Month: September 2026

Ranged Odds Ratios is another phrase to described the concept of dose–response odds ratios that is classically seen with how cancer odds change as smoking exposure increases. Instead of number of cigarettes consumed, we use the amount of each taxa to indicate the risk of a symptom or disease.

With taxa, literature suggests that there is a range for each taxa. Studies using averages are likely a poor choice. More of a taxa is neither better nor worse for most taxa, rather whether it is in a range.

My database of volunteered microbiome samples, Microbiome Prescription, contains two broad categories of data:

• 16s samples: typically from biomesight, ombre, ubiome, etc
• Shotgun samples: typically from CosmosID, PrecisionBiome.eu, Thorne, XenoGene, Tiny Health

I developed the candidate ranged odds ratios for healthy (no symptoms or diagnosis) v unhealthy (has a symptom or diagnosis) by pooling these samples and using percentile ranking as an adjustment mechanism. The underlying assumption was that converting relative-abundance results to percentiles would compensate, at least partially, for differences in laboratory processing, sequencing approaches, and reporting pipelines.

This analysis tests that assumption: do the resulting odds ratios distinguish healthy from unhealthy samples equally well in 16S and shotgun data?

A secondary objective was to identify the taxonomic rank—species, genus, family, order, class, or phylum—at which the odds-ratio model performs best.

Method

A modified 90/10 train–test approach was used. For each test sample, I calculated the log odds ratio at each taxonomic rank. I then summarized results separately for healthy and unhealthy samples using both:

• The arithmetic mean of the log odds ratios.
• The median log odds ratio.
• Restricting to P < 0.01 or Chi square > 6.635

Results are shown in the tables below as:

$$\text{mean} \; / \; \text{median}$$

Using 16s Samples

For 16s samples, the genus level produced the strongest separation between healthy and unhealthy samples. A skew to the lower values is apparent from median < mean.

Tax Rank	Healthy	Unhealthy
Species	-27 / -24.7	-31.7 / -30.6
Genus	-36.8 / -32.9	-44 / -42.5
Family	-19.1 / -17.2	-22.8 / -21.5
Order	-8.3 / – 7.1	-10.6 / – 9.2
Class	-2.6 / -2.4	-3.1 / -2.8
Phylum	-4.7 / -4.7	-5 / -4.6

At the genus level, the healthy versus unhealthy difference was approximately:

• 7.2 using the mean: $-36.8$−36.8 versus $-44.0$−44.0
• 9.6 using the median: $-32.9$−32.9 versus $-42.5$−42.5

This separation was substantially greater than at the other taxonomic ranks. Within this dataset and methodology, genus-level ranged odds ratios appear to be the most informative for 16S results.

Using Shotgun Samples

The shotgun results were unexpected. The model showed little or no ability to distinguish healthy from unhealthy samples, particularly at the species level, where I originally expected performance to be strongest.

Tax Rank	Healthy	Unhealthy
Species	-30.9 / -30.1	-30.9 / -30.1
Genus	-17.8 / -18.2	-16.7 / – 16.9
Family	-6.6 / -6.7	-6.2 / -6.3
Order	-1.9 / – 1.7	-2.1 / – 1.7
Class	0.2 / 0.8	0.2 / 0.8
Phylum	-1.8 / -1.5	-1.9 / -2
All (multiple counting via parent /child)	-97.3 / -84.2	-116.4 / – 115.8

At the species level, the healthy and unhealthy values were identical: $-30.9 / -30.1$−30.9/−30.1 for both groups. Other ranks showed only small, inconsistent differences that would not provide a reliable basis for classification.

Refinement of P / Chi 2

Using 16s and taxa rank of genus, we will explore that the impact of different Chi Square values. Going to higher chi square values does not appear to do better separation of the categories.

Chi Square Threshold	Healthy	Unhealthy
3.84 (P < 0.05)	-36.8 / -32.9	-44 / -42.5
6.35 (P < 0.01)	-36.8 / -32.9	-44 / -42.5
10	-23 / -19.6	-27.3 / -26/1
20	-8.3 / -7.1	-10.2 / -8.9
40	-5.4/ -4.2	-6 / -4.2

Interpretation

One likely explanation is that the candidate ranged odds ratios were derived primarily from 16s data. Even after percentile normalization, the underlying taxa distributions may differ too much between 16s and shotgun pipelines for a shared model to work well.

This result also suggests that the difference between 16s and shotgun data may be more substantial than expected. Percentile ranking may reduce some cross-platform variation, but it does not necessarily make taxa measurements directly comparable across fundamentally different sequencing and bioinformatics workflows.

The poor shotgun performance could reflect one or more of the following:

• The training data were dominated by 16S samples.
• Species-level identification differs materially between 16S and shotgun methods.
• Taxonomic assignments, reference databases, filtering thresholds, and abundance calculations vary across reporting platforms.
• Percentile normalization does not adequately account for platform-specific measurement characteristics.
• There may be insufficient shotgun data to estimate stable ranged odds ratios.

Ranged Odds Ratios for specific symptoms or diagnosis is expected to perform much better. The definition of healthy and unhealthy lacks precision and is also self-declared.

Conclusions

The preferred approach is to develop ranged odds ratios using samples generated through the same processing flow as the samples to which the model will be applied. In practice, that means separate models may be needed for 16s and shotgun data—and potentially for individual laboratories or reporting pipelines.

The difficulty is obtaining enough comparable samples within each processing category to calculate stable odds-ratio ranges. Pooling data was a pragmatic attempt to test the methodology despite this limitation.

For the current dataset:

• Genus-level odds ratios performed best for 16s samples.
• The pooled model did not effectively distinguish healthy from unhealthy shotgun samples.
• Overall discrimination between healthy and unhealthy samples was modest rather than strong.
• Being symptom or diagnosis based is likely to produce better results.

The next step should be to build and validate separate 16s- and shotgun-specific odds-ratio models for explicit symptoms or diagnosis, then compare their performance using the same held-out test methodology.

Using the above 16s data, we can see that it appears to identify healthy individuals reasonably well.

Threshold	Healthy Percentage	Unhealthy Percentage
-32.4	74.4%	53.8%
-21	30.8%	22.3%
-42	87.2%	80.2%
-51	93.6%	93.8%

The key advantage of Ranged Odds Ratio is determining which taxa is the most probable contributor (highest chi square) and the likely contribution to the symptom or diagnosis (odds ratio).