ElevateBio is now another. They will present at ASGCT as well.
The first set of data will be on how they applied an active learning framework to optimise novel large serine recombinases.
Recombinases can work (great) in bacteria, but disappoint in mammalian cells. So optimising in bacteria can be misleading, which is what ElevateBio is trying to fix.
Instead of generating a large number of variants, testing them (often in bacteria) and picking winners, they test a small set of variants in human cells, train an ML model, predict which new variants are most likely to improve performance and test only those. This is repeated.
If successful, it could reduce development time (dramatically), produce enzymes that are highly active and specific in human cells, and outperform (some of) the current methods used.
They explicitly say this can extend to transposases, R2 retrotransposases and other editors. That's big because it suggests this is not just a recombinase strategy - it is a platform for engineering other gene editing tools.
The second set of data will be on generating entirely new gene-editing enzymes that evolution never produced.
They collected ~370,000 deaminase sequences from a massive dataset (~10 billion proteins) and trained a protein language AI model to learn what makes an adenine deaminase function. This was then used as a model to generate ~2 million new protein sequences. These were filtered computationally (structure, folding, catalytic motifs) before hundreds of candidates were tested in mammalian cells. Ten functional enzymes (with <60% similarity to known proteins) were found.
This is a big deal as they created functional enzymes outside natural evolution. These are structurally valid, catalytically active, but largely novel in sequence space. They also proved the AI model works by generating, filtering, and validating in mammalian cells. The hit rate looks meaningful as well.
It allows them to move from engineering biology to designing biology. However, it is still early-stage. They showed activity, but not necessarily the efficiency needed for the clinic. Specificity and off-targets are unknown. The new enzymes may behave unpredictably. Delivery constraints remain. Scale-up will be challenging.
This approach could eventually produce smaller enzymes, improve specificity profiles, enable hard-to-target loci, as well as reduce IP constraints (tied to natural enzymes). The latter is one that is often overlooked but important commercially.
The third set of data is essentially describing a next-gen evolution of prime-like editing systems, but with a much broader and more industrialised discovery engine
What they are trying to change is some of the core limitations of current RT editors, including weak or suboptimal reverse transcriptases and targeting flexibility.
When it comes to the first, most current systems rely on a small set of engineered RTs (often viral-derived). Instead, billions of natural proteins were mined in order to identify new RT variants. These were then optimised for efficiency, fidelity, and compatibility with human cells.
As for the second, they pair these RTs with LEGs (Life Edit Genes), which are their proprietary CRISPR-like nucleases. So instead of being constrained by one targeting system (e.g., SpCas9), they have a diverse toolbox of targeting proteins and potentially better access to hard genomic sites.
In vitro, it resulted in >60% precise sequence replacement in primary hepatocytes. That's a strong result. Primary human hepatocytes are not trivial cells to edit. So this suggests high editing efficiency and good compatibility with therapeutically relevant cells. The important caveat is that it is in vitro and under optimised conditions. Historically, many editing systems drop significantly going from in vitro to in vivo, and dividing to non-dividing cells drops additionally. So real-world performance is still an open question.
Put together, they are building an end-to-end enzyme discovery plus optimisation engine, not just one modality.
Market Data 
Markets 
AI
