Min Xu
Rutgers University-New Brunswick, NJ·Mathematics & Statistics
Statistical methods decide what researchers can conclude from data, and two of the field's practical bottlenecks are the subject of this lab's work: how to pin down which genes cause a disease once you know how genes interact, and how to choose the mathematical function a model is trained to minimize. Both projects develop new theory and algorithms, then test them on real data.
- Fields
Mathematics & Statistics
Genetics
Computer Science
- Topics
statistical inference
Markovian network
gene-gene interaction network
- Grants
- 2 known grants, through Aug 2027
What they work on
- Develop a general Markovian network model for gene-gene interaction networks that captures the step-by-step growth of a network over evolutionary time and includes existing network models as special cases.
- Construct confidence sets for the central root nodes of a Markovian gene network using the repro samples framework, so candidate disease genes carry rigorous frequentist significance guarantees.
- Incorporate community and module structure into Markovian network models and devise methods to estimate functional gene modules and hidden gene nodes, then validate findings with bioinformatics criteria and laboratory experiments.
- Learn a data-driven convex loss function for linear regression by approximating the unknown noise distribution with a log-concave density under the Fisher divergence, improving on squared error loss when noise is non-Gaussian.
- Extend data-driven convex loss estimation to multi-task regression with multivariate responses, and deliver algorithms practitioners can apply to their own datasets.
- Analyze the theoretical properties of score matching, the estimation method that underpins the proposed convex loss estimation and many other statistical learning applications.
More about the research
The first project tackles disease gene discovery. Genes rarely act alone; they work in modules and pathways, so medical geneticists typically take candidate genes found in patient data and then study the network of interactions among them. How best to use that gene-gene interaction (GGI) network to pinpoint the genes actually driving disease remains unresolved. The proposal models a GGI network as a Markovian network, a model in which the network grows step by step over time, which suits gene networks because new interactions arise through evolution. Four linked components are planned: a general Markovian network model that includes several existing models as special cases; a method based on the repro samples framework for building confidence sets around the central root nodes of such a network, giving candidate disease genes rigorous frequentist significance guarantees that were previously hard to obtain; an extension that adds community or module structure to the model, with methods to estimate functional modules and infer hidden gene nodes; and a validation plan using standard bioinformatics criteria and laboratory experiments. The team combines statisticians and a geneticist with expertise in network analysis, statistical inference and machine learning.
Why it matters
Identifying which gene mutations cause complex diseases is a central problem in medical genetics; attaching rigorous statistical confidence to candidate disease genes found in interaction networks could make those findings more trustworthy and easier to follow up in the laboratory. Separately, automatically choosing a loss function suited to a dataset could give practitioners across data science more accurate predictions, tighter confidence intervals and greater robustness to outliers than the default squared error choice.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Vahid Tarokh
Duke University, NC·Engineering
Develop the theory of score-based quickest change detection for independent, identically distributed single-stream data under Bayesian, generalized Bayesian and minimax formulations, and establish that performance depends on the Fisher distance between distributions.
Yuxin Chen
University of Pennsylvania, PA·Mathematics & Statistics
Characterize the fundamental limits on sample and computational complexity of reinforcement learning, meaning how much data and computing an agent needs to learn good decisions in an unknown environment.
Weijing Tang
Carnegie Mellon University, PA·Mathematics & Statistics
Develop priority-aware rules for valuing training data based on the Shapley value, a game-theory method for fairly splitting credit, that account for originality, provenance and individual risk within a unified axiomatic framework.
The school
Rutgers University-New Brunswick: 57 labs on PI Pages, $837M a year on research, 713 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.