Absolutely. I’d organize yesterday’s lab meeting into two separate presentations, because they were doing rather different science but using a similar computational philosophy.
1. Kiloni — NutrientSignatureProfiler / autoimmune microbiome
The big question
This project is essentially asking:
Can we take existing microbiome datasets from different autoimmune diseases, harmonize them, identify reproducible disease-associated microbial signatures, and eventually translate those signatures into biologically plausible nutritional interventions?
The diseases visible in the presentation included RA, SLE, MS, and CFS, with additional diseases being screened in the automated pipeline.
Where the data come from
A major part of the project isn't generating new samples—it's finding and integrating existing datasets.
The dashboard showed searches across:
- NCBI GEO — processed/functional genomics datasets.
- NCBI SRA / BioProject — raw sequencing and microbiome studies.
- Literature-derived cohorts — datasets identified through published papers even when the accession isn't immediately obvious.
- GMrepo / BugSigDB — microbiome-oriented resources/signature databases.
- Potentially metabolomic/nutrition information layered onto those microbial findings.
The dashboard you photographed had 18 candidate studies, of which 13 were considered potentially validation-eligible at that stage.
That's important: public data mining itself is part of the research pipeline.
2. What happens after they find a microbiome dataset?
Think of the analysis as:
Raw/public studies → harmonize taxa → construct disease signatures → score patients → validate → interpret biology
They aren't just saying “Bacterium X is increased in RA.”
They create a set of organisms associated with disease versus health, and the spreadsheet you photographed is essentially the locked signature. It contained organisms such as Hungatella, Eggerthella, Allisonella, Coprobacter, etc., with effect/direction information.
Then they ask:
Does the combined signature distinguish disease from healthy subjects?
They tested several ways of calculating that signature score:
GSVA — Gene Set Variation Analysis. Originally designed for gene-expression signatures, but conceptually it asks whether a predefined feature set is enriched in an individual sample.
ssGSEA — single-sample gene-set enrichment analysis. Similar concept: calculate an enrichment/signature score for each individual sample.
PLAGE — another method for summarizing coordinated behavior of a feature set.
Z-score — simpler standardized combination of features.
Random Forest — supervised ML classifier; fundamentally different from the enrichment approaches.
That is why you saw five methods compared.
Their result
For RA, the signature separated disease from health impressively in the displayed analysis. The table showed approximately:
| Method | RA AUC |
|---|---|
| GSVA | 1.00 |
| PLAGE | ~0.72 |
| Z-score | 1.00 |
| ssGSEA | 1.00 |
| Random Forest | ~0.98 |
The smaller SLE/CFS datasets were much less convincing for some methods.
Important caveat: an AUC of 1.00 in a small discovery dataset doesn't mean they have invented a perfect diagnostic test. That's precisely why external validation matters.
3. The coolest part: they're building an automated discovery pipeline
This was the dark dashboard you photographed.
Instead of someone manually Googling:
“RA microbiome dataset”
the system searches databases/literature, identifies candidate studies, extracts accession numbers/sample sizes and prepares them for analysis.
But Kiloni emphasized a human validation checkpoint.
That's important because an automated/LLM-assisted pipeline can confidently produce the wrong accession number, sample count, disease assignment, etc. Their dashboard explicitly flagged discrepancies—for example, when a reported sample count didn't match the repository.
So their philosophy is:
AI finds candidates → human verifies them → verified data enter the analysis.
That's a useful research-design principle for anything you build with LLMs.
4. Translational endpoint: precision nutrition
The final goal isn't simply classification.
Their table proposed things like:
MS → SCFA/Treg-axis depletion → potentially fiber-based SCFA restoration
SLE → altered microbial functional profile → potential targeted micronutrient modulation
RA → pathway-specific microbial/metabolic abnormalities → potentially targeted nutritional intervention
Those are hypotheses, not established treatments.
So the full project is ambitious:
microbiome → microbial signature → metabolic pathway → disease biology → possible intervention.
Presentation 2 — Yaoan: Telomeres + All of Us
This one is much closer to genomics/bioinformatics than microbiome.
5. The central idea
We normally talk about someone's average telomere length.
But humans don't really have “one telomere length.”
We have chromosome ends, and individual chromosome arms can have different telomere lengths.
Traditional approaches such as qPCR generally collapse this into something resembling:
person → average TL
Long-read sequencing allows:
person → chr1p TL, chr1q TL, chr2p TL, chr2q TL …
That's a much richer phenotype.
The published All of Us analysis used 2,573 participants and Telogator2 to derive chromosome-specific telomere length from long-read whole-genome sequencing.
6. The datasets
There were really two key cohorts in Yaoan's analysis.
Karimian dataset
The slide showed a relatively small cohort containing healthy individuals plus 15 IPF/telomere-disease patients.
This gives them something extremely valuable:
actual pulmonary disease biology.
