Multi-modal meta-analysis of functional genomics data to identify regulatory relationships in the brain

Assigning target genes to a transcription factor (TF) is a persistent challenge in gene regulatory biology. No single assay directly observes a regulatory relationship, and the commonly used individual data types, perturbation-response transcriptomics, chromatin immunoprecipitation (ChIP-seq) binding, and gene co-expression, each carry substantial sources of noise and bias when used alone. In this thesis, I present an integrative computational pipeline that aggregates all three modalities into a single, transparent per-gene ranking for a queried TF, converting each modality's evidence to a common Z-score scale via empirical tail probabilities and combining them with Stouffer's weighted Z-score method. I applied the pipeline across 188 TFs and assessed its validity across these TFs. I assessed empirical validity from several independent angles: volume-matched null models, cross-modality concordance testing, comparison against an internally curated database of experimentally supported TF-target relationships, and an agentic PubMed/Europe PMC literature-screening pathway. The modality-swap null models showed that each individual modality carried substantial TF-specific information, and cross-modality concordance across the full candidate pool was significantly positive for all three modality pairs. Curated-database recovery, by contrast, was sparse and strongly bimodal across TFs, with only 132 of 188 TFs having any curated target at all in their candidate pool. I traced this gap back mostly to the reference database's own limited coverage rather than to inaccurate rankings, since the literature-screening pathway recalled curated targets at close to three times the rate it recalled the pipeline's own novel top-ranked candidates. In a single-TF case study of MEF2C, a well-studied neurodevelopmental TF, I found convergent evidence across all three modalities characterizing its top prioritized candidates. Together, these results characterize both the strengths and the specific limitations of an aggregative, hypothesis-generating approach to TF-target prioritization.