Interpreting ANCOMBC2 results for a continuous variable

[LLansing]

Hello,

In the experiment I'm working with, I wish to see which taxa are correlated with the factor of time. I have 8 timepoints from which my samples are taken, and I do not want to treat them as a categorical variable, since I am interested in taxa that are differentially abundant across time rather than between particular timepoints.

I've ran ancombc2 (and tested with ancombc as well) while treating the time variable as continuous, but I've been unable to find in the ancombc documentation or paper how to interpret the results (log-fold change, q_val) for a continuous variable. Could you please help me understand?

---

Additionally, I ask for your opinion on the following. The lfc values of the significant diff taxa are much lower than I would have expected, compared to the other non-time factors which were included in the formula. You can see in the following volcano plot that there are many taxa with strong q_values but weak lfc.

Plotting the raw reads over time of the significantly different taxa reveals that a number of them are taxa which have non-zero counts at early timepoints but have 0 (or very low) counts at later timepoints. (note: in the image, only one dot appears for later timepoints on the x-axis, but this is several samples overlapping)

I would have expected a stronger lfc for such taxa, given the strong drop off in counts over time. Am I misguided in this assumption?

[FrederickHuangLin]

Hello @LLansing,

Interpreting log fold-change and q-values for continuous variables is quite similar to how it's done for categorical variables. However, for continuous variables, the effect sizes correspond to a "one unit increase" in the continuous variable.

The difference between the log fold-change and the results you observed using the raw data likely arises from the fact that ANCOM-BC2 operates on bias-corrected abundances and, by default, works on complete data (excluding zeros). The analysis of zero counts is evaluated in pseudo-count sensitivity analysis.

If you have any further questions, please feel free to reach out.

Best regards,
Huang

[LLansing]

Hi @FrederickHuangLin,
Thank you for your response. I understand better now. The effect-size per "one-unit-increase" would seem to explain the small values in my volcano plot x-axis, as the effect per unit is quite small.
I've also discovered how to retrieve the the bias-corrected abundances from the ANCOMBC2 results, which I can use for plots to visualize the change in time of differentially abundant taxa while corresponding more directly to the ANCOMBC2 test result.

[LLansing]

@FrederickHuangLin
I have a follow-up about interpretation of ANCOMBC2 with continuous variables.
Am I able to extrapolate the effect-size value to more units of the continuous variable? For example, if I want to find the log fold-change for 10 units, would I simply multiply the per-unit log-fold value by 10?

[FrederickHuangLin]

You are absolutely right, @LLansing :)

[sophia-richter]

Hi @FrederickHuangLin I am still a bit confused about how to interpret log change when using a continuous variable. Does the log fold change mean that the taxa changed that "amount" between the lowest and highest values along the continuous range? So roughly a taxa with positive log fold change increases as the continuous variable increases, and a taxa with negative log fold change decreases as the continuous variable decreases?

I guess I am just a bit confused on how ANCOM runs on continuous variables in general - is it looking only at linear trends of taxon abundance?

Additionally, I noticed that many of my taxa that are differentially abundant do not pass the pseudo-count sensitivity test. How important is this in terms of reporting / considering results significant or not?

Thank you so much!