Applications of Machine Learning to Medicine

Some applications of machine learning to medicine.( slides )
[Summary: A summary of the algorithms below that apply cpmpressed sensing and large deviations techniques to medicine, especially genomic testing algorithms.].

Changes to the interactome in cancer.( slides )
[Summary: We test some of the algorithms below on TCGA and OHSU RNA-seq data to compute the perturbations to DNA translation Markov process in cancer datasets.].

Learning rate under distrbution shift (with an application to flow cytometry). Joint with M.Fahrbach, A.Javanmard, V.Mirrokni. ICML, 2023. ( long version, poster version)
[Summary: We show that optimizing the learning rate for continuously trained neural networks can lead to better accuracy. We provide an application of our algorithm to flow cytometry using single cell RNA-seq data from murine pancreas (Bergen et al, 2020)].

Enhancing selectivity using Wasserstein distance based reweighing. Under submission, 2022. .( preprint )
[Summary: We show that reweighting neural network training data can lead to much better chances of discovering selective drug molecules. In particular, we use our algorithm to predict small molecules which bind to MNK2 but not MNK1. Our predictions have a success rate of 5% (as opposed to less than 1% for non-ML methods) after synthesis and single point of concentration testing.]

Designing optimal tests for slow converging Markov chains. Joint with C.Stein. IMLH (ICML), 2023. ( short version, poster version)
[Summary: We design a hypothesis testing algorithm that can work with very small samples of data - as small as 20 cells. As a prototype example, using publicly available single cell RNA seq data (Bergen et al., 2020), we show that our algorithm can be much more stable than the usual hypothesis testing algorithm.]

Recovering approximate single cell distribution from aggregate measurements, under submission, 2023. ( long version, short version)
[Summary: We propose a hybridization assay based method using a new compressed sensing type algorithm that approximately recovers single cell RNA seq data from a small number of aggregate measurements. Using publicly available single cell RNA seq data (Bergen et al., 2020), we simulate our algorithm to demonstrate the recovery for genes like Cpe, that are involved in hormone secretion.]

Recovering a sparse linear dynamical system, under submission, 2023. ( short version)
[Summary: DNA transcription can be thought of as a Markov process where the underlying dynamical system (the interactome) is sparse. This dynamical system captures which genes directly influence a given gene. Using this assumption we recover the interactome from few samples. The bound on the number of samples needed follows from the parameters in Hormander's theorem - a surprising connection between compressed sensing and differential equations theory. Using publicly available single cell RNA seq data (Bergen et al., 2020), we compute genes that directly influence hormone secreting genes like Cpe and Nnat.]

Approximating a linear dynamical system from non-sequential data, joint with C.Stein, under submission, 2023. ( preprint)
[Summary: We design an algorithm to compute prominently pertrubed genomic pathways using RNA-seq data and apply it to TCGA datasets for breast cancer and AML. Unlike tools like DEseq2, which output differentiating genes, we can isolate differentiating pathways.]