Leveraging generative AI to design novel, functional deaminases for adenine base editing​

Current base editing systems are constrained by what nature has evolved, and experimental throughput makes it impossible to systematically explore the vast protein sequence space. We address this by pairing de novo enzyme design with computational filters and experimental feedback to rapidly discover active deaminases. 

ElevateBio has a versatile editing platform spanning a range of modalities, allowing us to match the right modality to each target. This work on de novo deaminases directly expands the capabilities of our base editing technology. 

We leverage ElevateBio’s extensive protein catalog to curate a diverse set of putative adenine deaminases, applying computational filters, including structure prediction, to arrive at 370,000 high quality training sequences. This carefully curated data set forms the foundation of our generative model. 

Using the training set, we fine-tuned a generative model to produce de novo deaminase sequences and observed the dramatic effect sampling parameters had on generated sequence quality and diversity. More deterministic settings led to high duplication and sequences that were too similar to the training set. While a balanced parameter set produced the most diverse candidates. 

As a quality check, we mapped each generated sequence back to its closest training cluster and found a strong correlation between training cluster size and how many sequences mapped to it. This confirms that our model learned some of the underlying biology of the training distribution, rather than just memorizing sequences. 

Starting from promising de novo candidates, we apply a machine learning pipeline trained on previous experimental based editing data to suggest optimal mutations. Incorporating more machine learning suggested mutations progressively improves editing performance. 

After just one round of machine learning guided engineering. Our top de novo candidate improved from less than 2% to greater than 20% on target A to G editing. Critically, bystander editing at flanking positions, including one only two nucleotides upstream remained extremely low. 

We evaluated the top candidate across a panel of therapeutically relevant gene targets and found that the minimal bystander editing profile is generally favorable across a wide range of unrelated targets. This demonstrates the generalizability and precision of this approach beyond a single locus. 

This work establishes an end to end, AI driven platform for designing novel, active therapeutic proteins that transcend natural sequence space. The result is a de novo deaminase achieving greater than 20% on target editing, with less than 1% bystander effects, positioning ElevateBio to unlock next generation-based editing strategies across a wide range of diseases.