Fine-tuning Protein Language Model

Contributors

Anup Kumar

Questions

• How to load large protein AI models?

• How to fine-tune such models on downstream tasks such as post-translational site prediction?

Objectives

• Learn to load and use large protein models from HuggingFace

• Learn to fine-tune them on specific tasks such as predicting dephosphorylation sites

Requirements

• slides Slides: Basics of machine learning
• tutorial Hands-on: Basics of machine learning
• slides Slides: Classification in Machine Learning
• tutorial Hands-on: Classification in Machine Learning

Language Models (LM)

• Powerful LMs “understand” language like humans
• LMs are trained to understand and generate human language
• Popular models: GPT-3, Llama2, Gemini, …
• Trained on vast datasets with billions of parameters
• Self-supervised learning
• Masked language modeling
• Next word/sentence prediction
• Can we train such language models on protein/DNA/RNA sequences?
• LM for life sciences - DNABert, ProtBert, ProtT5, CodonBert, RNA-FM, ESM/2, BioGPT, …
• Many are available on HuggingFace

Bidirectional Encoder Representations from Transformers (BERT)

Language Models (LMs)

Protein Language Model (pLM)

• Models trained on large protein databases
  • Big fantastic database (> 2.4 billions sequences), Uniprot, …
• Popular architectures such as BERT, Albert, T5, T5-XXL …
• Key challenges in training such models
  • Large number of GPUs needed for training: (Prot)TXL needs > 5000 GPUs
  • Expertise needed in large scale AI training
• Most labs and researchers don’t have access such resources
• Solution: fine-tune pre-trained models on downstream tasks such as protein family classification
• Benefits: requires smaller data, less expertise, training time and compute resources

Architecture: Protein Language Model (pLM)

T-SNE embedding projections

Challenges for downstream tasks

• Key tasks
  • Fine tuning
  • Embedding extraction
• Training challenges
  • ProtT5: 1.2 Billion
  • Longer training time
  • Training or fine-tuning cannot fit on GPU with 15 GB memory (~26 GB)

Low-ranking adaption (LoRA)

• Reduce model size
  • 1.2 Billion to 3.5 Million parameters
• Fits on GPUs with < 15 GB memory

Use-case: Dephosphorylation (Post-translational modification (PTM)) site prediction

• PTM: chemical modifications to a protein after systhesis
• Crucial for biological processes such as regulating proteins, gene expression, cell cycle, …
• Dephosphorylation
  • Removal of a phosphate group from a molecule
  • Is less studied and publicly available labeled dataset is small
  • Hard to train a large deep learning model
  • Fine-tuning might improve site classification accuracy

References

• BERT - https://www.sbert.net/examples/unsupervised_learning/MLM/README.html
• LLM sizes - https://microsoft.github.io/Workshop-Interact-with-OpenAI-models/llms/
• Compute sizes - https://ieeexplore.ieee.org/mediastore/IEEE/content/media/34/9893033/9477085/elnag.t2-3095381-large.gif
• ProtTrans - https://ieeexplore.ieee.org/document/9477085
• LoRA - https://magazine.sebastianraschka.com/p/lora-and-dora-from-scratch
• Dephosphorylation - https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8264445/

For additional references, please see tutorial’s References section

• Galaxy Training Materials (training.galaxyproject.org)

Speaker Notes

• If you would like to learn more about Galaxy, there are a large number of tutorials available.
• These tutorials cover a wide range of scientific domains.

Getting Help

• Help Forum (help.galaxyproject.org)

  

• Gitter Chat

  • Main Chat
  • Galaxy Training Chat
  • Many more channels (scientific domains, developers, admins)

Speaker Notes

• If you get stuck, there are ways to get help.
• You can ask your questions on the help forum.
• Or you can chat with the community on Gitter.

Join an event

• Many Galaxy events across the globe
• Event Horizon: galaxyproject.org/events

Speaker Notes

• There are frequent Galaxy events all around the world.
• You can find upcoming events on the Galaxy Event Horizon.

Key Points

• Training a very large deep learning model from scratch on a large dataset requires exertise and compute power
• Large models such as ProtTrans are trained using millions of protein sequences
• They contain significant knowledge about context in protein sequences
• These models can be used in multiple ways for learning on a new dataset such as fine tuning, embedding extraction, ...
• Fine-tuning using LoRA requires much less time and compute power
• Downstream tasks such as protein sequence classification can be performed using these resources

Thank you!