Transfer Learning for Disease Diagnosis and Therapy Development
Principal Investigator
Name
Herschel Rabitz
Degrees
Ph.D.
Institution
Princeton University
Position Title
Professor
Email
About this CDAS Project
Study
PLCO
(Learn more about this study)
Project ID
PLCO-1279
Initial CDAS Request Approval
Aug 8, 2023
Title
Transfer Learning for Disease Diagnosis and Therapy Development
Summary
We propose the use of transfer learning in metabolomics for disease diagnosis and therapy development. We will train a backbone neural network in an unsupervised fashion on a large (100K-sample) dataset of metabolite concentrations. Features extracted from the pre-trained backbone will then be used to train disease diagnosis classifiers. These classifiers are trained on small (100-1000-sample) labeled datasets. We believe that the backbone network, trained on a large number of samples, will aid in resolving the effects of various confounders. As a result the trained disease diagnosis models are expected to identify the key metabolites relevant to the corresponding diseases. These trained diagnosis models will then be studied by perturbative techniques to identify the metabolites most relevant to it diagnosis decisions. This additional outcome may constitute important information in identifying biological pathways relevant to particular disease therapeutic targets.
Aims
- Unsupervised training on metabolite concentration data forming a backbone neural network.
- Supervised training of disease diagnosis models.
- Identification of biological pathways for disease therapeutic targets.
Collaborators
Dr. Roberto Rey-de-Castro