# Chapter 125. Bulk RNA-Seq, Targeted RNA-Seq, Library Construction, Biases, and Differential Expression

## Scope Note

This chapter explains how bulk and targeted RNA sequencing convert RNA molecules into count-like evidence, how library construction choices shape what can be observed, and how computational models turn reads into expression inferences. The chapter treats RNA-seq as a chain of molecular, sequencing, and statistical transformations rather than as a black-box list of genes. It focuses on standard short-read bulk RNA-seq and targeted panel assays, while pointing to [Chapter 126](chapter1157.md) for specialized small and stable-RNA profiling, [Chapter 127](chapter1158.md) for RNA-end, tail, and cleavage profiling, [Chapter 128](chapter1117.md) for long-read and direct RNA sequencing, [Chapter 130](chapter1118.md) for single-cell and spatial assays, [Chapter 122](chapter1115.md) for RNA extraction and sample quality, and [Chapter 139](chapter1126.md) for broader experimental-design standards.

## Executive Summary

Bulk RNA-seq measures average RNA abundance across a population of cells, nuclei, tissue fragments, organisms, or environmental material. A standard experiment starts with RNA whose composition is already shaped by collection, preservation, extraction, degradation, and contamination, then enriches or depletes classes of RNA, fragments RNA or cDNA, converts RNA to cDNA, adds adapters and sample indexes, amplifies molecules when necessary, sequences short fragments, assigns reads to genomic or transcript features, normalizes count-like measurements across libraries, and tests whether observed differences are larger than expected from biological and technical variation. Each step changes the set of molecules that remain visible. A result table is therefore not a direct photograph of the transcriptome; it is an inference conditioned on the library design, reference annotation, computational model, and experimental design.

Library design is the first scientific decision. Poly(A) selection enriches mature eukaryotic mRNAs and many polyadenylated long noncoding RNAs, but it loses most bacterial transcripts, many non-polyadenylated RNAs, degraded molecules lacking intact tails, and some histone mRNAs. Ribosomal RNA depletion keeps a broader non-rRNA pool and is often preferred for degraded samples, bacterial samples, viral RNA, host-pathogen mixtures, and studies of nascent or noncoding RNA, but depletion depends on probe sequence complementarity and can leave residual rRNA or remove off-target transcripts. Targeted panels use probes or primers to concentrate specific genes, exons, fusions, pathogens, or RNA classes, which improves depth and clinical feasibility but narrows discovery. These tradeoffs are emphasized in RNA-seq bias reviews and multi-species RNA-seq best-practice discussions.

Library construction adds additional structure. Fragmentation controls insert-size distribution and transcript-body coverage. Random priming, reverse-transcriptase behavior, adapter ligation, PCR cycles, GC content, and strand-specific chemistry can cause nonuniform coverage and molecule-dependent capture. Unique molecular identifiers (UMIs) can distinguish original molecules from PCR duplicates when added before amplification, but UMIs do not remove selection bias, priming bias, or sampling noise. Computational analysis then introduces its own boundary cases. Reads may map uniquely to one gene, ambiguously to paralogs, ambiguously among isoforms that share exons, or to multiple organisms in a mixed sample. Alignment-dependent counting, transcript-level probabilistic quantification, and bias-aware quantification tools such as Salmon represent different ways to allocate ambiguous read evidence.

Differential expression is a statistical comparison, not simply a fold-change calculation. Robust analysis requires biological replication, a design matrix that separates the biological contrast from batch or nuisance variables, normalization that adjusts for library-size and composition differences, dispersion or variance estimation, multiple-testing correction, and careful interpretation of uncertainty. Transcript-level abundance estimates carry inferential uncertainty when reads cannot uniquely distinguish isoforms; recent tree-based and uncertainty-aware methods show that reporting at the right resolution may require aggregating transcripts when transcript-specific evidence is weak. In clinical or targeted RNA-seq, the same principles apply, but validation, limit of detection, reportable range, fusion-calling rules, specimen quality, and off-panel blind spots become central.

## Concept Inventory

- **Bulk RNA-seq:** sequencing-based measurement of RNA abundance in a pooled population of cells, nuclei, tissue, organisms, or environmental material; the output is a population-average signal.
- **Targeted RNA-seq:** sequencing restricted to a predefined set of transcripts, genes, exons, fusions, organisms, or RNA classes by capture probes, amplicon primers, or other enrichment strategies.
- **RNA selection:** enrichment or depletion that changes which RNA molecules enter the library, including poly(A) selection, rRNA depletion, globin depletion, small-RNA size selection, and targeted capture.
- **Library construction bias:** systematic distortion introduced during fragmentation, priming, reverse transcription, adapter ligation, amplification, size selection, sequencing, or computational processing.
- **Strand specificity:** preservation of the original transcriptional strand so reads can distinguish sense, antisense, overlapping, and convergent transcripts.
- **UMI:** a unique molecular identifier, usually a short random or semi-random barcode added before amplification to label original molecules for duplicate correction or molecule counting.
- **Read assignment:** the mapping of sequenced fragments to genes, transcripts, exons, splice junctions, organisms, or synthetic controls.
- **Normalization:** adjustment of count-like measurements to make samples comparable despite different library sizes, RNA composition, capture efficiency, or unwanted technical factors.
- **Differential expression:** statistical inference that an RNA feature differs in abundance between conditions after modeling biological variability, technical variability, and multiple testing.
- **Batch effect:** variation associated with processing time, reagent lot, operator, instrument, lane, site, extraction batch, or other non-biological factors that can mimic or obscure biology.

