Researchers Predict Zinc Fingers Using Transformers
NYU and Toronto Researchers team up to use transformers improve genetic Therapies
Created on January 31|Last edited on February 1
Comment
Researchers from NYU and the University of Toronto have introduced a transformer-based framework for designing Zinc Fingers that can target DNA within the human genome.
What are zinc fingers you may ask?
Zinc fingers are one of the most abundant proteins in the body, and interact DNA, RNA, and other proteins. For this reason, zinc fingers (ZNFs) are involved in DNA repair, cell migration, and many other processes.
Many diseases such as cystic fibrosis, Tay-Sachs disease, and sickle cell Anemia can be linked to errors in the DNA code of the host, and can theoretically be cured through the use of gene editing. Since ZNF’s naturally bind with DNA, they are less likely to trigger an immune reaction compared to other technologies like CRISPR, and are promising for use in gene therapies.
The major hurdle in utilizing ZNF’s has been that the process for designing as ZNF that can achieve a specific task is incredibly difficult due to the complex interactions they have with DNA. Previous efforts to design ZNF’s are hampered by the need for hand selected by a researcher, and oftentimes still fail to achieve desirable results.
The researchers use a unique model structure involving two transformers which learn the interaction patterns between the single ZNF’s, as well as the relationship between the ZNF’s and DNA. This intricate relationship of the individual ZNF’s as well as the ZNF’s relationship with DNA makes it a prime candidate for leveraging modern AI models that are normally used for language modeling.
The researchers created an architecture that consist of two transformers, which learn the intra-relationship of ZNF’s and the relationship between ZNF’s and DNA, respectively. The outputs of these two transformers are then concatenated and fed into a final transformer encoder block, to finally predict the ZNF’s using a feed-forward layer.
The model is already generating ZNF’s designs with clinical potential, and there is great potential for future application.
Reference
Add a comment
Tags: ML News
Iterate on AI agents and models faster. Try Weights & Biases today.