SEPAL:�Spatial Gene Expression Prediction from Local Graphs

1

9/29/23

Paper ID 3

2

9/29/23

• Spatial transcriptomics aligns gene expression with histology images
• These experiments are expensive and you require specialized equipment.

�Can we take the image �and predict gene expression?

3

9/29/23

4

9/29/23

5

9/29/23

1. Gene Selection

2. Relative Delta Prediction

Select genes with the most significant spatial patterns by Moran’s I value

Delta prediction focuses the model in biological nuances

6

9/29/23

We impute missing values with a modified version of an adaptive median filter

Improves training stability and gives 9.1% improvement in average gene correlation

Imputation only used in training. Evaluation done strictly on real data

Train image encoder to predict expression from single patch

Build a graph with features of the k spatial neighbors

Predict a spatial correction integrating graph information

8

9/29/23

SEPAL outperforms both local and global methods in 2 breast cancer datasets.

Predicting deltas already improves state-of-the-art

Introducing spatial context by increasing patch size is worse than using graphs

9

9/29/23

Easiest genes show good correlation but oversmoothed output

Hardest genes are predicted to be almost always constant

• Prediction of deltas instead of absolute values of expression improves performance.
• Using graphs to include spatial context improves over local prediction and outperforms increasing patch scale.
• SEPAL bridges the gap between local and global methods leveraging small datasets while also being able to include spatial context.

10

9/29/23

Thank You�for Your Time!�Questions?

11

9/29/23

{gm.mejia,p.cardenasg,da.ruizl1,a.castillo13,pa.arbelaez}@uniandes.edu.co

Welcome to the poster!