Caleb N. Ellington

GHC 7603

Gates Hillman Complex, 4902 Forbes Ave

Pittsburgh, PA 15213

Check out my research on contextualized.ml, a statistical machine learning toolbox for estimating powerful biological models at per-patient or per-cell resolution, enabling truly personalized modeling for healthcare and precision diagnostics.

My research interests span meta-learning, multi-task learning, graphical models, and biological modeling. I primarily work on adaptive and sample-efficient machine learning systems that respond to new challenges by using knowledge gathered across many different environments. I’m particularly interested in using these tools to study heterogeneous diseases, like cancer and COVID-19, where disease behavior and pathological mechanisms vary from patient to patient. Meta-learning helps us understand how contexts like genetics and environmental factors create patient-specific disease behaviors and pathologies.

I also enjoy rock climbing, guitar, learning languages, my cats Mimi and Goma, and my rats Boogie and Disco. I’ve lived in Austin TX, Seattle WA, and Pittsburgh PA and I call them all home.

news

Nov 16, 2022	Check out my talk from CSHL Biological Data Science this past week: Contextualized Graphical Models Reveal Sample-Specific Transcriptional Networks for 7000 Tumors
Oct 20, 2022	Our extended abstract Sample-Specific Contextualized Graphical Models Using Clinical and Molecular Data Reveal Transcriptional Network Heterogeneity Across 7000 Tumors has been selected for a talk at CSHL Biological Data Science and was also accepted to ML4H and GLIndA.
Apr 30, 2022	Our study on the variability of the effects COVID-19 treatment effectiveness is published in JBI.
Apr 26, 2022	The first version of contextualized.ml, our open-source machine learning toolbox for personalized modeling, has been released!
Nov 3, 2021	MultiDAG: Multi-task learning of order-consistent causal graphs is in NeurIPS.

selected publications

arXiv

NOTMAD: Estimating Bayesian Networks with Sample-Specific Structures and Parameters

Lengerich, Ben, Ellington, Caleb, Aragam, Bryon and 2 more authors

Abs HTML PDF

Context-specific Bayesian networks (i.e. directed acyclic graphs, DAGs) identify context-dependent relationships between variables, but the non-convexity induced by the acyclicity requirement makes it difficult to share information between context-specific estimators (e.g. with graph generator functions). For this reason, existing methods for inferring context-specific Bayesian networks have favored breaking datasets into subsamples, limiting statistical power and resolution, and preventing the use of multidimensional and latent contexts. To overcome this challenge, we propose NOTEARS-optimized Mixtures of Archetypal DAGs (NOTMAD). NOTMAD models context-specific Bayesian networks as the output of a function which learns to mix archetypal networks according to sample context. The archetypal networks are estimated jointly with the context-specific networks and do not require any prior knowledge. We encode the acyclicity constraint as a smooth regularization loss which is back-propagated to the mixing function; in this way, NOTMAD shares information between context-specific acyclic graphs, enabling the estimation of Bayesian network structures and parameters at even single-sample resolution. We demonstrate the utility of NOTMAD and sample-specific network inference through analysis and experiments, including patient-specific gene expression networks which correspond to morphological variation in cancer.