## What to Know Before Reading This Chapter

The reader should already know that RNA molecules differ in abundance, length, sequence composition, processing state, modification state, and cellular location. The key background from [Chapter 122](chapter1115.md) is that RNA quality and input composition are decided before sequencing begins. A degraded formalin-fixed tumor, a fresh cultured-cell pellet, a bacterial infection sample, and a sorted population of rare cells are not interchangeable starting materials. RNA-seq library construction can reduce some problems, but it cannot fully recover molecules that were selectively lost during collection or extraction.

The reader should also know that short-read sequencing does not usually observe full RNA molecules. Most bulk RNA-seq reads are short fragments sampled from longer RNAs. To infer gene or transcript abundance, a computational method must decide which feature produced each fragment, or must distribute fragment evidence among several possible features. A read that crosses a splice junction can be highly informative for an isoform, while a read from an exon shared by many isoforms may be uninformative for transcript-specific quantification.

Finally, the reader should separate three quantities that are often confused. A raw read count is a number of sequenced fragments assigned to a feature. A molecule count estimates how many original molecules were represented before amplification, often with UMIs. A normalized expression estimate is a model-based quantity intended to make samples or features comparable. None is automatically a direct measure of absolute cellular RNA copy number.

## 125.1. Library design and RNA selection strategies

Bulk RNA-seq begins with a design question: which RNA molecules should be visible in the final data? A whole-transcriptome experiment on mammalian cell lines may be designed to profile mature polyadenylated messenger RNAs, whereas an infection experiment may need host mRNAs, bacterial transcripts that lack stable poly(A) tails, viral RNA, and noncoding RNAs from both host and pathogen. The choice between poly(A) enrichment, rRNA depletion, total RNA sequencing, size selection, targeted capture, and specialized depletion is therefore a biological choice, not a purchasing detail. RNA-seq library-bias reviews emphasize that enrichment, depletion, fragmentation, priming, amplification, and sequencing all shape the final dataset.

![Figure 125.1. Library Design Determines the Observable RNA Population.](../assets/figures/chapter1116_figure1.png)

**Figure 125.1. Library Design Determines the Observable RNA Population.** Figure 125.1. RNA-seq library design as a decision tree. The biological question, organism mix, RNA integrity, target RNA class, input amount, and discovery-versus-validation goal determine whether the experiment uses poly(A) selection, rRNA depletion, total RNA, targeted capture, amplicon panels, or another specialized enrichment.

Poly(A) selection uses oligo(dT) sequences to capture RNAs carrying sufficiently intact poly(A) tails. In many eukaryotic samples, this approach strongly enriches mature mRNAs and reduces the dominance of ribosomal RNA. Poly(A) selection also simplifies analysis because the resulting library is enriched for coding transcripts and polyadenylated long noncoding RNAs. The boundary is equally important: bacterial mRNAs are generally not stably polyadenylated in the eukaryotic mRNA sense; many regulatory RNAs, histone mRNAs, degraded transcripts, nascent RNAs, viral RNAs, and partially processed RNAs can be underrepresented or lost. In degraded samples, such as many formalin-fixed paraffin-embedded tissues, an RNA molecule may have lost its 3′ tail or may be too fragmented for efficient oligo(dT)-based capture, so rRNA depletion or targeted short-amplicon designs may be more appropriate.

Ribosomal RNA depletion takes the opposite route: it removes abundant rRNA and leaves a broader pool of non-rRNA molecules. This can preserve non-polyadenylated RNAs, prokaryotic mRNAs, viral RNAs, pre-mRNAs, many long noncoding RNAs, and degraded fragments. The cost is that rRNA depletion depends on the match between depletion probes and the sample's rRNA sequences. A depletion kit optimized for human, mouse, or rat may not perform well in a mixed fungal, bacterial, plant, or environmental sample. Probes can also bind unintended RNAs, and residual rRNA can consume sequencing capacity. Multi-species RNA-seq guidance therefore recommends thinking about organism composition, expected RNA mass per organism, and the desired read depth for the least abundant member before choosing a selection strategy.

