PMF-GRN and GLM-Prior: New AI frameworks decode gene regulation from DNA sequence

Gene regulatory networks (GRNs) link transcription factors to their target genes, but reconstructing them from genome-wide data is notoriously difficult due to methodological constraints and incomplete reference networks. Existing methods often couple modeling assumptions to specific inference procedures, rely on heuristic model selection, and produce point estimates without uncertainty. Additionally, prior knowledge of TF-gene interactions is often assay-dependent and poorly transferable across species. In her PhD thesis, Claudia Skok Gibbs tackles these challenges with two novel frameworks.

The first, PMF-GRN, casts GRN inference as a probabilistic graphical model optimized via variational inference, enabling principled model selection and uncertainty quantification in every edge estimate. The second, GLM-Prior, addresses the prior bottleneck by fine-tuning the pretrained Nucleotide Transformer to predict TF-target gene interactions directly from nucleotide sequence, and generalizes successfully across yeast, mouse, and human settings. Together, these methods propose a dual-stage view: sequence-derived priors provide a transferable scaffold for less-characterized systems, while probabilistic inference refines regulatory estimates with quantified uncertainty. This work marks a step toward robust, scalable GRN reconstruction from sequence data alone.