Replies to post #1038 on Sangamo Biosciences Inc (SGMO)

[jondoeuk]

Other companies, including Basecamp Research, are using AI to learn the rules for how recombinases target different sequences of DNA. Earlier this month, Basecamp announced its AI model could create new recombinases from scratch https://www.prnewswire.co.uk/news-releases/basecamp-research-launches-world-first-ai-models-for-programmable-gene-insertion-302657979.html https://www.biorxiv.org/content/10.64898/2026.01.12.699009v1.full

Seamless also uses AI, but to keep track of which mutations improve or hinder the activity of its enzymes, and feeds those results into a machine learning program that helps it pick out a library of enzymes for its next round of directed evolution. Since Seamless launched in 2023, it has sped up the process of evolving recombinases to target new sites from two years to under six months.

[jondoeuk]

Seamless ESGCT 2024 (found in a PDF on their site): Highlights

P0670 - Reprogramming Large Serine Recombinases (LSRs) for Site-Specific Integration

Key Points:

Seamless uses a proprietary search motif to identify target sites in almost all coding genes and human safe harbour sites, giving broad applicability beyond sequences similar to natural recombinase targets.

Through directed evolution and rational design, they engineered LSRs to target four selected sites in the human genome.

Engineered LSR clones showed high activity in bacteria, which translated well to human cells.

A leading engineered enzyme, IntSTX-SH2, could precisely integrate DNA into the SH2 locus in human cells.

Optimisation to IntSTX-SH2 2.0 produced a three-fold activity improvement over the first generation, with room for further enhancement.

P0590 - Zinc-Finger Dependent Recombinases (ZFD-Conditional Editing)

Key Points:

Seamless combined Zinc-Finger Binding Domains (ZFDs) with tyrosine recombinases (Y-SSRs) and LSRs.

Engineered variants were inactive without the correct ZFD target site and reactivated only when the site was present, ensuring high specificity.

This strategy increases efficiency and precision of recombinases.

Expands the suite of DNA editing techniques, giving the ability to program edits in a conditional, highly controlled manner.

So Seamless is moving beyond traditional recombinase limits by searching almost all genes for editable sites, programming LSRs to target chosen human sequences, enhancing efficiency via iterations (like IntSTX-SH2 2.0), and increasing specificity with zinc-finger conditional activation.

This positions Seamless with a platform capable of safe, precise, and flexible gene integration, competitive with Sangamo’s MINT, but with greater programmability for new target sites and conditional editing capability (via ZFDs).

All this builds directly on the foundational work described in these papers https://www.nature.com/articles/s41587-023-02121-y https://link.springer.com/article/10.1186/s13059-023-03097-3 https://academic.oup.com/nar/article/51/10/5285/7157522

[jondoeuk]

Stylus Medicine can join the list. Stylus' approach centres on an extensive, proprietary library of therapeutic-grade recombinases, optimised through computational design and machine learning for protein engineering.

There are at least two high-impact foundational studies that are directly relevant to Stylus Medicine's recombinase engineering platform and reflect the kind of scientific work that underpins their proprietary library and AI/ML-driven design strategy.

The first paper describes a new class of recombinases - called bridge recombinases - designed to perform large-scale programmable DNA rearrangements in human cells. These tools go well beyond simple insertions, with the potential to move or rearrange very large DNA segments. The bridge recombinases described were optimised for human cell function and can drive massive programmable DNA rearrangements (e.g., multi-megabase regions) with precision. It expands the landscape of recombinase-based editing beyond classic serine/tyrosine systems - showing that more complex rearrangements are possible, and not limited to small insertions or excisions https://www.science.org/doi/10.1126/science.adz0276

The second paper focuses on engineering large serine recombinases (LSRs) to achieve efficient and highly specific DNA insertion at defined human genomic loci without the need for pre-installed landing pads. It combines directed evolution, machine-learning-guided mutation combinations, dCas9 fusions, and donor DNA optimisation to enhance both efficiency and specificity of recombinase-mediated insertions. They demonstrated up to ~53% insertion efficiency and ~97% genome-wide specificity in human cells with engineered variants - making site-specific insertion of large DNA cargo (up to ~12 kb) practical for therapeutic and research applications https://www.nature.com/articles/s41587-025-02895-3

[jondoeuk]

Brink Therapeutics is yet another. The French company uses directed evolution. To speed things up, Brink also uses a technique called in vitro compartmentalisation, a way of running billions of enzyme reactions in parallel inside microscopic droplets. It's a miniaturised system that accelerates how quickly recombinases can be tested and optimised. By repeating these cycles, Brink can evolve recombinases that target specific genomic sites with high precision. The resulting data, on which enzymes work, and why, is being compiled into a proprietary library that will also feed into AI-driven design tools. This opens the door to eventually designing enzymes computationally before testing them in the lab, speeding up discovery even further.

The combination of in vitro directed evolution, metagenomic discovery and genAI platforms allows them to rapidly screen and capture activity data for billions of synthetic and natural recombinase sequences against diverse DNA target sequences, while continually learning from the results . The immediate goal is to validate five recombinases this year, demonstrating their ability to make precise and safe DNA edits in human cells. The company is also building a library of data that will underpin its AI-assisted enzyme design.

[jondoeuk]

Most recombinase-based genome insertion systems require large double-stranded DNA (dsDNA) donors that are often long. But in mammalian cells large cytosolic dsDNA is a danger signal. Key innate immune sensors include pathways related to the cGAS-STING pathway. Different pathways are activated, including the interferon pathway and cells shut down translation or undergo apoptosis (programmed cell death). This means editing efficiency collapses.

This led to INSTALL (Integration through Nucleus-Synthesized Template Addition of Large Lengths), which avoids dsDNA-triggered innate toxicity and is compatible with recombinases https://www.nature.com/articles/s41586-026-10241-z

However, efficiency is still a huge problem, with less than 1% of mouse liver cells showing successful DNA integration. So INSTALL today is still an order of magnitude or more below where it needs to be. Even if LNP delivery works all the steps involved multiply inefficiency. As for recombinase activity, all the steps have failure probabilities.

While the study uses standard LNPs designed for delivering mRNA, the vehicle can be optimised. Also, if recombinases become at least 10-100x more active, which could be possible with groups using ML/AI models, INSTALL-type systems could suddenly cross therapeutic thresholds.

So INSTALL is a big advance but not a near-term clinical solution.