From reads to a result a biologist can question

A PDF that lists twelve genes is a slide. The result a biologist can question is the chain from a read to the filter that kept those twelve and dropped the rest. Each step throws information away on purpose. The comparison this page makes is between the file you still hold at each step and the question that file can still answer. Production of the reads themselves is next-generation sequencing from library to reads. The nouns for the files are collected in a glossary of sequencing terms.

QC asks whether the run failed before the biology did

Raw FASTQ in hand, the first question is dull and necessary. Do the base qualities collapse halfway along the read? Are adapter sequences a large fraction of the bases? Does one sample have far fewer reads than its neighbours? A quality report of the FastQC class summarises those patterns. It does not decide biology. It decides whether you are allowed to interpret biology.

A biologist should ask which sample failed QC, and what the fail rule was. A sample that was carried forward "so the table would be complete" will dominate a later argument. If the failure is a quantification or a dimer problem, the bench-side reasoning is in troubleshooting a failed library quantification. If the failure is a control, stop, as described for lane controls elsewhere in this cluster. Do not trim your way out of a lane that never worked.

Trimming asks what sequence was chemistry

Adapters and poor-quality tails are not genome. Trimming removes them. The question a biologist can ask is how much was removed, and whether inserts were so short that the two reads of a pair now overlap in adapter rather than in genome. Aggressive trimming can make reads too short to align, which then looks like missing coverage. Mild trimming can leave adapter that aligns nowhere or, worse, aligns somewhere by chance.

Compare the trimmed FASTQ with the raw one by counts, not by a vibe. If a gene you care about is built from reads that were mostly adapter, the gene is not a result. Record the trimmer and the version. A later analyst who trims differently will not reproduce your table, and both of you can be internally consistent.

Alignment asks where the reads were allowed to land

Alignment places reads on a reference you must name. The biologist's question is: which build, and what did you do with reads that fit in more than one place? A coordinate means nothing until the build matches the browser session, on Ensembl or the UCSC Genome Browser. A mean depth of "thirty" is not a description of a locus that the aligner refused because it was a repeat.

This is also where GC bias masquerades as absence. A hole can be chemistry. The comparison between a GC slope and a local deletion is in GC bias and coverage holes. Ask for that plot before you accept a missing exon as a result. Ask for MAPQ behaviour in repeats before you accept a pileup there as depth.

The BAM or CRAM is the file that still lets you look. A coverage bigWig is a picture of the alignment after someone chose window sizes. Keep the picture if it helps, and keep the alignment if you intend to argue.

Duplicates ask whether depth is molecules

PCR copies and optical duplicates pile onto the same start site. Marking them, and telling the caller to ignore them, is how depth becomes closer to a count of independent fragments. The biologist should ask for the duplicate rate per sample. A very high rate means the library had too few molecules, and the remaining unique depth may be too thin for the claim. A caller that was run without the duplicate flag will report confident variants from copies.

Deduplication is not free of error. In some amplicon designs every molecule starts at the same primer, so a start-site duplicate rule will treat true independent molecules as copies. The question to ask is whether the duplicate definition matches the library. If it does not, the step is theatre.

Calling, annotation, and filtering are where the claim is made

A caller proposes differences from the reference. Annotation attaches a gene, a consequence, a transcript. Filtering throws away calls that fail a rule: low depth, low allele balance, a common population variant, a strand bias. The biologist's questions are the ones that change a paper. Which transcript was used for the consequence? Was the population filter appropriate for this species and this cohort? Did the allele balance rule assume a diploid germline and then get applied to a tumour, a mosaic, or a pooled culture?

Compare a raw call set with a filtered one on a locus you understand. If a variant you can see in the alignments disappears, the filter is the result, and it should be written down. A PASS flag means "passed this pipeline". It does not mean "true in the cell".

Sanger still has a place as a comparison method for a small number of those calls. A clean trace under a heterozygous call, or a mixed trace you were told was homozygous, is information. Discussing that confirmation as a method belongs with the Sanger DNA sequencing enquiry reference. It is a second assay, not a decoration on the PDF.

A gene list cannot survive contact with a new question

The comparison that matters is between objects. A FASTQ can answer a new trimming question. A BAM can answer "show me the reads". A filtered VCF can answer "what did this pipeline emit?" A PDF gene list can answer "what did the slide say?" When a collaborator returns with a new question, only some of those objects survive. Data retention and who holds the raw files is the practical consequence. If the only archived object is the list, the new question is already lost.

Where the chain usually breaks

The report starts at annotation. Nobody can say which build was used, so the gene names do not match the ones in your notebook. Duplicates were ignored and a rare allele became common. A population filter removed the variant that defines your inbred line, because the filter was built for a different species. Strand bias was real, and the call was an artefact at the end of a read. The PDF was generated from a VCF that has since been overwritten.

A biologist does not have to run the aligner to ask these questions. They do have to refuse a result that arrives with none of the files in the table. Reanalysis is a different piece of work from sequencing, and the statement of work should already say who does it.

A filtered call is a research statement

Nothing in the chain diagnoses a patient, approves a biosafety level, or validates a clinical assay. Annotation databases are incomplete. Filters encode assumptions. A research VCF of human DNA is still research. Institutional review decides whether that DNA should have been sequenced. This page does not. If reads are deposited, archives such as the European Nucleotide Archive need metadata and permission. A gene list in a supplement is not a deposit of the evidence.

Keep an intermediate when the power may drop

Analysis jobs die when a building loses power. The failure mode is a half-written BAM treated as finished the next morning. Write outputs to a second location only after the tool has closed the file, and keep the log that records the reference and the version. A laptop that holds the only VCF is not an archive. In a shared facility, agree who may delete intermediates. Deleting the FASTQ the day the PDF is emailed is how a question becomes unaskable.

Ask for the files when you enquire

State that you want FASTQ, an alignment against a named build, and a VCF with the filter text, not a gene list alone. State who will QC, who will call, and which duplicate rule matches the library. Consumables for producing the reads are in the genomics and sequencing catalogue. The scientific setting is the genomics research pathway. A genome-scale chain can be discussed through the whole-genome sequencing enquiry reference. A locus you want as a trace, beside the short-read call, can be discussed through the Sanger DNA sequencing enquiry reference. Put the file list in the quote request. The pages let you discuss the method. They are not a pipeline you have already run.