Summary of a thesis. Maria T. Carilli. "Genetic Interrogation of Expression Regulation." PhD thesis, California Institute of Technology, Pasadena, California, 2026. Defended May 18, 2026. (original). Rights: CC BY-NC-ND 4.0 — no adaptation permitted, not reproduced. This is a short account of it in our own words; the work itself is not reproduced here.

Genetic Interrogation of Expression Regulation

What this covers

A Caltech PhD thesis in computational biology developing stochastic models of transcription and a genetic framework for cis versus trans gene regulation, combined to map regulatory differences across eight genetically diverse founder mouse strains, eight tissues, and tens of thousands of genes.

The question

GWAS link genetic variants to disease and traits, but most significant variants sit in noncoding DNA whose function is opaque. The usual intermediate step, linking variants to gene expression (eQTL studies), is largely restricted to variants near a gene (cis), because testing all distal (trans) candidates against all genes in all cell types is computationally and statistically prohibitive — and even a successful eQTL hit says nothing about which cellular process (transcription, splicing, degradation) a variant disturbs, since it rests on mean expression alone. The thesis targets three linked problems: establishing that trans regulation exists for a gene without an exhaustive variant search, resolving regulation at single-cell and tissue resolution, and moving past mean expression to identify which biophysical process a genetic difference affects.

The approach

Single-cell RNA counts are modeled as outcomes of stochastic (chemical master equation) processes of transcription, splicing and degradation, rather than measurements collapsed to a per-gene mean; fitting such models to the full count distribution recovers kinetic parameters — burst size, splicing rate, degradation/export rate — invisible to mean-based differential expression. Since most such models lack closed-form solutions, the thesis builds neural-network approximators (kernel weight regression) fast enough for genome-wide use, and embeds the same likelihoods in a variational-autoencoder architecture (biVI) to scale to modern datasets. Separately, it develops a coordinate-geometry and regression framework, applied to homozygous parental strains and their F1 hybrids, testing whether an expression difference is cis or trans regulated without identifying the causal variant. An earlier chapter applies a related idea to dimensionality reduction: contrasting the variance structure of target and background data via a Rayleigh quotient ("rhoPCA") recovers signal that comparing means, or standard PCA, misses. The final chapters combine these pieces, applying the cis/trans test to fitted biophysical parameters across eight founder mouse strains.

What it found

Chapter II reformulates contrastive PCA as a generalized eigenvalue (Rayleigh quotient) problem, avoiding the original's non-physical negative variances and tunable contrast parameter. On a mouse kidney dataset it recovered known sex-biased expression patterns and ran orders of magnitude faster than standard contrastive PCA as sample/feature counts grew; an extension handles longitudinal data.

Chapter III shows fitting chemical master equation models (via the authors' Monod package) to scRNA-seq distributions surfaces changes invisible to mean expression. In a radiation-recovery dataset, 380 genes showed a significant change in a kinetic parameter but not in mean mature expression, versus 157 the reverse and 127 both; comparing minimal transcription models across cortical cell types showed consistent differences in best-fitting model, suggesting real biological variety in mechanism.

Chapter IV develops kernel weight regression networks approximating the joint nascent/mature distribution these models predict. The fastest variant was about eight times faster than a moderately fine numerical solver at comparable accuracy, and nearly an order of magnitude more accurate than a coarser, similarly fast solver; a second variant trades accuracy for further speed.

Chapter V embeds the same likelihoods in a scVI-style variational autoencoder (biVI), learning cell-type structure and per-gene kinetic parameters jointly. On mouse cortex data it recovered marker-gene-specific parameter differences and, via Bayes-factor testing, found hundreds of genes per cell subclass differing significantly in burst size or degradation rate with no corresponding mean-expression difference — changes a standard pipeline would miss.

Chapter VI introduces a transformed coordinate system for parent/hybrid expression ratios, with binomial and GLM-based tests, assigning cis/trans regulation from cross data without a tunable threshold. Reapplied to two published datasets, it revised the original calls substantially: one dataset's 57,253 reported trans-regulated genes fell to 39,063, and another's 478 cis-assigned genes rose to 877. Estimating "proportion cis" from a regression slope, rather than the geometrically correct angle, can be off by up to 57% in relative terms.

Chapter VII applies the biophysical models and cis/trans test jointly to eight founder mouse strains across eight tissues (92 cell types, 30,763 genes) — called the largest analysis of single-cell biophysical mechanism the author knows of. Cis regulation dominates differences in burst size, splicing rate and export rate (52.2%, 52.8%, 62% of assignments), with trans next most common; patterns vary by strain, tissue and cell type, and genes such as Ptprd show strain- and cell-type-specific signatures in cortical neurons.

Chapter VIII closes by reviewing the arc from Mendel to genome-scale single-cell biophysics, reflecting that simple, few-parameter models, not large ones, most changed the author's own assumptions about a regulatory process during the PhD.

Limits and context

The thesis cautions against over-reading its own tests: rejecting a "purely cis" or "purely trans" null shows only that the alternative cannot be the sole explanation, not that it is true, and comparing "significant" gene counts across cell types with very different nuclei numbers risks mistaking low statistical power for absence of regulation. Chapter VII calls its splicing model deliberately simple relative to what longer-read sequencing could support, an experimental rather than theoretical limitation. More broadly, it argues against analyzing single-cell data via heuristic summaries (fold-change on means) detached from a model of how the data were generated, framing this as a tension between an "algorithmic modeling" and "stochastic data modeling" culture in genomics, and offers this work as a genome-scale argument for the latter.

Citation

Maria T. Carilli. Genetic Interrogation of Expression Regulation. PhD thesis, California Institute of Technology, Pasadena, California, 2026. Defended May 18, 2026. Available at https://thesis.library.caltech.edu/18729/.