Archived

This content is available here for research, reference, and/or recordkeeping.

Author ORCID Identifier

https://orcid.org/0009-0004-0543-7081

Date Available

7-13-2026

Year of Publication

2026

Document Type

Doctoral Dissertation

Degree Name

Doctor of Philosophy (PhD)

College

Engineering

Department/School/Program

Computer Science

Faculty

Qiang Cheng

Faculty

Simone Silvestri

Abstract

The integration and modeling of high-dimensional, heterogeneous biological data remain central challenges in computational biology due to complex feature dependencies and pervasive missingness. This dissertation addresses these challenges by developing novel generative frameworks for multimodal data reconstruction, imputation, and interaction prediction. In generative modeling, we focus on capturing structural and causal dependencies in sparse biological systems. We first introduce CausalGeD, a causality-aware diffusion framework that leverages Granger-causal attention for biologically coherent spatial gene expression generation. Next, we propose CausalGenDiff, which combines VAE-guided latent representations with causal diffusion to enable robust reconstruction across spatial and single-cell modalities. We further present DepMicroDiff, a dependency-aware diffusion model for microbiome imputation that effectively handles missing data by modeling microbial co-occurrence structures. For multimodal interaction modeling, we address dynamic molecular ``crosstalk'' across biological entities. We introduce CrossLLM-Mamba, which formulates interaction prediction as a state-space alignment problem using bidirectional Mamba encoders. This design enables efficient hidden-state propagation across embeddings from specialized biological language models, achieving strong performance and generalization across RNA-protein, RNA-RNA, and RNA-molecule interactions. Finally, we explore LLM-based reasoning through CrunchLLM, a multitask framework that integrates structured features with textual metadata. It introduces hierarchical input encoding to balance structured and unstructured signals, along with a self-verifiable objective where justification loss constrains prediction consistency. While validated on business intelligence data, this framework provides a general blueprint for interpretable multimodal reasoning in biological domains. Overall, this dissertation demonstrates that causality-aware diffusion models, state-space architectures, and multimodal LLM frameworks offer improved efficiency, accuracy, and interpretability for complex multimodal learning tasks.

Digital Object Identifier (DOI)

https://doi.org/10.13023/etd.2026.350

Archival?

Archival

Download

Share

COinS