites, and core folds required for functional deaminases.  

While generating new sequences, selection of sampling parameters is what separates useful diversity from noise. More deterministic settings produced high duplication rates and sequences nearly identical to training data. Balancing creativity with determinism produced more diverse protein sequences ready for experimental validation. 

Generation, Filtering, and Optimization Need to Work as One Loop 

Generating novel sequences is only the first step. We pair sequence generation with strict computational filtering and a machine learning optimization pipeline that uses experimental feedback to improve candidates quickly. Using the “generate, test, optimize” feedback loop, we mitigate much of the inherent risk of de novo generation by treating protein design as an iterative engineering problem.  

Although here we chose to focus on base editors, this integrated platform enables us to design purpose-built editors that can be matched to specific therapeutic targets or disease mechanisms across the full spectrum of gene editing modalities. 

One Round of Optimization Creates a Precision Editor 

Our top de novo deaminase candidate started at less than 2% on-target A→G editing. After a single round of ML-based optimization, it exceeded 20%. The results: 

• Greater than 20% on-target A→G base editing after one optimization cycle (view poster and data) 

• Minimal bystander profile across a panel of therapeutically-relevant genes. Examples include less than 1% bystander editing just two nucleotides from the target base 

The specificity profile is key. Bystander editing at nearby positions is the factor that makes many genomic targets intractable with current base editing tools. A deaminase that edits its target base while leaving the +2 position unchanged opens sites that were previously untouchable.  

The next leap in base editing will not come only from finding better natural enzymes. It will come from designing fit-for-purpose deaminases, testing them quickly, and improving them through experimental feedback. That shift moves enzyme discovery from a search problem to an engineering discipline. 

View the Related Poster and Video Presentation