Zero-Shot Antibody Design with Generative Methods
Estimated Reading Time: 4 minutes
TL;DR
We discuss an intriguing recent paper that makes progress towards AI-guided antibody design from AbSci, and highlight an interesting recent mathematical result.
Zero-Shot Antibody Design?
AbSci this week announced a new preprint that claims the capability to do zero-shot antibody design. Antibody therapeutics form an increasingly important class of compounds that have found a broad range of uses in the last few decades. Classical methods of antibody discovery typically depend on random searches through a combinatorial space of potential antibody sequences. These approaches have been very effective, but guided antibody design could potentially design novel antibodies that are not easily discoverable by random search techniques.
AbSci’s recent preprint claims to achieve a practical generative design tool. The work uses a generative model to generate 400K potential HCDR3 regions for antibody variants designed to bind to HER2. These compounds were screened in AbSci’s proprietary assay ACE. 421 potential hits were further validated with a surface plasmon resonance assay with 3 very tight (sub-nanomolar) binders discovered. AbSci claims that this is a zero-shot antibody design, since all HER2 binding sequences were removed from the training data for the model.
AbSCi has also validated its approach on a few other targets such as VEGF-A and the Omicron SARS-CoV-2 spike protein.
AbSci has open sourced the discovered sequences (https://github.com/AbsciBio/unlocking-de-novo-antibody-design). AbSCi also provides extensive details about its experimental setup, but reveals very few details about the actual model it uses in its paper. As a method designer, I cannot evaluate the merit of the claimed zero-shot methodology due to the absence of details; it is not clear whether the machine learning approach is necessarily an advance over basic off-the-shelf methods without details. However, the claimed experimental results suggest that there was some methodological innovation behind the scenes.
Antibody machine learning has been booming as a field in the last few years. For this reason, I have been interested in improving DeepChem’s support for antibody handling for some time now. If you are a student or open source contributor who is interested in aiding this effort, please email me at bharath@deepforestsci.com!
Update: Some readers alerted me to this discussion of potential issues with the zero-shot claims in AbSci’s paper. In particular, the thread below argues that a random search could have yielded comparable results (the author is the CEO of a competing company and himself notes the conflict of interest). I am inclined to believe AbSci’s results are still noteworthy, but agree that more details on methodology may be required to assess novelty claims.
Interesting Links from Around the Web
https://www.quantamagazine.org/finally-a-fast-algorithm-for-shortest-paths-on-negative-graphs-20230118/: A very neat result on an old and hard problem in algorithms theory. I remember learning about this problem in my first algorithms course in undergrad. It’s amazing to hear about a practical and simple solution emerging!
Feedback and Comments
Please feel free to email me directly (bharath@deepforestsci.com) with your feedback and comments!
About
Deep Into the Forest is a newsletter by Deep Forest Sciences, Inc. We’re a deep tech R&D company building Chiron, an AI-powered scientific discovery engine. Deep Forest Sciences leads the development of the open source DeepChem ecosystem. Partner with us to apply our foundational AI technologies to hard real-world problems such as drug discovery. Get in touch with us at partnerships@deepforestsci.com!
Credits
Author: Bharath Ramsundar, Ph.D.
Editor: Sandya Subramanian, Ph.D.
This "de novo" antibody design claim is disputed by other experts in the space: https://twitter.com/SurgeBiswas/status/1613232556673224705