Deciphering the Language of DNA to Identify Regulatory Elements and Classify Transcripts Into Functional Classes (LanguageOfDNA)
- Project Identification
- Project Period
- 6/2020 - 5/2022
- Investor / Pogramme / Project type
- European Union
- MU Faculty or unit
- Central European Institute of Technology
"The Book of Life is written in a four letter alphabet, A, G, C aThe Book of Life is written in a four letter alphabet, A, G, C and T, - with additional marks for DNA structure, methylation, sites conservation etc. In many aspects, understanding of DNA sequences is analogous to understanding natural languages. Machine Learning methods like Recurrent Deep Neural Networks have been successfully applied to both. Examples of use in Genomics include identification of protein binding sites, transcription factors, promoters / enhancers, functional elements like mRNA or lncRNA and even metagenomics classification.
Last two years revolutionized deep learning methods for natural language processing and methods like ELMO (???), BERT (March 2018, Google), ULMFit (January 2018, fast.ai) and LASER (January 2019, Facebook) now provide language model even for cases of limited labeled data size, several meanings of the same word and an attention mechanism focusing on right part of sentence when interpreting given word.
For genomic application, we also often have a limited size of training data (but the whole genome of unlabeled corpus to learned from), the same DNA sequence can have different consequences based on context and we need to know to look for this context. The hope that the newest Deep Learning methods can be useful for genomic data is further strengthen by first experiments, like K. Heyer's Genomic ULMFit, beating several state of the art benchmarks: https://github.com/kheyer/Genomic-ULMFiT However, the number of modern DL methods' application is still very limited.
The aim of this proposal is to change this. Primary goal should include protein binding sites identification. While it is easy to find motifs of the binding sites, the task of prediction whether protein binds to a given DNA location is till not satisfactory solved because the problem cannot be simplified so much. Neural networks previously proved to bring qualitative improvement exactly to areas like this."