\documentclass[11pt]{article}
\usepackage{amsmath, amssymb, hyperref}
\usepackage{darkmode} % click, or press d, to enable
\title{A Life, Partial and Still Compiling}
\author{Joon Hwan (Justin) Hong}
\date{April 23, 2026}
\begin{document} \maketitle
A Life, Partial and Still Compiling
The inconvenient thing about Alzheimer’s disease is that the brain does not give up its explanations in one piece. Some evidence arrives while a person is living: a scan, a cognitive score, a curve traced across repeated visits. Some of it arrives later, in tissue: plaques, tangles, cell types, molecular profiles, and the small biological particulars that rarely fit neatly into the same table. My work begins in that gap.
1 Writing §
Short essays written without the deadline a paper enforces — on the habits of the field, on cities and language, on whatever is interesting enough on a given evening to be worth typesetting.
- On Korean, English, French, and the indignity of being answered in English. Written in all three, with apologies to two of them.
- A civic grievance about Montréal transit — fare zones, missing platform doors, and a half-billion-dollar bus garage.
- On p-values, FDR, and the quiet request that machine-learning results behave like a Western blot.
- A complaint about biological foundation models, and the fanfare that arrives before the baseline.
- A small confession about scientific figures, and the craft the field never quite teaches.
- A satirical defense of the eternal promise that one more lane will cure traffic at last.
2 Research §
Some evidence arrives while a person is living: a scan, a cognitive score, a curve traced across repeated visits. Some of it arrives later, in tissue: plaques, tangles, cell types, molecular profiles. I am a PhD candidate in Quantitative Life Sciences at McGill University, working in the NeuroPM Lab on machine-learning approaches to neurodegeneration and Alzheimer’s disease. Broadly, I study cognitive resilience: why similar burdens of neuropathology can lead to different cognitive outcomes, and how brain networks, cellular organization, and multimodal data may help explain that unevenness.
The methods are computational—PLS-SVD, machine learning, structural and functional connectivity, single-nucleus RNA sequencing, ligand–receptor features, brain atlases, and the usual unglamorous business of making datasets speak to one another without pretending they were born speaking the same language. The aim is biological: to better understand vulnerability, resilience, and the organization of neurodegenerative disease.
3 How I Got Here §
I arrived at neuroscience by accident, which is probably the usual way one arrives at anything sufficiently interesting. The first nudge came in high school, from a biology teacher who made the nervous system seem less like a chapter and more like an inside joke played by nature on itself: the mind, after all, studying the organ that makes studying possible. That curiosity followed me to McGill, where I studied both computer science and biology, with a growing habit of choosing neuroscience whenever the curriculum allowed it.
Since then, the route has been less straight than useful. I worked on depression and suicide research at the McGill Group for Suicide Studies, then on epilepsy and automatic sleep-state labelling with recurrent neural networks, before finding my way to disease trajectory modelling and Alzheimer’s disease in the NeuroPM Lab. The thread through these projects was not simply the brain as an elegant object, though it is that on its better days. It was the question of dysfunction: how biological systems become disordered, how those changes appear in data, and why the same broad disease label can conceal very different trajectories.
My work brings together neuroimaging, structural connectivity, single-nucleus RNA sequencing, cognitive evaluations, longitudinal trajectories, cell-type annotations, ligand–receptor communication, and brain atlases—a collection of measurements that are useful, partial, and not especially inclined to agree with one another. That mismatch is where much of the work lives. In a field increasingly fond of large models and larger claims, I am drawn to methods that earn their complexity and remain biologically interpretable.
4 Selected Projects §
There is a familiar temptation in modern machine learning—biomedical or otherwise—to add more data, more parameters, more attention heads, and hope the subject matter becomes impressed. Most of what follows takes a more skeptical route. And most of it contains a great deal of work that does not appear in the title: brain regions must be reconciled, atlases translated, matrices persuaded into the same order, cell-type labels made comparable, organizational messages made legible to a machine, and features made to disagree in useful rather than accidental ways. The infrastructure underneath the analysis—atlas harmonization, connectivity alignment, feature construction, single-cell preprocessing, machine-learning pipelines, the small JSON-shaped traps between a diagram and a runnable workflow—is not always the most glamorous layer of the work, but it is often where the work is either preserved or quietly lost.
4.1 OrgFlow: workflows from the administrative sediment
An organization rarely explains itself in diagrams. It explains itself in smaller and less dignified ways: an email asking whether something has been approved, a message saying the invoice has gone astray, a call in which three people discover that the “usual process” was usual only to one of them. Reminders, exceptions, confirmations, apologies, attachments, and that most revealing of institutional artifacts: the follow-up. OrgFlow starts there — a prototype (PDF) that asks whether those traces can be recovered into a process model, and from there into a workflow that an automation platform can actually run. My contribution lived mostly in the last stretch: the conversion agent that translated formal process elements into n8n workflow JSON, the structural-validation layer, the code-interpreter environment that let the agent assemble workflows through helper functions rather than hand-writing brittle objects, and the trace checks that compared what the workflow actually did against what the process model said it should do. Less about asking a language model to be clever, and more about giving it tools, guardrails, and a narrow corridor in which cleverness could do less damage. Its failures are part of the point: a workflow that looks plausible on paper becomes, in execution, a very patient machine waiting for a node that will never arrive.
4.2 Cognitive resilience in Alzheimer’s disease
Alzheimer’s disease is often described through its pathological marks—amyloid, tau, atrophy, decline—but the clinical story is less obedient than the list suggests. Some people carry considerable pathology and remain cognitively steadier than expected; others do not. Using multimodal data from cohorts including ROSMAP and TRIAD, I examine how neuropathology, brain activity, functional connectivity, and cognition relate to one another across individuals and across time. The goal is not simply to find a protective region or a convenient biomarker. It is to ask whether resilience has a network-level organization.
4.3 Cell-type-resolved molecular connectomics
A connectome is usually drawn as regions joined by edges. It is a useful picture, though one with a habit of leaving out the fact that every edge is made from biology. This project places a molecular layer beneath the familiar connectome diagram. Using the Allen Brain Cell Atlas, ligand–receptor interaction maps, structural and functional connectivity, and Alzheimer’s disease cohort data, we construct a cell-type-resolved directed molecular connectome—a way of asking how cells in one brain region may communicate with cells in another, and whether those molecular conversations relate to the brain’s larger wiring and synchrony.
5 Correspondence §
Correspondence is welcome on matters relating to this document or adjacent. Responses may be slow but are sincere.
| joon.hong@mail.mcgill.ca | |
| github | github.com/Joon-Hwan-Hong |
| linkedin.com/in/joon-hwan-hong | |
| lab | neuropm-lab.com |