**Table 125.1. RNA Selection and Targeting Strategy Matrix.** Table 125.1. RNA selection strategies and their interpretive boundaries. Each strategy enriches a different molecule set and creates different negative-space claims.

| Strategy | Best-fit questions | Retained molecules | Commonly lost or underrepresented molecules | Common artifacts | Useful QC metrics | Related chapters |
| --- | --- | --- | --- | --- | --- | --- |
| **Poly(A) selection** | Mature eukaryotic mRNA abundance in intact samples | Polyadenylated mRNAs and many polyadenylated lncRNAs | Bacterial mRNAs, histone mRNAs, many noncoding RNAs, degraded RNAs lacking intact tails | Poly(A)-tail enrichment, loss of non-polyadenylated RNA classes, degraded-sample dropout | Insert size, tail-end bias, mapping rate, exonic fraction, rRNA carryover | `[Chapter 29](chapter1028.md)`, `[Chapter 90](chapter1085.md)`, `[Chapter 122](chapter1115.md)` |
| **rRNA depletion or total RNA-seq** | Broader non-rRNA profiling, degraded specimens, bacteria, viruses, host-pathogen mixtures | mRNA, pre-mRNA, many lncRNAs, viral RNA, bacterial RNA, degraded fragments | RNAs removed by off-target probes; low-abundance RNAs when residual rRNA dominates | Residual rRNA, probe mismatch in non-model organisms, intronic or immature-RNA enrichment | Percent rRNA, organism read fractions, intronic fraction, transcript-body coverage | `[Chapter 42](chapter1039.md)`, `[Chapter 111](chapter1106.md)`, `[Chapter 122](chapter1115.md)` |
| **Targeted hybrid capture** | Deep evidence for selected genes, fusions, pathogens, or isoforms | Probe-covered exons, junctions, gene panels, pathogen targets | Off-panel transcripts, poorly captured regions, unknown targets outside probe design | Probe-efficiency bias, uneven on-target coverage, false negative outside reportable range | On-target rate, uniformity, target coverage, duplicate or UMI rate, limit-of-detection controls | `[Chapter 128](chapter1117.md)`, `[Chapter 139](chapter1126.md)`, `[Chapter 163](chapter1150.md)` |
| **Amplicon or anchored multiplex panel** | Fast focused detection of known regions or fusions with limited input | Primer-defined regions; anchored designs can recover unknown partners adjacent to target primers | Regions without primer support; transcripts with primer-site variants or degraded primer sites | Primer dropout, amplification bias, chimera or fusion artifacts, overcounting without UMIs | Amplicon coverage, UMI family support, negative controls, fusion read support, specimen pass rate | `[Chapter 139](chapter1126.md)`, `[Chapter 121](chapter1148.md)`, `[Chapter 163](chapter1150.md)` |
| **Size or fraction selection** | Small RNA, long RNA, nuclear, cytoplasmic, or compartment-focused questions | Molecules in the selected size range or fraction | RNAs outside the chosen size range or fraction; fragile RNP-associated RNAs lost during handling | Fragment-size skew, fraction contamination, selective loss during extraction or cleanup | Size profile, fraction marker RNAs, recovery yield, library complexity | `[Chapter 89](chapter1084.md)`, `[Chapter 122](chapter1115.md)`, `[Chapter 132](chapter1120.md)` |

RNA selection strategy interacts with read depth. A library with 40 million paired-end reads can still fail for a pathogen if host RNA contributes nearly all molecules. Conversely, a targeted capture panel can produce high coverage for a rare organism or a short gene list, but only for sequences represented by probes. In host-pathogen or symbiosis experiments, the experimenter may need pilot sequencing or quantitative reverse-transcription PCR to estimate host and microbe RNA proportions before choosing a final depth and enrichment scheme. Counts per million thresholds and saturation curves are useful diagnostics only after the researcher remembers what population of molecules was allowed to enter the library.

The design boundary should be written into every interpretation. A poly(A)-selected human RNA-seq result can support claims about expressed polyadenylated transcripts under the sampled conditions, but it should not be used to infer that a non-polyadenylated RNA is absent. A ribo-depleted total-RNA library can support broader RNA-class discovery, but it often carries more intronic, nascent, degraded, or immature signal that must be interpreted with the correct gene model. A targeted assay can be excellent for a clinical fusion or an actionable cancer-gene panel, but the absence of an off-panel transcript is not evidence that the transcript is absent from the specimen.

> **Box 125.1. Negative Results Depend on the Observable Molecule Pool**
>
> **Negative evidence in RNA-seq is conditional.** A transcript can be absent from a result table because the RNA was not present in the specimen, but it can also disappear because the assay never gave that molecule a fair chance to be measured. Poly(A) selection underrepresents many non-polyadenylated RNAs and degraded molecules lacking intact tails. rRNA depletion can fail if depletion probes poorly match the organism, or can consume depth if residual rRNA remains. Targeted panels only test sequences inside the reportable design. Alignment and annotation choices can also hide molecules that are missing from the reference or difficult to distinguish from paralogs and repeats. A careful negative statement therefore names the measurement boundary: "not detected in this poly(A)-selected, short-read, human-annotation analysis" is stronger and more honest than "not expressed."

