Link: silask.github.io/publication/...

Because our subspecies are genome-based, we can identify specific genes or mutations that explain differential associations. This provides a direct path to mechanistic understanding.

We found many cases where only one subspecies associates with disease while sibling-subspecies don't. Sometimes no association is detected at species level because opposing subspecies associations cancel each other out.

Why it matters: Subspecies explain disease associations invisible at the species level. For example, in colorectal cancer, subspecies-based ML models consistently outperformed species-level ones.

We created a sourmash-based method for fast & accurate subspecies quantification, enabling large-scale analysis across entire datasets, e.g., the whole SRA.

We built HuMSub: the most comprehensive catalog of human gut microbiota subspecies (5,361 OSUs across 977 species).
We found that 1/3 of species in the human gut have subspecies, most of which we knew nothing about.

Going down to the strain-level isn't useful for most comparative studies, because everyone has person-specific strains for most of their microbes!
We need an intermediate level: Subspecies.

Most microbiome studies stop at the species level. But strains within a species can differ dramatically in function. This obscures associations and limits reproducibility.

Yes, the church leaders acepted heliocentrism once the evidence convinced them that's better for calculating dates. I don't think any evidence could convince RFK Jr

I find surprising that CLR is not helpful for the associations, rather it is counterproductive.

But then they didn't used the multiplicative replacement for zero-imputing.