Question: How to determine optimal K using ADAMIXTURE?

[rellaovo]

Hi ADAMIXTURE team,

Thank you for developing and sharing this excellent tool! I’m writing to ask for some guidance regarding model selection.

Background

In ADMIXTURE, I usually determine the optimal number of clusters (K) using cross-validation, for example:

admixture pop.bed 16 --cv=10 -j8 --seed=12345

which reports:

CV error (K=16): 0.21043

This makes it straightforward to compare different K values and select the one with the lowest CV error.

My question

In ADAMIXTURE, I did not find a --cv option in the CLI. I understand that ADAMIXTURE reports log-likelihood and other metrics, but I am unsure about the recommended approach for selecting the best K.

Could you please advise:

• What is the recommended way to choose the optimal K in ADAMIXTURE?
• Is there an equivalent to ADMIXTURE’s cross-validation workflow?

If there is documentation, an example workflow, or a recommended strategy that I may have overlooked, I would greatly appreciate being pointed to it.

Thank you very much for your time and help!

Best regards,
Aaron

[xiekunwhy]

Also want to know.

[joansaurina]

Hi @rellaovo and @xiekunwhy,

Thank you very much for your kind words and for the thoughtful question — this is extremely helpful feedback.

We’re happy to share that a new release of ADAMIXTURE will be available later this week, including cross-validation support and several additional improvements.

I’ll post an update here as soon as it’s released.

Best,
Joan

[joansaurina]

Hi @rellaovo and @xiekunwhy,

We’ve just released ADAMIXTURE v1.5, which brings many improvements—including PGEN/VCF support, better initialization for high K, and GPU acceleration.

Cross-validation isn’t included yet, but we’re actively working on it and plan to add it soon. In the meantime, feel free to try v1.5 and explore its enhancements.

Best,
Joan