## 125.2. Fragmentation, priming, amplification, and strand specificity

After RNA selection, short-read RNA-seq must convert molecules into sequenceable fragments. Fragmentation controls the insert-size distribution and the spacing of read evidence across each transcript. Fragmentation can occur at the RNA stage before reverse transcription or at the cDNA stage after reverse transcription. RNA fragmentation may reduce 3′ bias when compared with oligo(dT)-primed full-length cDNA workflows, but chemical, enzymatic, and physical fragmentation methods can have different sequence and length preferences. cDNA fragmentation can be useful when starting material is scarce or workflows combine tagmentation with adapter addition, but sequence-specific transposase or nuclease preferences can leave uneven coverage.

![Figure 125.2. Bias Entry Points During Library Construction.](../assets/figures/chapter1116_figure2.png)

**Figure 125.2. Bias Entry Points During Library Construction.** Figure 125.2. Biases enter at multiple library-construction steps. RNA or cDNA fragmentation, priming, reverse transcription, adapter ligation, PCR amplification, size selection, and sequencing chemistry each alter which fragments are overrepresented, underrepresented, duplicated, or missing.

Priming is the step that starts complementary DNA synthesis. Random hexamers are often described as random, but the priming process is not perfectly uniform. Primers have sequence preferences, RNA secondary structure can block access, and reverse transcriptase can stall, template-switch, or misprime. Oligo(dT) priming focuses the library near the poly(A) tail and is useful for intact polyadenylated transcripts, but it can produce 3′ end enrichment and poor recovery of degraded molecules. Gene-specific priming can improve sensitivity for targeted assays but narrows the assay to expected sequences and makes primer design a source of interpretive bias. The evidence basis for these statements comes from comparative library-preparation work summarized in bias reviews, but the local bibliography still needs more landmark primary keys for random-priming and reverse-transcription bias. Curation-deferred source need: add local keys for Hansen et al. random-hexamer bias and Levin et al. strand-specific protocol comparisons.

PCR amplification is usually necessary when input is limited, but PCR is also a major source of library distortion. Templates do not amplify with equal probability. GC-rich and AT-rich fragments can be inefficient, early stochastic differences are magnified during later cycles, and overamplified libraries accumulate duplicate reads that do not represent independent starting molecules. Reducing cycle number, using polymerases and buffers that better handle sequence extremes, increasing input when possible, and preserving molecular complexity can mitigate but not eliminate these effects. A PCR-free or low-amplification workflow can reduce bias when enough input material exists, but this tradeoff is not available for many clinical, rare-cell, or degraded specimens.

Strand-specific library construction preserves whether a fragment came from the sense or antisense strand of the original RNA. This matters in compact genomes, overlapping genes, antisense transcription, intronic transcription, enhancer RNAs, bacterial operons, and fusion detection. Several common strand-specific designs mark the second cDNA strand, for example through dUTP incorporation followed by selective degradation or suppression of second-strand amplification. The main interpretation rule is simple: if the library is not strand-specific, reads from overlapping genes on opposite strands may not distinguish sense from antisense transcription. If the library is strand-specific but strand information is incorrectly declared in the analysis software, expression can be assigned to the wrong strand. Reporting the strandedness assay and verifying it with diagnostic read distributions are therefore part of data quality, not optional metadata.

## 125.3. UMIs, deduplication, and molecule counting

A unique molecular identifier is a short barcode attached to an RNA-derived molecule before substantial amplification. If two reads map to the same position and carry the same UMI, the reads are usually interpreted as amplification descendants of one original molecule rather than as independent molecules. UMI-based deduplication can therefore convert a read pile into a molecule-count estimate and can reduce PCR-driven overcounting. UMIs are especially valuable in low-input, targeted, single-cell, and amplicon-like assays where amplification duplicates can dominate the library.

![Figure 125.3. UMI-Based Molecule Counting and Deduplication.](../assets/figures/chapter1116_figure3.png)

**Figure 125.3. UMI-Based Molecule Counting and Deduplication.** Figure 125.3. A UMI labels an RNA-derived molecule before amplification. Reads with the same mapping position and UMI can be collapsed as amplification descendants, while UMI collision, sequencing errors, and assay-specific fragment boundaries determine how deduplication should be performed.

UMIs are not magic labels for truth. A UMI must be attached before the amplification step whose duplicates it is meant to identify. If a barcode is added after reverse transcription, it cannot distinguish losses or biases that occurred during RNA capture or cDNA synthesis. UMI length and diversity must be large enough for the number of molecules being tagged; otherwise different original molecules can receive the same barcode, a problem called UMI collision. Sequencing errors in the UMI can split one molecule into several apparent molecules unless the deduplication method clusters similar barcodes. Conversely, aggressive clustering can merge distinct molecules when a highly expressed feature has many real molecules with related UMI sequences.

