Bioinformatics analysis as a separate line

A folder of FASTQ files is a sequencing deliverable. Alignment, quantification, and a statistical contrast are a second deliverable, and they need their own line on the quotation. Mixing them is how a laboratory pays for "analysis" and receives reads, or pays for reads and expects a finished figure. This protocol is how to specify the second line so it can be checked. Commissioning the wider study, including sample QC, is covered in commissioning a sequencing or proteomics study. The enquiry pattern is how to write a laboratory sourcing enquiry. A scope can be discussed through laboratory procurement or academic research, and placed by quotation request or contact. Specifying analysis does not mean a sourcing desk runs the aligner.

What the analysis line is for

You are deciding which interpretive choices are fixed before anyone looks at the outcome. Reference, contrast, and file type are choices. They change the table you will defend. If they stay inside a provider's default, you cannot explain them to a reviewer, and you cannot see that a quotation omitted them. Write them as performance the way you would write a purity class: something the returned files either match or do not.

Raw acquisition remains necessary. You cannot align what you do not have. The point of a separate line is to stop the acquisition from pretending to be the interpretation. NIST laboratory metrology is a public reminder from instrument science that a number needs a stated method. A fold-change needs a stated contrast and a stated reference in the same spirit.

What belongs on the line

Alignment is the placement of reads on a reference, or the equivalent search of spectra against a sequence database. Name the reference. For expression work that may be a genome build plus an annotation, or a transcriptome. For proteomics it may be a protein database. UniProt is one public protein sequence resource; if you use it, name the release, the species, and whether isoforms or a contaminant list are included. "The human database" is not a release.

Quantification is the step that turns alignments or identifications into numbers per sample: counts, abundances, or a precursor intensity. Say the unit. A normalised value and a raw count are different columns. If you need both, say both.

Statistics is the contrast. Name the groups and the comparison: treatment versus control, time point two versus time point one, the covariate you insist on keeping. Say what you want in the table: an effect size, a measure of uncertainty, and the multiple-testing behaviour if you have agreed one. You do not need to paste a software manual. You do need the contrast to be something you designed, not something a default menu suggested after the fact.

File types close the line. Ask for the alignment files if you will inspect tracks, the quantification matrix in a table you can open, and the contrast table as a table rather than only as a heatmap image. Images are views. Tables are the deliverable you can re-filter. Say whether a short methods paragraph will list the reference version and the contrast. That paragraph is how the line stays auditable.

Workflow from reads to a table

Start with primary files that match the sample sheet. If a sample failed process QC, the sheet should say so. Do not begin alignment by dropping that sample in a footnote nobody sees.

Confirm the reference. One person reads the build or the database release out loud against the statement of work. Mismatch is a stop, described below. Match means you may align or search.

Quantify with the rule you wrote. For RNA, that rule might be counts per gene under the annotation you named. For peptides, it might be intensities for identifications that pass the threshold the statement of work names. Keep the threshold in the file, not in a slide.

Only then compute the contrasts that were pre-specified. A contrast that occurs to you because a volcano plot looks exciting is a new analysis. It can be honest if you label it as exploratory and you do not pretend it was the line you commissioned. The commissioned line is the one in the statement of work.

Deliver the files in the formats named. A portal full of plots, with the matrix withheld, has not met a line that asked for the matrix. Public protocol notes on protocols.io illustrate how many small choices sit between a read and a table. Your statement of work will not list every flag. It will list the choices that change the scientific claim: reference, contrast, and the files you can re-open.

A worked expression contrast

Four biological replicates of a treated cell line and four of a vehicle control were sequenced under a separate acquisition scope. The analysis line now says the following. Reference: the species genome build and gene annotation you name, with version identifiers the provider writes back and you accept. Quantification: gene-level counts, plus a normalised matrix if you asked for one. Contrast: treated versus vehicle, with the four and four replicates as the groups, no samples removed unless process QC failed under a rule you already wrote. Files: count matrix, contrast table, and a methods paragraph that repeats the reference version. Alignments are included if you will look at tracks, and excluded if you have said you will not. A heatmap may be included as a picture and is not the only file.

The quotation says "standard differential expression". You treat that phrase as silence. You ask for a revision that names the reference version, the contrast treated versus vehicle, and the count matrix. You do not accept a heatmap as closure.

If someone later asks for a second contrast, resistant versus sensitive, that was not in the design. It is a new line. It may be worth doing. It is not a correction to the first invoice, and it needs its own groups defined before anyone sorts the table to suit a story.

Proteomics follows the same protocol with spectra in place of reads, a database release in place of a genome build, and identification plus quantification as the middle step. The separate line still names the database, the contrasts, and the tables.

The branch when the reference is wrong

Stop interpreting. A wrong build or a wrong species database does not become slightly wrong. Gene or protein identifiers will not line up with the rest of your notebook. Write the correction, confirm you still have the primary files under the custody terms, and rerun. Keep the mistaken output under a name that marks it as superseded. Do not edit identifiers by hand to force a match. Hand edits do not survive the next student, and they do not match the methods paragraph.

If the reference was right and the contrast was wrong, the same branch applies to the statistics only. You may keep the quantification matrix and recompute the contrast if the line allows you to, or you commission that recompute. You do not quietly change group labels in the plot.

If a sample fails a process metric the provider named, such as too few assigned reads under their written rule, stop and decide under the rule. Reinterpret with the sample held out only if the rule said so, and show the hold-out. Do not drop it because it was inconvenient for the volcano plot.

Failure modes

The first is the default pipeline. It may be competent and still be the wrong reference for your species or your annotation habit. Competence is not identity. Read the version.

The second is the orphan contrast. The table compares groups you did not design, or it drops two replicates to improve a picture. Treat that table as a different analysis. Ask for the pre-specified contrast on the full qualified set.

The third is file substitution. You receive a PDF and a screenshot. The line asked for a matrix. The job is not done. Ask for the matrix while the project is still open, which is also why retention of primary files matters on the custody line.

Research limits

An analysis line does not make a biomarker, a diagnosis, or a paper. It makes a table under stated choices. Biological replication was your design. The provider cannot add replicates in software. Human-derived data stay inside the ethics approval that allowed the study. Do not widen a contrast into a clinical claim because the software printed a small number. Small numbers are a statistical output, not a medical instruction.

A note on versions and the next student

The student who joins next year needs the reference string more than they need the heatmap. In a laboratory where people share a drive and power cuts interrupt long jobs, store the primary files and the matrix in the place your institution already backs up. Do not keep the only copy on a laptop that cannot finish a rerun. The statement of work should say you will receive those files, so the backup has something official to hold. Do not invent a compute duration. If turnaround for the analysis matters, ask the quotation to write it, and treat silence as an unknown date.

What to ask for in the enquiry

Ask for two lines. The acquisition line names samples and raw formats. The analysis line names the reference and version, the quantification, the contrasts, and the file types, including a methods paragraph. Ask the provider to propose the version in writing if you do not know the current build, and accept it explicitly. Mark exploratory plots as optional. Send the pair of lines through laboratory procurement or academic research, using contact or a quotation request. A scope can be discussed. When the tables return, check the methods paragraph against the line before anyone drafts a sentence about biology.