Fwd: Mechanick Lecture at UMD

Neurogenomics and 2 topics inspired by this, Quantum Computing & Privacy

Abstract:

My seminar will discuss various data-science issues related to
neurogenomics. First, I will focus on classic disorders of the brain,
which affect nearly a fifth of the world's population. Robust
phenotype-genotype associations have been established for several
psychiatric diseases (e.g., schizophrenia and bipolar disorder). However,
understanding their molecular causes is still a challenge. To address
this, the PsychENCODE consortium generated thousands of transcriptome
(bulk and single-cell) datasets from nearly 2000 individuals. Using these
data, we have developed interpretable machine learning approaches for
deciphering functional genomic elements and linkages in the brain and
psychiatric disorders. Specifically, we developed a deep-learning
model embedding the physical regulatory network to predict phenotype
from genotype. Our model uses a conditional Deep Boltzmann Machine
architecture and introduces lateral connectivity at the visible layer
to embed the biological structure learned from the regulatory network
and QTL linkages. Our model improves disease prediction (6X compared
to additive polygenic risk scores), highlights key genes for
disorders, and imputes missing transcriptome information from genotype
data alone. Next, I will look at the "data exhaust" from this activity
- that is, how one can find other things from the genomic analyses
than what is necessarily intended. I will focus on genomic privacy,
which is a main stumbling block in large-scale
genomic studies. In particular, I will look at how the quantifications
of expression levels can reveal something about the subjects studied
and how one can take steps to sanitize the data and protect patient
anonymity. Lastly, I will look at how the kinds of interpretable
deep-learning models we have developed can be extended to incorporate
Quantum Machine Learning approaches. We have developed Quantum
Boltzmann machines and Quantum Neural Network analogues of the
architectures from the first part of the talk, which we test using
simulations and implementations on quantum hardware. We show hints
that such models may outperform matched classical models on certain
genomics tasks, and show how parameters unique to the quantum versions
of the models can be interpreted in the context of gene regulation
networks.

==
i0umd22