Results#

This section documents experimental results and analysis for compound profiling using different self-supervised learning approaches.

Model Performance Comparison#

The following plot shows the accuracy comparison between different models using cosine similarity as the distance metric:

Accuracy comparison using cosine similarity

Detailed confusion matrices for model evaluation. We show how treatments of each MOA are classified:

Confusion matrices with cosine similarity

We compare similarity stats between same and different MOAs here:

Similarity comparison between models

t-SNE visualization comparing treatment embeddings from different models (Base ResNet, SimCLR variants, and WS-DINO):

t-SNE comparison of model embeddings