De novo antibodies are here.
AI and synthetic biology are unlocking new opportunities to create de novo antibodies, which could potentially reduce the time it takes to get new drug candidates into the clinic by more than half, while also increasing their probability of success in the clinic.
Absci achieved a breakthrough in generative AI drug creation: We are the first to design and validate de novo therapeutic antibodies with zero-shot generative AI.
- What is zero-shot? A method that involves designing antibodies to bind to specific targets without using any training data of antibodies known to bind to those specific targets.
- Why does it matter? Absci’s zero-shot model generates antibody designs unlike those found in existing antibody databases, including de novo versions of all three heavy chain CDRs (HCDR123), the antibody regions most critical to target binding.
- How effective is it? We validated against more than 100,000 antibodies and found our hit rate to be up to five to 30 times greater than biological baselines examined. This data demonstrates the effectiveness of Absci’s generative AI platform at generating de novo antibodies that bind to the target of interest.
Creating antibodies in silico with generative AI represents a major industry breakthrough on the path to fully de novo antibody design and our vision to deliver breakthrough therapeutics at the click of a button, for everyone.
Read the bioRxiv preprint manuscript here
Please note that the preprint manuscript has not undergone peer review, the findings are provisional, and the conclusions may change.
Absci creates the largest expression database of its kind to overcome a longstanding codon optimization problem.
With ultra-high-quality expression data from just three well-studied proteins, Absci’s AI model could generalize accurate predictions relating codon optimization to protein yield.
Codon optimization is the process of finding the perfect DNA sequence to maximize the production of the desired protein therapeutic in a host cell. There are a lot of codon optimization tools out there, but what are they doing?
In this preprint manuscript, we describe how we used the largest expression database of its kind to train our AI models to make accurate predictions relating codon optimization to protein yield. The work represents a robust, accurate AI model to optimize DNA codon sequences to maximize therapeutic protein yield.
The technical details are in the manuscript below. Here are a few takeaways:
- We used our scalable wet lab technologies to generate the largest synonymous mutant expression dataset we know of – a feat on its own.
- We used large language models (LLMs) to learn natural patterns of codon usage, predict expression levels, and ultimately design high-expressing coding sequences (CDSs) on proteins outside our training set.
- We measured functional activity of three different proteins to ensure we produced properly folded, soluble molecules — not misfolded ones.
- Our model outperformed commercially-available algorithms, suggesting it had learned fundamental rules governing codon optimization.
- AI-based codon optimization could theoretically be applied across protein classes to save significant time and money by maximizing protein production at scale.
- In drug creation, this is an exciting tool for increasing expression levels of recombinant proteins, including biologics such as antibodies. Increasing production yields of therapeutic antibodies can increase the availability and accessibility of drugs to patients.
The bottom line: Absci has demonstrated the application of AI to solve another longstanding challenge in the field – creating a robust, accurate AI model to optimize DNA codon sequences to maximize therapeutic protein yield. This could potentially save significant resources in drug creation.
Read the bioRxiv preprint manuscript here
Related news
- Absci recently announced being the first to design and validate de novo therapeutic antibodies with zero-shot generative AI, creating novel antibodies whose in-silico designs were tested and validated in the wet lab — without further lab optimization or affinity maturation. You can read more about that work here.
- Absci also recently showed its ability to simultaneously optimize multiple parameters important to drug developability, including binding affinity and Naturalness score – a measure associated with drug developability and immunogenicity. More details on that work can be found here.
Please note that the preprint manuscript has not undergone peer review, the findings are provisional, and the conclusions may change.
This preprint manuscript, which we will be submitting for peer review, demonstrates how we were able to deploy our AI drug discovery platform to rapidly optimize multiple parameters important to drug development for the antibody trastuzumab. Key highlights from this work include:
- A robust, validated workflow that combines AI models and proprietary wet lab assays to accurately predict an antibody’s binding affinity given only its amino acid sequence
- Investigation of “naturalness”, a metric that scores antibody variants for similarity to natural immunoglobulin repertoires, and characterization of its association with downstream outcomes related to developability and immunogenicity
- Simultaneous multiparameter optimization of binding affinities and naturalness parameters for a given target
Compared to traditional antibody discovery methods, this work shows that our platform can:
- Explore a larger, more diverse area of antibody variants – orders of magnitude greater than experimental methods alone – which may improve the chances of finding more antibodies with better properties
- Potentially save resources and mitigate the risks associated with optimizing one property at the expense of another, by consolidating sequential optimization steps into one multiparameter optimization run
Read the full manuscript here: https://www.biorxiv.org/content/10.1101/2022.08.16.504181v1
Please note that the preprint manuscript has not undergone peer review, the findings are provisional, and the conclusions may change.