A pipeline produces numbers. Statistics helps make sense of them.
Variant calls, expression counts, biomarker measurements. In biology the statistics are unforgiving, because the datasets are high dimensional, the sample sizes are usually small, and the temptation to find something is strong.
The multiple testing problem
Genomic data tests thousands of hypotheses at once. Twenty thousand genes checked for differential expression means twenty thousand statistical tests, and at a conventional significance threshold roughly a thousand of them will appear significant by chance alone.
Every serious biological analysis corrects for this, most often with the Benjamini Hochberg procedure, which controls the false discovery rate rather than the per test error. Reporting uncorrected p values from a genome wide screen is one of the most common errors in the field, and one of the easiest for a reviewer to spot.
False positives expected by chance alone from 20,000 uncorrected tests at a conventional threshold.
Benjamini Hochberg — controls false discovery rate.
Normalization and batch effects
Raw counts from sequencing are not directly comparable across samples. Library size differs, sequencing depth differs, and technical variation from different runs, different days, or different labs — batch effects — can dwarf the biological signal. Normalization corrects for the first set of problems. Batch correction addresses the second, but only if the experimental design allows it.
A study where all treated samples were run on one day and all controls on another cannot separate batch from treatment, no matter what statistics are applied afterward. Design has to be right before analysis starts.
Common analyses
Differential expression
Comparing gene expression between conditions — the workhorse of RNA sequencing analysis. Both standard tools model the count data appropriately rather than treating it as continuous.
Dimensionality reduction
Lets a team see structure in high dimensional data, spot outlier samples, and check whether samples cluster by biology or by batch before formal testing begins.
Survival analysis
Connects molecular findings to clinical outcomes, which is where biomarker work usually has to end up to matter.
Association testing
Links genetic variants to traits or outcomes across a cohort, with the same multiple testing discipline applied at genome scale.
Power and sample size
Small sample sizes are the norm in biology, and underpowered studies produce findings that do not replicate. Power analysis before an experiment estimates how many samples are needed to detect an effect of a given size. It is frequently skipped, and the cost shows up later as a result that cannot be reproduced.
SequoiaAT applies statistical methods and data analysis techniques to biological datasets as part of its genomics and biomarker analysis work — the layer that sits after the pipeline and turns its output into a finding a team can report and defend.
Bioinformatics and Genomics →