PCR vs Optical Duplicates: Key Differences

Marking preserves the full record: the duplicate reads stay in the BAM file with the 0x400 SAM flag set, so any downstream tool can choose to include or exclude them, and the original alignment is never destroyed. Most variant callers (including GATK) simply skip flagged duplicate reads by default, which gets the same practical effect as removal without permanently discarding data that might be useful for troubleshooting, re-analysis with different parameters, or auditing how many reads were actually duplicated. Removing duplicates outright (Picard's REMOVE_DUPLICATES option) is possible but less common, because it makes it impossible to later reconsider the duplicate call without re-running from the original BAM.

No — both categories get the same 0x400 flag in the FLAG field. Picard MarkDuplicates does track the distinction internally for QC reporting purposes (visible in the DuplicationMetrics output, e.g. the READ_PAIR_OPTICAL_DUPLICATES field, and via the optional DT tag under certain tagging-policy settings), but a tool reading only the SAM flag treats all duplicates identically regardless of origin.

It's the maximum physical distance, in pixels on the flow-cell image, that two reads on the same tile can be apart and still be classified as an optical duplicate rather than an independent PCR duplicate. Picard's OPTICAL_DUPLICATE_PIXEL_DISTANCE parameter defaults to 100 pixels. Because different sequencer/flow-cell generations pack clusters at different densities, this value sometimes needs adjusting to match the platform in use — too low and genuine optical duplicates get missed (misclassified as PCR duplicates); too high and independent, unrelated clusters can be wrongly merged into the optical-duplicate count.

The practical effect on a variant caller is similar even though the cause differs: both inflate the apparent read depth supporting one version of the DNA fragment at a locus, which can distort allele-balance estimates and push a caller's confidence in a variant higher (or lower, for the reference allele) than the true independent sampling would support. That's why both categories are marked and excluded by default in standard variant-calling pipelines, not just PCR duplicates.

Yes, in the same direction as PCR duplicates: an unmarked optical duplicate adds an extra read count to a transcript's tally, inflating its apparent expression level in raw counts, TPM, or FPKM. Because optical duplicates are a flow-cell artifact rather than a biological amplification signal, leaving them unmarked doesn't just add noise — it adds it in a way that has nothing to do with actual transcript abundance, which is exactly why they get marked and excluded from counting the same way PCR duplicates are.

Yes. Both Picard and samtools markdup expose the pixel-distance threshold as a configurable parameter precisely because cluster density varies by flow-cell generation. samtools markdup's own documentation suggests roughly 100 for older, non-patterned HiSeq-style flow cells and roughly 2500 for newer, higher-density NovaSeq-style patterned flow cells. Using a threshold tuned for older, sparser flow-cell hardware on a modern patterned flow cell risks under-counting true optical duplicates, since genuinely adjacent clusters can sit farther apart in pixel terms than the default assumes.

UMIs address PCR duplicates specifically, not optical duplicates. A UMI is a short random barcode attached to each original molecule before amplification, so reads that share a UMI (and the same coordinates) can be confidently collapsed as PCR duplicates of one true molecule, even distinguishing true duplicates from reads that coincidentally start and end at the same position. That does not help with optical duplicates, which arise after sequencing begins and have nothing to do with the molecule's barcode — coordinate- and pixel-distance-based duplicate marking is still needed for those regardless of whether a UMI protocol is in use.