// Jul 2026
The Next Base Editors Will Be Engineered, Not Found
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.
Matt Nethery, Principal Data Scientist, Computational Biology
Matt Nethery works as a Principal Data Scientist in ElevateBio's gene editing services, responsible for the implementation of generative AI for discovery and engineering of novel genome editing tools. He has 15 years of wide-ranging industry experience, from building software applications and mobile analytics in the healthcare industry, to bioinformatics and generative AI. Matt completed his PhD in Functional Genomics at North Carolina State University researching microbial genomics, CRISPR biology, and phage engineering.
