Comparisons and the evidence ladder
“High expression” is incomplete until we know compared with what, in which cells, and measured how. The October 4 reports repeatedly show why those questions change the conclusion.
Read the comparison
| Term | Meaning |
|---|---|
| Comparator / reference | The group or baseline against which a result is interpreted |
| Fold change (FC) | Ratio of expression in one group to another under the analysis's normalization |
| log2FC | Base-2 logarithm of that ratio: +1 is twice as high, +2 is four times as high, −1 is half as high |
| Percentile | Position within a defined reference distribution; the 90th percentile is not 90% target-positive cells |
| Batch effect | Systematic differences associated with preparation, run, site, or another technical factor |
| Capture chemistry | How molecules are selected for sequencing; targeted hybrid capture and poly-A selection sample RNA differently |
| Normalization / correction | A method to make measurements more comparable; it cannot remove every source of bias |
Fold changes depend on count normalization and low-count handling. A ratio with a nearly zero denominator can look dramatic while resting on very little signal. edgeR methods guide
The reports compare Personalis ACE4 hybrid-capture bulk RNA with poly-A-based public data and nuclei with whole-cell data. A global scaling correction cannot guarantee that every gene becomes comparable. Treatment exposure, biopsy site, and different tissue pieces add biological differences too. A before/after claim requires a reliable specimen identity and collection date, not only a matching genotype. October 4 problems report
Statistical and technical confidence
Differential expression is a modeled difference between groups. A p-value quantifies incompatibility with a specified null model under its assumptions; it is not the probability a target is useful. FDR, false discovery rate, addresses multiple testing. Small or contaminated comparison groups and unmodeled preparation differences can still undermine biological interpretation even when a statistical threshold passes. Bioconductor comparison methods
QC, quality control, checks whether the material and measurements are fit for the question. A positive control checks that expected signals can be recovered; a negative control helps assess background or false signals. Controls support the assay’s interpretation, rather than automatically validating every result.
Enrichment, selectivity, and dependence
Tumor-enriched means higher in the chosen tumor comparison. Tumor-selective is a stronger claim about discrimination from relevant normal tissue. Off-tumor effects occur when a treatment acts on normal tissue; RNA comparisons alone cannot quantify that risk.
Hypothetical example: A marker is high in malignant epithelial cells, low in fibroblasts, and similarly high in normal luminal cells. It is epithelial-enriched in that comparison; it has not demonstrated tumor selectivity. If the normal group is small or contaminated, even that conclusion needs qualification.
A target is something a drug or immune therapy acts on. A dependency is something the cancer needs for survival or growth, demonstrated through suitable perturbation experiments. A biomarker is a measured characteristic; whether it predicts response depends on validation for the assay, treatment, and setting. A protein can be an accessible delivery address for an ADC without being the tumor's essential growth driver.
Keep each measurement at its own step
For a potential cell-surface target, expression, protein production, localization, and functional dependence need separate measurements.
| Evidence | What it supports | Next question |
|---|---|---|
| DNA variant or gain | A sequence/copy change in the tested material | Is the call authentic and in the relevant cancer cells? |
| RNA abundance | Captured transcripts | Are intact protein and the relevant isoform—a particular molecular version—produced? |
| Total protein | Protein present in the assayed mixture | Which cells contain it, and where is it located? |
| Membrane localization | Protein on the cell surface under the localization assay | Can the treatment bind/access it, and what normal tissues share it? |
| Activation / phosphorylation | A signaling-associated protein modification | Is the cancer dependent on that activity? |
| Functional response | An effect under specified experimental conditions | Does it reproduce and translate to the clinical setting? |
| Clinical validation | Evidence in the intended assay, disease, and treatment context | Does it apply to the current decision? |
“Concordant DNA, RNA, and protein” means agreement across layers. Shared specimens, assumptions, or technical biases can make those layers partly dependent. Agreement is useful evidence; it is not automatic proof of treatment benefit.
When the question is whether a variant is real
A variant caller is software proposing sequence changes. PASS means its filters passed, not clinical confirmation. Consensus means agreement under a stated rule. Read support counts supporting observations; fragment support avoids counting two paired-end reads of one DNA fragment as independent evidence. An alignment artifact is a misleading match to the reference. Positive controls are expected real signals used to check that a test can recover them.
The October 4 audit recovered positive-control variants while failing to reproduce the EVEE headline variants. A pathway story built on those specific calls therefore loses its patient-specific premise; general pathway biology still has its own evidence. Problems report
Say it in your own words
“The RNA result nominates a question. I want the right normal comparator, protein location, and evidence of function before calling it a selective target or a dependency.”
Check yourself: Does a nine-copy gain prove its named gene drives the tumor? No. Several genes can share a gain; expression and functional evidence help identify what matters. Return to the course and discussion card.