Inference of biophysical models of transcription from expression data

Synopsis

Transcriptional regulation controls when, where, and how strongly genes are expressed. This process is central to cell identity and development, and its disruption can contribute to diseases such as cancer. A major challenge is to understand how transcription factors (TFs) bind regulatory DNA sequences, interact with one another, and regulate RNA polymerase II (Pol-II) to produce transcriptional output. Although modern reporter assays can measure the activity of many regulatory sequences in parallel, interpreting these measurements requires quantitative models that connect sequence-level information to gene expression. This thesis develops biophysically motivated and interpretable models for inferring enhancer regulatory architecture from transcriptional output data. First, a mean-field approximation of an Ising-type model of transcription is derived and applied to self-transcribing active regulatory region sequencing (STARR-seq) measurements from overlapping enhancer fragments. This approach allows functional regulatory sites and their effective roles in transcription to be inferred directly from fragment-level activity. The method is tested on simulated data and then applied to experimental androgen receptor-bound enhancer sequences. Second, the same biophysical framework is extended to infer effective interactions between TFs. After validation on simulated data, the model is applied to fluorescence measurements of the Drosophila even-skipped stripe-2 enhancer, allowing effective interactions between gap-gene transcription factors to be inferred from spatial expression patterns. Finally, a quadratic approximation of the model is applied to hormone-responsive enhancers measured by STARRseq in treated and untreated conditions. Across androgen- and glucocorticoid-responsive datasets, the inferred interaction maps reveal a reproducible butterfly-like pattern centered on the hormone response element. This suggests that hormone-dependent enhancer activity depends not only on the central receptor-bound region, but also on structured interactions with surrounding regulatory sequences. Model predictions are further tested using synthetic interaction maps and STARR-seq mutagenesis experiments. Together, these results show that transcriptional output measurements can be used not only to identify functional regulatory regions, but also to infer effective interaction architectures that shape enhancer activity.