Deduplication also depends on how molecule identity is defined. In a random-fragment library, reads with the same UMI but different start positions may represent different molecules, while in an amplicon panel every true molecule may share the same primer-defined boundaries. A deduplication rule that works for one assay can be wrong for another. For gene-level bulk RNA-seq with high-complexity random fragments, simple coordinate-based duplicate removal without UMIs can remove real biological molecules from highly expressed genes. For targeted panels with short amplicons, failure to account for UMIs can inflate counts and variant evidence. The correct reporting unit should therefore state whether the table contains raw reads, deduplicated reads, UMI-collapsed molecules, or model-estimated abundances.

UMI molecule counts reduce one class of technical artifact but leave several others intact. They do not correct for RNA degradation, inefficient rRNA depletion, sequence-specific priming, capture-probe dropout, transcript-length effects, multimapping, or batch effects. A model can use UMIs as a stronger measurement layer, but it still requires biological replication and a defensible statistical design. Curation-deferred source need: add local method keys for UMI-tools, molecular-barcode error correction, and UMI collision models before treating UMI-specific claims as fully supported.

## 125.4. Alignment, transcript assignment, and quantification

Sequencing produces reads, not gene counts. Alignment and quantification decide how those reads support features. In a genome-alignment workflow, reads are trimmed, quality-checked, mapped to a reference genome, and assigned to annotated genes, exons, splice junctions, or other features. Eukaryotic mRNA analysis usually requires splice-aware aligners because reads may cross exon-exon junctions. Bacterial, archaeal, many viral, and organellar workflows often use splice-agnostic mapping, but operons, overlapping genes, antisense transcription, and compact genomes still create assignment problems. In a transcriptome-quantification workflow, reads are compared to transcript sequences, and ambiguous fragments are distributed among transcripts by a probabilistic model.

![Figure 125.4. From Reads to Features to Uncertainty.](../assets/figures/chapter1116_figure4.png)

**Figure 125.4. From Reads to Features to Uncertainty.** Figure 125.4. Short reads become expression evidence through reference selection, mapping or transcript compatibility, feature assignment, quantification, and uncertainty estimation. Reads unique to a gene or isoform carry different information from reads shared among isoforms, paralogs, repeats, or organisms.

Multi-mapping is not a minor edge case. Reads can map equally well to paralogous genes, repetitive elements, pseudogenes, conserved domains, duplicated exons, shared untranslated regions, or several isoforms of the same gene. Gene-level counting often hides isoform ambiguity by summing reads over a gene, but this can obscure transcript switching. Transcript-level quantification tries to resolve isoforms, but short reads from shared exons may not contain enough information. Salmon and related transcript-quantification methods model fragment compatibility and bias-aware abundance estimation rather than treating each read as a uniquely observed transcript molecule. The correct feature resolution depends on the data: sometimes a transcript is well supported; sometimes the honest result is a gene-level signal or a transcript-group signal.

Reference choice is part of the measurement. If the genome assembly lacks an isoform, a noncoding RNA, a pathogen strain, or a fusion partner, reads from that molecule may be discarded or misassigned. Annotation versions also affect gene counts because exon boundaries, transcript models, merged genes, and pseudogene labels change across releases. In multi-species RNA-seq, a composite reference containing all expected organisms can reduce incorrect organism assignment, whereas separate mapping can assign ambiguous reads to the wrong species when homologous sequences exist. For host-associated bacteria, lateral gene transfer, conserved rRNA, and shared repeats can still complicate classification.

Quantification output uses several scales. Counts are integer or near-integer assignments used by many differential-expression models. Transcripts per million and related abundance units adjust for feature length and library composition, making them useful for within-sample transcript-abundance summaries, but they are not usually the preferred input for count-based differential testing. Effective length correction matters because longer transcripts generate more fragments at the same molar abundance. Bias correction can account for fragment sequence, GC content, positional effects, and fragment-length distribution, but correction is model-dependent and cannot recover molecules absent from the library. Curation-deferred source need: add local landmark keys for RSEM, kallisto, STAR, HISAT2, featureCounts, tximport, and transcript-to-gene summarization.

> **Box 125.2. Do Not Swap Counts, TPMs, and Molecule Counts**
>
> **Ask what the number counts before comparing it.** A raw count is the number of sequenced fragments assigned to a feature under a specific reference and assignment rule. A UMI-collapsed molecule count estimates how many tagged molecules survived the library workflow, but it does not correct for RNA selection, reverse-transcription loss, capture dropout, or biological sampling. TPM and related length-normalized abundance units are useful for asking which transcripts dominate one library, because they adjust for feature length and library composition. Count-based differential-expression tools usually expect count-like values because their variance models describe sampling and overdispersion across replicates. Feeding TPMs into a count model, or treating UMI-collapsed counts as absolute cellular copy number, mixes measurement scales and can produce confident-looking but poorly grounded conclusions.

