Loading...

Learning Interpretable Representation of Drugs based on Microscopy Images

Sanian, Mohammad Vali | 2023

127 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 56191 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Rohban, Mohammad Hossein
  7. Abstract:
  8. In this study, we aim to learn representations from microscopic cell images that effectively capture the features of drugs affecting the cells, allowing us to identify effective drugs for treating a disease. We employ two parallel learning paths using predictive and generative models. Specifically, we have achieved a predictive model on the RxRx19a dataset that, unlike previous models, is interpretable, optimized, and robust to dras- tic changes in drug properties. Additionally, we have developed the first generative model on this dataset, which not only generates high-quality images but also discovers a meaningful latent space. This latent space divides the representation into relevant features for image generation, filtering out noisy information. We have obtained the first predictive model that calculates the impact score of a drug, ranging from zero to
    one, for each cell. This scoring system allows our network to have an inherently interpretable structure, and by assigning high scores to specific drugs, we can identify thecells that are affected by them. Furthermore, we introduce a confidence score for each drug prediction, representing the network’s confidence, and demonstrate that incorpo- rating this confidence score enhances the Jaccard similarity between the top-ranked drugs predicted by our model and previous models while worsening the ranking of toxic drugs. This indicates that our network has become resilient to drastic cellular changes, and we interpret this increased similarity as a non-noisy and meaningful feature space. In parallel with the predictive models, we have achieved a self-supervised autoen coder generative model that not only produces high-quality images with FID scores of 62.4 and 16.05 on the RxRx19a and BBBC021 datasets, respectively but also possesses a meaningful latent space. Using this latent space, we were able to separate healthy and diseased cells and rank the top-performing drugs on the RxRx19a dataset. This space follows mathematical operations in a way that adding the vector of healthy cells to the vector of an unhealthy cell in the latent space and then decoding it into the image space generates features related to healthy cells in the image space. This latent space is also disentangled, meaning that each part of the latent layer controls a specific part of the image, and we have demonstrated this through attention maps between the feature space and the image space. Essentially, each part of the latent layer controls the generation of a specific part of the image with a specific dimension.
  9. Keywords:
  10. Microscopic Image ; Interpretable Models ; Generative Models ; Drug Resistance ; Cytoxicity ; Drastic Cellular Changes ; Jaccard Similarity Criteria

 Digital Object List

 Bookmark

...see more