Research

Generalizing MPRA activity prediction to data-limited cell types

We benchmarked machine-learning, deep-learning, and fine-tuned models (e.g., AlphaGenome and Borzoi) for massively parallel reporter assay (MPRA) activity prediction. We tested generalization across neuronal contexts to determine how moderate-sized MPRA datasets can support genome-wide prediction of regulatory activity in data-limited cell types.

Regulatory rewiring and network robustness in familial ASD

We compared changes in gene regulation during neuronal differentiation across individuals with differing familial liability for autism spectrum disorder (ASD). Most changes altered the strength of cis-regulatory interactions rather than the network structure. When rewiring occurred, regulatory architecture mattered. Non-redundant architectures were more vulnerable to disruption, whereas redundant architectures were more resilient to regulatory loss.

MPRabc: predicting enhancer–gene regulatory interactions using MPRA data

[Paper]

We developed MPRabc, a machine-learning model for predicting cell type–specific enhancer–gene (E–G) regulatory interactions. We showed that incorporating MPRA data improved the prediction of experimentally validated E–G interactions compared with existing state-of-the-art methods. Using sequence-based models, we extended experimentally measured MPRA activity to genome-wide E–G prediction.

Regulatory network dynamics during early neuronal differentiation

[Paper]

We dissected changes in gene regulation during early neuronal differentiation using E–G regulatory interaction networks. Using multi-omics data, we constructed networks at seven stages and tracked how regulatory interactions changed over development. We identified an ASD-associated variant in a predicted cis-regulatory element governing CHD8 expression, a known risk gene, within a genomic region independently implicated in clinical cases of ASD.

Machine-learning methods for biomarker discovery and disease classification

[Paper]

We designed a machine-learning approach to rank transcriptomic biomarkers and utilize them for patient-level disease prediction. In patients with cardiovascular disease, an 18-biomarker panel was sufficient for high-accuracy disease classification. We also developed IntelliGenes, an open-source tool that integrates statistical and machine-learning methods for biomarker discovery and disease classification.