## 125.5. Normalization, differential testing, and uncertainty

Differential expression asks whether an RNA feature differs between conditions after accounting for sampling, biological variation, technical variation, and the large number of features tested. The most basic input is a matrix with features in rows and samples in columns. Before testing, features with negligible counts are usually filtered because they carry little power and inflate the number of hypotheses. Samples should be inspected by library size, mapping rate, RNA selection metrics, strandedness, duplication, transcript-body coverage, principal-component structure, and known covariates. A sample that is a clear outlier because of degradation, contamination, failed depletion, low complexity, or a lane problem should not be rescued by a statistical model without explanation.

Normalization is needed because samples differ in sequencing depth and RNA composition. If one condition strongly induces a small set of highly abundant transcripts, the remaining transcripts may appear to decrease as a compositional consequence even when their absolute molecule numbers did not change. Size-factor and trimmed-mean approaches estimate scaling factors from the distribution of counts rather than assuming that total reads alone are sufficient. Improved normalization methods continue to be an active area because assumptions about most genes being unchanged, control genes being stable, or unwanted factors being separable are not always satisfied. Curation-deferred source need: curate local full-text or landmark references for TMM, median-ratio normalization, upper-quartile normalization, RUV, SVA, and spike-in normalization.

Differential-expression models for bulk RNA-seq usually estimate mean abundance and biological variability for each feature across replicates. Negative-binomial models became common because RNA-seq count variability is larger than simple Poisson sampling would predict. Replication is therefore not a luxury. Without biological replicates, the analysis cannot estimate condition-specific variability from the experiment itself and must borrow assumptions from other data or report descriptive differences rather than strong inferential claims. Multiple-testing correction is required because thousands of genes or transcripts are tested. False discovery rate procedures control the expected fraction of false positives among called discoveries under the model assumptions, but they do not protect against confounded designs, mislabeled samples, unmodeled batches, or wrong feature assignment.

**Table 125.2. Differential-Expression and Uncertainty Reporting Checklist.** Table 125.2. A defensible RNA-seq differential-expression report documents design, normalization, variance modeling, uncertainty, effect size, multiple testing, batch covariates, and feature annotation.

| Decision or metadata item | Why it matters | Example values | Common failure mode | Related claim IDs |
| --- | --- | --- | --- | --- |
| **Biological replication** | Estimates condition-level variability and supports inferential testing | 3 or more independent biological replicates per condition when feasible | Single sample per condition reported as differential expression | |
| **Design matrix and covariates** | Defines the biological contrast and separates modeled nuisance factors | Condition plus batch, sex, donor, lane, RIN, or site when identifiable | Batch perfectly confounded with condition or covariates overfit in small designs |, |
| **Feature definition and annotation version** | Determines which reads count toward each gene, transcript, exon, or junction | GENCODE or Ensembl release, RefSeq set, custom host-pathogen reference | Gene counts change because annotation or organism reference is undocumented |, |
| **Filtering and normalization** | Removes uninformative features and adjusts for depth and composition | Low-count filter, size factors, trimmed mean, spike-in-aware scaling when justified | Raw counts, TPMs, or total-read scaling used as generic DE input without model fit | |
| **Dispersion or variance model** | Models overdispersion beyond Poisson sampling in count data | Negative-binomial model, mean-variance trend, empirical Bayes borrowing | Read depth treated as a substitute for biological variability | |
| **Inferential uncertainty** | Records ambiguity from isoforms, paralogs, multimapping, and transcript groups | Bootstrap or Gibbs replicates, transcript-group tests, gene-level fallback | Transcript-level calls reported without acknowledging shared-exon ambiguity | |
| **Effect size and multiple testing** | Separates magnitude, uncertainty, and genome-wide error control | Log2 fold change, confidence or credible interval, adjusted p-value or FDR | Small adjusted p-value interpreted as large biological effect |, |
| **Reproducibility metadata** | Makes the measurement auditable and reanalyzable | RNA quality, selection chemistry, strandedness, UMI status, read geometry, software versions, accessions | Published gene list lacks sample, library, reference, or analysis provenance | |

Uncertainty exists at more than one level. A gene-level count can be uncertain because reads map ambiguously to paralogs or pseudogenes. A transcript-level estimate can be uncertain because isoforms share exons. A condition effect can be uncertain because biological replicates are few or variable. Inferential replicate methods propagate transcript-quantification uncertainty into downstream analysis, and tree-based methods can report differential signal at transcript-group resolution when individual transcript resolution is not trustworthy. This is not a failure of RNA-seq; it is a recognition that short reads sometimes support a broader feature than the feature the researcher initially hoped to test.

Effect size and significance answer different questions. A small log-fold change with a tiny adjusted p-value may be reproducible but biologically modest. A large fold change for a low-count gene may be unstable or driven by a few reads. Shrinkage estimators, confidence intervals, independent filtering, and diagnostic plots help prevent overinterpretation. Pathway or gene-set enrichment can summarize coordinated changes, but gene-set results inherit biases from gene length, expression level, annotation coverage, pathway redundancy, and the differential-expression threshold. A transparent differential-expression result should therefore report the design formula, sample exclusions, normalization method, feature annotation version, filtering criteria, multiple-testing method, effect-size estimates, uncertainty estimates, and software versions.

