Which brain measurements share the same genes?
Brain scans produce thousands of measurements, and many of them are shaped by the same genetic variants. DIMPLE-GWAS sorts those measurements into groups by their shared genetic signal, so researchers can study a few dozen meaningful traits instead of thousands of overlapping ones.
Figures from Bhattacharyya, John et al., A Pleiotropic Map of Brain Imaging Genetics Reveals Biologically Distinct Latent Endophenotypes, medRxiv preprint, 2026 (not yet peer reviewed).
Too many measurements, too much overlap
A GWAS (genome-wide association study) scans hundreds of thousands of spots in the DNA where people differ, and asks: is this spot linked to the trait? Brain imaging hands us a huge number of traits, called IDPs (image-derived phenotypes): the volume of each region, the thickness of the cortex, how strongly two networks fire together at rest, and so on.
Running one GWAS per IDP creates two problems. First, with thousands of tests the bar for "significant" becomes very high, and real but modest signals get lost. Second, many IDPs are close cousins: the left and right halves of a structure, or neighbouring patches of cortex, are often influenced by the same genetic variants. Analysing them one by one repeats the same discovery many times over.
DIMPLE-GWAS asks the data itself: which measurements behave alike genetically? It then groups them so each group can be studied as one, stronger trait.
Reviewing a playlist song by song
Imagine rating 4,000 songs one at a time. You would hear the same sound again and again, and anything subtle would get lost in the pile.
Sorting it into genres first
Group songs that sound alike, then describe each genre. You see the overall shape of the collection, and the shared pattern is clearer than in any single song.
Letting the music define the genres
The genres here are not taken from a brain atlas. They come from the genetic data. That is why the groups only partly match textbook brain regions.
Five steps from DNA signals to trait groups
Each step shows real numbers from the test dataset bundled with the software: 164 resting-state brain-activity measurements, each scored against 86,885 DNA variants.
The genetic map of 164 brain-activity traits
Each dot is one brain measurement. Dots that sit close together have similar genetic fingerprints. Colour and shape show the group the pipeline assigned. Drag to rotate, then hover over or tap a dot to see what it measures and which traits are its closest genetic relatives.
What is in each group?
All traits
| Trait | Type | Group |
|---|
How does it decide how many groups there are?
The grouping step has settings that change the result, such as the smallest group size allowed. DIMPLE-GWAS tries hundreds of combinations and scores each one on three things. Every dot below is one combination it tried on this dataset. Change how much each goal matters and watch which combination wins.
See the blocks the grouping picks up
This grid compares every trait with every other trait: how closely their genetic signals line up across all 86,885 DNA variants. Red means they rise and fall together; blue means they move in opposite directions. Sorted by group, related traits form visible blocks along the diagonal.
Coloured bars along the edges mark each group, in the same colours as the 3D map. Grey marks unassigned traits.
From thousands of measurements to 25 genetic brain traits
The demo above runs on a small test set. The published analysis started from about 4,000 brain IDPs and applied the same pipeline to the 2,315 with significant heritability, covering structure, white matter and brain activity.
Genetically defined groups
15 first-pass groups, with the large white-matter and brain-activity groups split further. Cortical volume and surface area grouped mainly by lobe.
More heritable than the parts
Every group was more heritable than its individual measurements. The median gain was 0.075.
Independent DNA regions
Genome-wide significant loci across groups, 104 of them not reported in earlier GWAS of single IDPs.
Replicated in children
Group-based polygenic scores carried over to the independent ABCD cohort.
The same 164 brain-activity traits, in the paper
For the resting-state fMRI subset, the paper's best setting left 102 of 164 traits unassigned (62%) and placed the remaining 62 in five small groups. The authors flag these groups as exploratory because the structure is weak:
- Networks linking primary and higher-order cortex
- Visual cortex and cross-hemisphere networks
- Cortex–cerebellum networks
- Default mode and other within-network connections
- Motor networks
The 20-run demo above lands close to this (5 groups, 91 unassigned). The 100-run demo's automatic weights favour fewer unassigned traits, so it picks 2 broad groups. Try both in the tuning panel.
What this page does not show
- The test runs use 20 or 100 UMAP runs. The paper used 1,000, so exact assignments here will differ.
- Group labels (A, B…) are arbitrary numbers from the software, not biological names. Naming a group means checking which brain networks it contains.
- The similarity grid is a simple correlation of association Z-scores. It is not a formal genetic correlation (such as LDSC rg).
- Downstream steps from the paper (group-level GWAS, heritability, ABCD replication) are not part of this demo.
Terms in plain words
Run it on the test data
DIMPLE-GWAS is a Python command-line tool. Your input is a table of GWAS Z-scores (rows are DNA variants, columns are traits) plus a short key that names each trait.
git clone https://github.com/mlamcogent/DIMPLE-GWAS
cd DIMPLE-GWAS
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python DIMPLEGWAS.py \
--datainput test_data/test_dat.csv \
--idp_key test_data/IDP_KEY.csvTo use a combination other than the top-scoring one, pass --override_rank N.