Transformer-based tool recommendation system in Galaxy.

Area of Science:

• Computational Biology
• Bioinformatics
• Scientific Workflow Management

Background:

• Galaxy is a popular web-based platform for scientific analysis.
• Researchers utilize numerous tools and workflows within Galaxy.
• A tool recommender system aids in extending existing analyses by suggesting relevant tools.

Purpose of the Study:

• To develop a novel tool recommender system for Galaxy using transformer neural networks.
• To compare the performance of the transformer model against recurrent neural networks (RNN), convolutional neural networks (CNN), and dense neural networks (DNN).

Main Methods:

• Training a transformer neural network on existing workflows from Galaxy Europe.
• Evaluating the transformer model's convergence speed, model usage time, generalization capabilities, and recommendation quality (precision@k).
• Benchmarking against RNN, CNN, and DNN models for tool recommendation.

Main Results:

• The transformer model demonstrated two times faster convergence compared to RNN.
• Transformer models exhibited significantly lower model usage time (reconstruction and prediction) than RNN and CNN.
• The transformer achieved a higher precision@k metric (approx. 0.98) than DNN (approx. 0.9), indicating superior recommendation quality.
• Transformer outperformed CNN and DNN on key performance indicators, including convergence speed and recommendation accuracy.

Conclusions:

• Transformers offer a novel and effective approach for building robust tool recommendation systems in Galaxy.
• The developed transformer model provides faster training, lower computational overhead, and higher recommendation accuracy, benefiting researchers in scientific workflow creation and exploratory data analysis.
• The enhanced scalability of transformers allows for training on larger datasets, paving the way for more comprehensive and accurate tool suggestions in scientific research.
• Open-source scripts for the recommendation model are available under the MIT license.