Logo image
Back
Mechanistic triangulation across molecular, circuit, and behavioral scales: how our genes, brains, and words relate to thinking and medicine
Dissertation   Open access

Mechanistic triangulation across molecular, circuit, and behavioral scales: how our genes, brains, and words relate to thinking and medicine

Muhammad Elsadany
University of Iowa
Doctor of Philosophy (PhD), University of Iowa
Spring 2026
DOI: 10.25820/etd.008404
pdf
02_Elsadany_final-submission50.97 MBDownloadView
Open Access

Abstract

Modern neuroscience can map a genome, image a brain network, and measure behavior at ever finer scales. What it cannot yet do is trace how a genetic variant alters a specific cell type, how that change reverberates through a circuit, and how that circuit-level shift ultimately shapes the words someone says, or why a child with ADHD cycles through three medications before one works. This gap is not just a matter of missing data; it is a structural problem I call triple attenuation: imprecision at each measurement level compounds, and even large sample sizes cannot rescue signal that has been degraded at multiple stages. This dissertation addresses that problem through an approach I call mechanistic triangulation: generating convergent evidence across molecular, circuit, and behavioral scales to narrow down plausible links between genes, brains, and behavior. Rather than trying to trace a single chain from gene to behavior, I assemble a puzzle where each layer constrains the others. I start where a perturbation meets the tissue. In neurosurgical patients, brief electrical stimulation leaves a cell‑type‑specific molecular footprint. Excitatory neurons upregulate the expected activity‑regulated genes, but microglia mount a larger, unexpected response rich in cytokine and chemokine pathways. That signal is stimulation‑specific, not a byproduct of disease. The observation reframes a basic question: when a drug or a pulse alters brain function, which cell types actually respond? To answer that for chemical perturbations, I bridge molecular signatures to anatomy. I built a deep learning model that learns to predict gene expression across the whole brain from sparse post‑mortem data, then align those spatial maps with drug‑induced signatures from the Connectivity Map. The result is a set of drug–brain maps that suggest where in the brain a compound is most likely to act. When I correlate those maps with meta‑analytic fMRI term maps from thousands of studies, the predicted spatial footprints align with cognitive networks in ways that match each drug’s known clinical profile. Methylphenidate, for instance, aligns with attention and language networks; bupropion, sertraline, and lamotrigine map onto affective and limbic circuitry. The framework passes a face‑validity check. But drugs do not work the same way in everyone. I turn to genetics. Using cell‑type‑specific eQTLs, I impute individual excitatory neuron transcriptomes from genotype data and test whether a person’s transcriptomic alignment with methylphenidate’s signature predicts stimulant response in ADHD. The effect sizes are small, but the pipeline runs at scale in real‑world cohorts and shows exactly where signal degrades: eQTLs derived from adult tissue applied to developing brains, drug signatures from cancer cell lines rather than neurons, and outcome measures that collapse dose, duration, and polypharmacy into binary reports. The machinery works; it needs better inputs. At this point, I have molecular tools that generate spatial predictions and individual‑level hypotheses. To test them, I need behavioral phenotypes that are precise enough to distinguish among the dimensions those predictions target. I turn to language. In ten minutes of speech using the Iowa Speech Sample, I extract over ninety linguistic, acoustic, and temporal features. These features recover much of the structure of a full IQ battery and preserve individual differences. Using canonical correlation analysis, I decompose the joint space into a dominant shared language‑cognition axis and three domain‑specific components. The language‑specific components capture dimensions invisible to traditional scoring, such as whether a person’s first association to “paper” is “cut, copy, paste” or the idiosyncratic “fibrous.” Finally, I bring these behavioral components back to the brain. In a 7T fMRI study, I ask whether they have distinguishable neural correlates. They do. Task‑rest reconfiguration tracks the language‑specific axis in language networks, not in the shared general factor. White matter profiles differ across components. Even the GLM residuals, usually discarded as noise, carry systematic signal about individual differences. The brain distinguishes these behavioral dimensions in ways a single composite score would miss. The contribution is not a single discovery but a demonstration that this kind of multi-scale work can be done. The studies show that principled multi‑scale integration is feasible even with noisy, heterogeneous data, and that holding the layers together generates testable, converging hypotheses that no single scale could provide. The microglia finding reframes which cells might matter for drug response; the behavioral components offer new phenotypes for linking speech to brain circuits; the drug maps provide a scalable way to generate spatial predictions before clinical trials. None of these alone would have raised the same questions. The triple attenuation problem remains unsolved. But by documenting where and why signal degrades, this thesis clarifies what improvements, such as prospective designs, better cell‑type models, and denser phenotyping, will be needed to move toward clinical utility. The next step is a prospective study that puts all three scales in the same individuals. That is the experiment this framework now makes possible.
 Neuroimaging Behavior Cognition Language Medication response Neurosciences

Details

Metrics

1 Record Views
Logo image