## 125.6. Targeted panels, capture methods, and clinical assays

Targeted RNA-seq restricts sequencing to expected regions so that a smaller assay can achieve higher depth, faster turnaround, or clinical validation for a defined question. Hybrid-capture panels use biotinylated probes to pull down cDNA or RNA-derived library molecules from target genes or exons. Amplicon panels use primers to amplify specific regions. Anchored multiplex designs can detect fusions when only one partner is known, because a gene-specific primer can be paired with a universal adapter primer. Each design answers a different question: expression of a gene list, detection of known splice events, discovery of fusion partners for selected genes, detection of pathogen transcripts, or isoform characterization in a focused gene set.

![Figure 125.5. Targeted RNA-Seq Evidence Ladder.](../assets/figures/chapter1116_figure5.png)

**Figure 125.5. Targeted RNA-Seq Evidence Ladder.** Figure 125.5. Targeted RNA-seq concentrates evidence within a designed reportable range. Probe or primer design, RNA quality, on-target rate, molecule or read support, fusion-calling rules, validation thresholds, and clinical interpretation define what a positive or negative result can mean.

Capture is most valuable when the target RNA is rare relative to background RNA. In multi-species RNA-seq, targeted capture can enrich transcripts from a minor organism when host or major-organism RNA overwhelms the library. In cancer research, targeted long-read RNA sequencing methods such as TEQUILA-seq use tiled probes to enrich defined genes and then sequence longer molecules to reveal isoforms and aberrant transcript structures within the targeted set. In clinical oncology, targeted RNA panels are often used to detect fusions, expressed variants, exon skipping, and diagnostically informative expression signatures from limited or degraded tissue.

Clinical targeted RNA-seq has a different evidence standard from exploratory bulk RNA-seq. A clinical assay must define specimen requirements, input range, library metrics, accepted failure modes, sensitivity, specificity, limit of detection, reportable genes, variant or fusion calling thresholds, confirmation policy, and how results are interpreted with pathology and other molecular tests. A recent pediatric rhabdomyosarcoma study illustrates this clinical use: a targeted RNA-seq panel applied to formalin-fixed and fresh frozen tumors identified point mutations and gene fusions that supported molecular classification and potential management decisions, while also noting that targeted RNA-seq does not provide genome-wide copy-number information.

Targeted assays also create blind spots. A fusion cannot be detected if neither partner or breakpoint-adjacent region is represented by the assay design, unless the chemistry is specifically designed for unknown partners. A capture panel cannot quantify transcripts absent from the probe set. Amplicon dropout can occur when mutations, RNA editing, degradation, or polymorphisms disrupt primer binding. A high on-target rate is not the same as unbiased representation, because different probes capture with different efficiency. Targeted RNA-seq is therefore strongest when the clinical or biological question is explicit and the negative space of the assay is reported.

## 125.7. Biases, batch effects, reproducibility, and reporting

RNA-seq bias is any systematic distortion that causes the measured data to differ from the underlying molecules in a reproducible direction. Some biases are molecular: degradation favors short or protected fragments; extraction can lose small or structured RNAs; rRNA depletion can leave residual rRNA; reverse transcription can stall at structured or modified sites; PCR can favor sequence compositions; and adapter ligation can prefer particular ends. Other biases are computational: low-complexity reads may be filtered, multi-mapped reads may be discarded, annotation may miss isoforms, and normalization may assume a stable majority that does not exist. The final result combines all of these effects.

Batch effects are systematic differences associated with processing rather than the biological contrast. They include extraction date, library kit lot, operator, instrument, sequencing lane, flowcell, site, storage time, RNA integrity, and analysis version. Batch effects become especially dangerous when they are confounded with condition, such as all controls processed in January and all treated samples processed in March. In that design, no statistical method can reliably decide whether a difference is biology or month. The remedy is prevention: randomization, balanced blocking, consistent protocols, sufficient replication within batches, recorded metadata, and diagnostic plots before final testing.

> **Box 125.3. Batch Confounding Cannot Be Corrected Afterward**
>
> **A confounded design removes the comparison the model needs.** Suppose all control samples are extracted by one technician using one reagent lot, and all treated samples are extracted a month later by another technician using a new lot. If treated samples differ from controls, the data contain no internal evidence that separates treatment from extraction month, operator, or lot. A batch-aware model can help when each condition is represented within each batch, because the model can compare treated and control samples that share processing history. It cannot solve a design in which biology and processing are perfectly aligned. The practical rule is to randomize, block, and record metadata before sequencing. After sequencing, inspect principal components, library metrics, and known covariates, but do not present software correction as proof that a fully confounded contrast was rescued.

