GTEx Tissue Modeling with Galaxy Image Learner

Contributors

Paulo Cilas Morais Lyra Junior Allissa Dillman Natalie Kucher Jeremy Goecks

Questions

• What is machine learning, and why is it useful for classification?

• How does GTEx provide data for biological machine learning examples?

• How can Galaxy Image Learner train and evaluate image classification models?

Objectives

• Explain the difference between traditional machine learning and deep learning.

• Describe how GTEx gene expression and sample metadata support tissue classification.

• Identify the main Image Learner inputs, settings, and outputs.

Before the Hands-on

• Machine learning can learn patterns from examples.
• GTEx provides gene expression data from many human tissues.
• Image Learner makes image classification workflows available in Galaxy.

Speaker Notes

These slides introduce the ideas needed to understand the tutorial.

What is Machine Learning?

• A way to learn patterns from data instead of writing every rule by hand.
• A model is trained with examples and then used on new examples.
• In supervised learning, each training example has an input and a known label.
• The goal is to make useful predictions on data the model has not seen before.

Prediction and Classification

• Prediction: estimate an outcome from input data.
• Classification: predict a category or label.
• Binary classification: two possible labels.
• Multi-class classification: more than two possible labels.

In this tutorial, the model predicts the tissue label for each GTEx sample.

Traditional Machine Learning

• Often uses human-designed features.
• Works well with smaller or structured datasets.
• Can be easier to interpret than deep learning.
• Examples: logistic regression, random forests, support vector machines.

Tradeoff: feature design can require strong domain expertise.

Deep Learning

• Uses neural networks with many trainable parameters.
• Can learn useful features directly from raw or lightly processed data.
• Strong for images, text, audio, and large complex datasets.
• Usually needs more data, computation, and careful validation.

Traditional ML vs Deep Learning

Both approaches require good data, clear labels, and careful evaluation.

Why Validation Matters

• A model can memorize training data and still fail on new data.
• Validation data helps tune model behavior during training.
• Test data estimates performance after training is complete.
• Large gaps between training and validation performance may indicate overfitting.

What is GTEx?

• GTEx stands for Genotype-Tissue Expression.
• It is a reference project for studying human gene expression across tissues.
• GTEx links molecular measurements with sample and tissue metadata.
• Researchers use GTEx to study tissue-specific expression and regulation.

GTEx Data Products

• Gene expression matrices: expression values for many genes and samples.
• Sample annotations: tissue labels and technical metadata.
• eQTL resources: links between genetic variation and gene expression.
• Histology and metadata resources: biological and quality context.

For this tutorial, the key pieces are gene TPM values and tissue labels.

GTEx in This Tutorial

• Each sample has thousands of gene expression values.
• Each sample also has a detailed tissue label.
• The model learns whether expression patterns can predict tissue type.
• This is a supervised, multi-class classification task.

What is Image Learner?

• A Galaxy tool for training image-based deep learning models.
• It accepts images plus a metadata table containing labels.
• It can train models for classification tasks from the Galaxy interface.
• It produces reports that summarize training and model performance.

Image Learner Inputs

The image_path column points to each image, and the label column tells the model what to predict.

Important Settings

• Task type: binary, multi-class, or other supported task.
• Label column: the target variable the model should predict.
• Image column: the filename for each input image.
• Model architecture: the deep learning model used for training.
• Epochs and early stopping: how long training can continue.

Good settings depend on the dataset, task, and available compute.

Why Use Image Learner?

• Keeps training inside a reproducible Galaxy history.
• Reduces the amount of code needed to train a model.
• Standardizes inputs, model settings, and outputs.
• Makes reports easier to share, inspect, and compare.

This helps researchers focus on the biological question and model interpretation.

Reading Model Results

• Accuracy: how often predictions are correct overall.
• Precision and recall: how errors differ across classes.
• Loss: how uncertain or incorrect model predictions are.
• Training curves: whether learning improves over time.
• Confusion patterns: which classes are mixed up.

Always compare training, validation, and test performance.

Takeaways

• Machine learning uses examples to learn prediction rules.
• Deep learning is powerful for complex data but needs careful validation.
• GTEx provides expression data and tissue labels for supervised learning.
• Image Learner lets Galaxy users train and inspect image classification models.

Thank you!