Keywords: ai, elixir, ml

Learning objectives:

• Compare the embeddings of wild-type and mutated sequences to quantify the impact of mutations using L2 distance as a metric.
• Concept of RNNs
• Concept of attention
• Concept of backprop and epochs
• Concept of filters
• Concept of pooling layers
• Configure and implement data collation to organize tokenized data into batches for efficient training.
• Define and configure hyperparameters for pretraining a model, such as learning rate and batch size.
• Define and configure training parameters to optimize the model's performance on the classification task.
• Describe the process of generating synthetic DNA sequences using pre-trained language models and explain the significance of temperature settings in controlling sequence variability.
• Develop a complete workflow for training a language model on DNA sequences, from data preparation to model evaluation, and apply it to real-world bioinformatics tasks.
• Develop a comprehensive plan for documenting and sharing AI model configurations, datasets, and evaluation results to enhance transparency and reproducibility in their research.
• Develop a pipeline to detect open reading frames (ORFs) within generated DNA sequences and translate them into amino acid sequences, demonstrating the potential for creating novel synthetic genes.
• Develop a script to automate the process of predicting mutation impacts using zero-shot learning, enabling researchers to apply this method to their own datasets efficiently.
• Evaluate the fine-tuned model's accuracy and robustness in distinguishing between different classes of DNA sequences.
• Explain the concept of zero-shot learning and its application in predicting the impact of DNA mutations using pre-trained large language models (LLMs).
• Explain the importance of data provenance and dataset splits in ensuring the integrity and reproducibility of AI research.
• Explain the role of a tokenizer in converting DNA sequences into numerical tokens for model processing.
• Forward step
• How model parameters are learned
• Identify and load a pre-trained language model (LLM) suitable for DNA sequence analysis.
• Implementation of RNN (code)
• Implementation of attention mechanism (code)
• Implementation of fine-tuning (code)
• Initialising a model with conv layers (code)
• Initializing model with a single layer (code)
• Initializing model with multiple layers (code)
• Input data representation
• Interpret the results of the L2 distance calculations to determine the significance of mutation effects and discuss potential implications in genomics research.
• Learn the fundamentals of programming in Python
• Load a pre-trained model and modify its architecture to include a classification layer.
• Loss function
• Model as equation
• Monitor and evaluate the model's performance during training to ensure effective learning.
• Perform BLAST searches to assess the novelty of generated DNA sequences and interpret the results to determine the biological relevance and uniqueness of the synthetic sequences.
• Predictions and save+load models
• Prepare and preprocess labeled DNA sequences for fine-tuning.
• Prepare and tokenize DNA sequence datasets for model training and evaluation.
• Set up a computational environment (e.g., Google Colab) and configure a pre-trained language model to generate synthetic DNA sequences, ensuring all necessary libraries are installed and configured.
• Training (code)
• Training steps (code)
• Use k-mer counts and Principal Component Analysis (PCA) to compare generated synthetic DNA sequences with real genomic sequences, identifying similarities and differences.
• Use the trained model to generate embeddings for DNA sequences and interpret these embeddings for downstream bioinformatics applications.
• Utilize a pre-trained DNA LLM from Hugging Face to compute embeddings for wild-type and mutated DNA sequences.
• cross validation and a test set
• finetuning LLM
• general sklearn syntax intro
• metrics and imbalance
• overfit/underfit
• pretraining LLM for DNA
• the need for regularization
• to do
• zeroshot prediction for DNA variants and synthetic DNA sequence generation.

Event types:

• Workshops and courses