In this repository, we will follow a training for large language models (LLMs) for genomics. The training comprises a short lecture and several lab classes.
You can download the lecture note here.
The data can be found in the file:
- data/genome_sequences/hg38/sequences_hg38_200b_verysmall.csv.gz
The file contains 100,000 non-overlapping DNA sequences of 200 bases, corresponding to around 1% of the human genome. For instance, here is one DNA sequence of 200 bases:
We will pretrain an LLM from scratch (a simplified mistral model, see folder data/models/Mixtral-8x7B-v0.1/) on the 100,000 DNA sequences from the human genome. The LLM is pretrained with causal language modeling using 200b DNA sequences from the human genome hg38 assembly.
We will use a pretrained LLM from huggingface (https://huggingface.co/RaphaelMourad/Mistral-DNA-v1-17M-hg38) and finetune it for DNA sequence classification. The aim is to classify a DNA sequence depending on whether it binds a protein or not (transcription factor), or if a histone mark is present, or if a promoter is active.
We will use a pretrained LLM from huggingface (https://huggingface.co/RaphaelMourad/Mistral-DNA-v1-17M-hg38) to predict the impact of mutations with zeroshot learning (directly using the pretrained model for DNA sequences). Here, we compute the embedding of the wild type sequence and compare it to the embedding of the mutated sequence, and then compute a L2 distance between the two embeddings. We expect that the higher the distance, the larger the mutation effect.
We will use a pretrained LLM from huggingface (https://huggingface.co/RaphaelMourad/Mistral-DNA-v1-138M-yeast) to generate artificial yeast DNA sequences.
We will use a finetuned LLM for promoter or transcription factor binding.