Advances in Bayesian Latent Factor Interaction Models for Ecological and Environmental Health Applications

Abstract

This dissertation develops novel latent space statistical methodology for imputation in complex interaction networks characterized by sparsity, observational bias, spatiotemporal variation, and functional-valued edges. First, we address the challenge of imputing a meta-network from multi-source ecological data subject to extreme observational bias. This work is motivated by frugivory interaction networks, where most source studies focus on a limited number of taxa, leading to incomplete species co-occurrence information and uninformative non-edges. We introduce the Extended Covariate-Informed Link Prediction (COIL+) framework which flexibly borrows information across studies to reduce bias due to uncertain species occurrence. This allows non-edges arising from species non-overlap to be more effectively distinguished from true absences of interaction.Second, we consider spatiotemporal populations of highly sparse networks, where both node set variability and under-sampling impede inference. In this regime, existing dynamic latent factor models are over-parameterized, resulting in overfitting and poor predictive performance. To address this, we introduce the Nested Exemplar Latent Space (NEX) model, driven by a low-rank factorization of the latent trait tensor. NEX decomposes node-level latent features into a lower-dimensional feature space governed by a small number of exemplar curves. We demonstrate the performance and interpetability of this approach through simulation and application to a collection of Arctic dynamic plant-pollinator networks. Finally, we examine to chemical-gene interaction networks where edges represent dose-response functions measured via high-throughput screening (HTS). These networks are characterized by extreme sparsity, with the vast majority of chemical-gene pairs still untested. We propose the Dose-Activity Response Tracking (DART) method which integrates existing knowledge of chemical structural covariates and gene ontologies into a latent factor model to borrow information across dense and sparse regions of the dose-response surface. DART enables prediction of dose-response curves for untested pairs, facilitating prioritization of substances for further screening. We evaluate the performance of the proposed method with respect to simulated data and a new PFAS-HepG2 dose-response dataset.