How would you estimate library complexity and decide whether deeper sequencing is worthwhile?

Why

Library complexity is the number of distinct original molecules available to sequence. Daley and Smith define the complexity curve as the expected number of distinct molecules seen as a function of reads sequenced, and show it can be extrapolated from a shallow run. Their warning is that early curves can mislead: two libraries can look linear at 5 million reads and then cross, so a naive comparison picks the wrong one.

Picard reports PERCENT_DUPLICATION and an ESTIMATED_LIBRARY_SIZE in unique molecules, based on paired-end duplication. When a library is low in complexity, the preseq authors note that deeper sequencing either gives redundant data that is discarded or introduces bias. So the decision is economic and statistical.

If projections show 20 percent more unique molecules for doubling the reads, resequencing is poor value. If the curve is still rising steeply, more reads pay. Duplicates have more than one origin.

Some come from PCR, some are optical duplicates from the instrument, and in high-coverage targeted or RNA-seq data many identical start sites arise naturally from abundant molecules.

How to work it through

1. Mark duplicates and read the duplication rate and estimated library size.
2. Run a complexity extrapolation on the aligned reads from the current run.
3. Compare the projected distinct molecules at the planned depth with the number needed for the study.
4. Check whether duplication is concentrated in a few regions, which suggests biology or capture bias rather than low complexity.
5. Decide between resequencing the same library and preparing a new one with more input or fewer PCR cycles.
6. Record the projection used to make the decision.

Where the answer stops holding

Extrapolation from very shallow data is uncertain, and estimates assume the next sequencing samples the same library. Natural duplicates in RNA-seq or amplicon data inflate apparent duplication without meaning low complexity. Estimated library size from duplication assumes paired reads and random fragmentation.