Understanding tumors with diffusion models. Just getting this project started. Stay tuned for updates.
Introduction
Tumors are made up of many cell types with different mutations. These types are known as clones. Understanding these clones and their evolutionary relationships with eachother is essential for predicting therapy resistance, understanding what goes on inside of a tumor, and guiding precision oncology.
Problem
Methods exist [such as PhyloWGS] that do a great job at using biomarkers [mutations] and their respective frequencies [from whole genome sequencing], to reconstruct/deconvolve subclonal populations.
The problem with older methods is that they are dependent on outdated software, and are very slow due to the use of Markov-chain Monte-Carlo engines. This leads to inference times of upwards of 4 hours for a single patient. Moreover, these methods don't work all the time becuase they require very clean, "good" data. Large scale studies and clinical decision making is bottlenecked by these limitations. How can we make these methods better?
Solution
Let's create a diffusion [Denoising Diffusion Probabilistic Model] model to deconvolve subclonal populations. In theory, this can help us create a faster, stable method, allowing for scalable, efficient processing.