Keeping popANI honest across depth

popANI creeps upward as sequencing depth goes up, far enough to push a genuine same-strain pair over the "same clone" line for no biological reason.

This is one of the profiling parameter guides. It covers two choices you make together: how strict to build the sequencing-error null model (its error rate, plus an optional min_freq post-filter), and which ANI metric to pair it with (popANI, cosANI, or conANI). Get these right and the score stays stable across sequencing depth. The rest of the page shows the effect, measures three ways to remove it, and recommends what ZipStrain ships with.

The problem

popANI calls a position "shared" when the two samples have at least one allele in common. Deeper sequencing turns up more alleles, both real rare variants and leftover sequencing error, so more positions clear that low bar and the score creeps toward 100.

To pin this down, we ran a simple experiment. We took one pair of samples that we know carry the same strain and subsampled their raw reads to four increasing depths (f005 → f100). That gives us a depth axis while the biology stays completely fixed: the same pair, the same genome, just more or fewer reads. So anything that moves in popANI across those depths is a depth artefact, not real signal. On this pair, popANI climbs steadily from 99.9227 at the shallowest depth to 99.9920 at full depth, crossing the 99.99 "same clone" line for no reason other than the reads we added. Everything that follows is about removing that artefact: flattening the depth curve while leaving the real same-versus-different-strain signal alone.

Seeing it: ANI across depth

Each line is one ANI method. A flat line means the method is depth-invariant; a rising line is the inflation. The dashed and dotted curves are the compare-time min_freq references. The y-axis is zoomed into the top fraction of a percent so the small differences are actually visible.

The headline: two knobs, the same cure

You can attack the bias at two stages: a stricter sequencing-error null model (the rows) or a stricter ANI metric (the columns). The grid below shows the residual depth bias for every combination, and darker means flatter.

Residual depth bias, measured as max minus min ANI across depth. Reading down a column shows what the null model buys you at a fixed metric; reading across a row shows what the metric buys you at a fixed null model. Both routes reach the same near-flat floor, and they overlap: they are correcting the same error, so stacking them buys you very little.

Two settings are enough on their own: a null model built at a 1% error rate (span 0.009) and cosANI at a threshold of 0.01 (span 0.002). Either one pulls the bias down from 0.069 into the range of ordinary measurement noise.

The cost: essentially none

A fix that flattens the curve by throwing away data would be a hollow win. That is not what happens here. Every correction leaves the number of compared positions basically untouched (around 1.71M), so the choice between them comes down to where the correction lives, not to lost data.

Each fix on its own: bias removed on the left (lower is better) against positions kept on the right (higher is better). The right-hand bars are all about the same height, so none of the fixes costs meaningful breadth on this pair.

Recommendation

---

Depth-bias parameter study. One same-strain pair (SRR30025063 × SRR30025566, GCF_900475675.1), four read fractions, swept over null-model strictness, ANI metric, and a compare-time frequency filter. The figures are interactive, so hover for exact values.