Some chromosome arms had previously been identified as consistently shorter—e.g. 17p, 20q, 12p—while others such as 3p, 4q and 12q tended to be longer.
All of Us
The much larger validation/population dataset.
The published analysis contains 2,573 people with usable long-read chromosome-specific TL measurements.
All of Us v8 includes long-read WGS generated using PacBio and Oxford Nanopore sequencing, alongside short-read WGS and extensive phenotype information. Access to genomic data is through the Controlled Tier Researcher Workbench.
That's a major resource for this lab.
7. Telogator2
This is worth remembering because you'll probably hear it again.
Telogator2 = software that extracts telomeric long reads and assigns them to chromosome arms to estimate chromosome-specific telomere length.
It uses the telomeric/subtelomeric sequence information in long reads to determine which chromosome end the read came from.
The published All of Us pipeline even has an available community workspace and public code.
All of Us chromosome-specific telomere analysis code
8. What did Yaoan actually find?
There were several findings.
A. Telomere length decreases with age
No surprise biologically.
But the interesting question was:
Which part of the telomere-length distribution changes?
They looked beyond the mean.
B. ~5–6 kb appears to be an informative threshold
The percentage of reads below roughly 5.96 kb increased with age.
When they swept thresholds from approximately 1–10 kb, the age association peaked around that region.
So rather than:
mean TL → age
they can derive:
% of telomeres <5.96 kb → age.
That's a distributional biomarker.
C. The whole distribution shifts with age
The density curves showed older age groups shifting toward shorter telomere lengths and becoming narrower.
So aging affects more than one summary statistic.
9. Then comes the clinically interesting IPF result
This was probably the most surprising slide.
They compared different percentiles of telomere length:
10th percentile → little healthy/IPF separation
50th percentile → some separation
90th percentile → much clearer separation
Hence their slide title:
“Signal lives at the top of the distribution, not the bottom.”
Their SVM analysis likewise suggested better IPF discrimination as they moved toward the upper percentiles.
That's interesting because the intuitive hypothesis would be:
telomere disease → shortest telomeres provide the strongest signal.
Their preliminary analysis suggests the overall shape/upper tail may actually carry substantial information.
With only ~15 IPF subjects in that disease cohort, though, this is something I'd regard as a hypothesis requiring validation rather than a settled biological conclusion.
10. Spearman correlation slide
You asked about this previously.
They were comparing chromosome-specific telomere measurements between the Karimian and All of Us datasets.
Spearman correlation asks whether two variables move monotonically together based on their ranks rather than requiring a linear relationship.
Their conclusion:
moderate correlation + similar correlation structure across the two cohorts.
That matters because it suggests the chromosome-arm patterns aren't completely dataset-specific.
The published All of Us work independently found substantial chromosome-arm effects, individual-level variation, and inverse associations with age across chromosome arms.
11. Resources worth remembering
If you're going to keep working with this group, I'd remember these rather than every graph:
| Resource | What you'd use it for |
|---|---|
| GEO | Processed genomic/transcriptomic/microbiome studies |
| SRA | Raw sequencing reads |
| BioProject | Organizes sequencing projects/studies |
| BugSigDB | Published microbial signatures |
| GMrepo | Curated human gut microbiome data |
| All of Us | Huge EHR + survey + genomic research cohort |
| long-read WGS | Large DNA reads; enables structural/chromosome-specific analyses |
| Telogator2 | Chromosome-specific telomere lengths |
| GSVA / ssGSEA / PLAGE | Convert sets of biological features into sample-level signature scores |
| Random Forest/SVM | Supervised ML classification |
And All of Us is particularly powerful because it's not merely sequencing: it is a longitudinal resource combining EHR, genomic and survey information, which creates many opportunities for phenotype-genotype analyses.
12. What I think you should retain from the entire meeting
You don't need to remember every organism or every chromosome arm.
The five concepts worth carrying into your next Johnson meeting are:
1. Public data can itself generate projects.
You don't necessarily need to collect patients.
2. Harmonization is a major scientific problem.
Combining cohorts correctly is often harder than running the final ML model.
3. Don't analyze individual features only—construct biological signatures.
That's what Kiloni is doing with microbial taxa.
4. Don't automatically collapse complex biology to averages.
Yaoan's telomere work shows why distributions, percentiles, thresholds and chromosome-specific measurements can reveal information the mean misses.
5. Discovery → external validation is critical.
Both presentations are ultimately about finding a pattern in one place and asking whether it survives somewhere else.
And that is probably the most useful framework for your own ATS brainstorming:
Pulmonary clinical question → identify public dataset(s) → derive phenotype/signature → validate in an independent cohort → interpret biologically.
For you, I would specifically bring ARDS/critical illness microbiome, IPF/telomere biology, and potentially All of Us pulmonary phenotypes to Dr. Johnson and ask which has both accessible data and somebody in the lab who can work with you. That would turn “I need an ATS project” into three concrete research directions.
Comments
Post a Comment