Analytical Tools

The Center develops and disseminates statistical, computational, and AI-driven tools so that individual research teams can get full value from complex genetic, genomic, and phenotype data without building the whole stack themselves. Our team works with addiction science researchers to implement and customize analytical methods to fit the scientific question at hand.

Below, we list some areas of expertise of our team and current projects.

Genetic mapping

Genetic mapping is used to discover genetic regions (loci) that may be causing variation in quantitatively defined trait such as drug intake. High-throughput traits such as proteome or transcriptome may also be used as quantitative traits. Mapping populations segregating genetic variation such as heterogeneous stock (HS) rats, recombinant inbred rat strain panels (such as the HRDP) or crosses between rat strains may be used.

  • Genome scans for high-throughput traits can be time consuming. We developed BulkLMM, a fast linear mixed model (LMM) implementation to perform genome scans of a large number of quantitative traits. It is suitable for genome scans for whole transcriptome data (eQTL scanning) or other high-throughput traits and is orders of magnitude faster than standard LMM implementations designed for scanning one trait at a time.

  • Use of strain-level summary data for genetic mapping (as would be the case when using HRDP), we can provide recommendations on what strain summaries should be used and how.

  • Fine mapping. We have performed variant-level work which builds on joint SNP and indel calls across more than 100 inbred strains, generated with DeepVariant and GLNexus from IROC's SRA data, with effect prediction by SNPEff, ANNOVAR, Ensembl, PolyPhen-2, and SIFT.

  • Many traits of interest to addiction researchers are longitudinal or multivariate. For example we may measure nicotine intake over time. Mapping genetic factors of the longitudinal trait is generally more powerful than mapping at single timepoints; it also provides a more holistic view of the trait. We have developed FlxQTL for this purpose.

Multi-omic integration

Modern high-throughput assays (transcriptomics, metabolomics, proteomics, etc) provide an unprecedented molecular window into biological tissues and organisms. The goal of mutli-omic integration is to provide a comprehensive understanding ("systems-level") of biological processes by combining information from disparate molecular views of the same processes or organisms.

Multi-omic integration is an active research area; the methods used depend on the nature of the omics data and the purpose of the study.

  • Data reduction and matrix decompositions are widely used as building blocks for multi-omic integration. We have developed BigRiverEssence, a lean and fast implementation of common matrix reduction techniques such as principal components analysis, penalized matrix decomposition (PMD), sparse cannonical correlation analysis (CCA), joint and individual variance explained (JIVE), and non-negative matrix factorization (NMF).

  • Matrix linear models are a family of bilinear models that can be used for matrix-valued high-throughput biological data. It is useful when both rows and columns of a data matrix have associated covariates. Our MatrixLM package provides a unified interface for specifying models, fast estimation, and inferential tools.

  • We are developing methods for integrative analysis for identifying tissue- and cell-type-specific promoter usage from bulk ATAC-seq, bulk RNA-seq, and scRNA-seq. and will use Multi-Omics Factor Analysis to tie the SMART core's long-read methylation, direct RNA, and snRNA-seq layers together.

  • We have developed a kernel-based approach for using external information to improve metabolite effect estimation. This is potentially also useful for other omics data as well and are exploring such possibilities.

AI-assisted analysis

  • Literature synthesis with a retrieval-augmented generation framework that summarizes existing knowledge for genes of interest and generates testable hypotheses about their roles in SUD.

  • Reproducible analysis pipelines in which an AI agent works through the analysis plan up front, maintains version control of all code and outputs, and validates results across multiple programming languages.