Reproducibility requires enough information for another laboratory or analyst to understand what was measured. Minimum reporting should include sample source, preservation, RNA integrity or degradation metrics, input amount, selection strategy, library kit or chemistry, strandedness, UMI status, fragmentation method, amplification cycles when available, read configuration, sequencing depth, adapter and quality trimming, alignment or quantification software, reference genome and annotation version, filtering criteria, normalization method, statistical model, covariates, multiple-testing method, and data-accession information. The reporting burden is higher for targeted and clinical assays because the reportable range and validation evidence are part of the scientific claim.

The most common interpretation error is treating RNA-seq as direct causality. Differential expression does not prove direct transcriptional regulation, protein abundance, pathway activity, or phenotype causation. A gene can increase in RNA abundance because transcription increases, decay decreases, a cell population expands, an isoform switches, or a sample contains more of a tissue compartment. A targeted panel can detect a fusion transcript but cannot alone describe all genomic rearrangement complexity. A bulk tissue RNA-seq result can reflect changing cell-type composition rather than changing expression within each cell type. Orthogonal validation, perturbation, protein measurement, imaging, and single-cell or spatial assays may be needed depending on the claim.

## Experimental Foundations and Evidence Standards

Good RNA-seq evidence begins with the match between question, sample, and measurement. If the question is mature mRNA abundance in a cell line, poly(A)-selected paired-end short-read RNA-seq with biological replication may be adequate. If the question is bacterial gene expression inside host tissue, poly(A) selection is usually inappropriate because bacterial mRNAs lack stable poly(A) tails, and a composite host-pathogen reference or enrichment strategy may be required. If the question is fusion detection in a pediatric tumor, a validated targeted RNA panel may be more useful than low-depth whole-transcriptome sequencing. If the question is isoform structure, [Chapter 128](chapter1117.md) methods may be needed because short reads often cannot resolve full transcript architecture.

Evidence also depends on controls. External spike-ins can monitor technical performance, but spike-ins do not necessarily mimic endogenous RNA extraction, fragmentation, priming, or capture. Synthetic controls are useful for calibration and troubleshooting; they are not a substitute for biological replicates. Negative controls can reveal contamination, barcode bleed, or index hopping. Technical replicates can evaluate library or sequencing variability, but biological replicates are needed for condition-level inference. Orthogonal validation should be chosen for the claim: reverse-transcription quantitative PCR for a specific transcript junction, digital PCR for low-level fusion evidence, northern blotting or capillary electrophoresis for size, proteomics for protein consequence, or perturbation for causal regulation. [Chapter 123](chapter1155.md) explains the targeted detection assays and their quantitative boundaries in depth.

An RNA-seq result is strongest when the same conclusion survives reasonable changes in analysis choices. This does not mean trying many pipelines until the desired result appears. It means testing whether conclusions are robust to defensible alternatives: gene-level versus transcript-level summarization, different filtering thresholds, inclusion of known batch covariates, removal of obvious failed samples, or use of independent validation cohorts. Sensitivity analysis is especially important for low-count features, repetitive genes, transcript isoforms, targeted panels, and multi-species samples.

## Biological, Clinical, and Technical Contexts

Bulk RNA-seq is well suited to questions about average transcript abundance across a defined sample, such as treatment response in a homogeneous cell line, organ-level response to infection, tumor expression signatures, environmental metatranscriptomic activity, or expression quantitative trait studies. Its weakness is that a bulk profile mixes cell states and cell types. A higher inflammatory RNA signal in diseased tissue can mean that each resident cell increased inflammatory transcription, that more immune cells entered the tissue, or both. [Chapter 130](chapter1118.md) discusses single-cell and spatial strategies for separating these possibilities.

Organism biology changes library design. Eukaryotic splicing requires splice-aware analysis and makes isoform ambiguity central. Bacterial transcripts are often polycistronic, compact, non-polyadenylated, and organized around operons, so gene-level assignment can be affected by overlapping transcription and incomplete transcript annotations. Viral RNA can be highly structured, overlapping, segmented, subgenomic, or host-derived depending on the virus. Plant, fungal, and protist transcriptomes add distinct rRNA sequences, organellar RNAs, cell-wall extraction issues, and annotation limitations. A generic mammalian mRNA-seq workflow should not be assumed to transfer intact across these contexts.

Clinical specimens add constraints. Formalin fixation fragments and chemically modifies nucleic acids. Biopsies can have low input, tumor purity variation, necrosis, blood contamination, and stromal admixture. A clinical report may need a clear answer about a fusion, expressed variant, or therapy-associated marker from a marginal specimen. Targeted RNA-seq can be appropriate in that setting because it concentrates evidence where the assay has been validated, but the result must be interpreted with pathology, DNA testing, copy-number methods, and clinical context.

## Recent Consensus

The current consensus is that RNA-seq is a mature and powerful measurement system when its design assumptions are explicit. Library construction choices are not interchangeable, and no single RNA-seq protocol measures "the transcriptome" in an unrestricted sense. Poly(A) selection, rRNA depletion, targeted capture, short-read sequencing, long-read sequencing, and UMI-based counting all define different observable molecule sets. Bias should be minimized, measured, and reported rather than imagined away.

For computational analysis, the consensus is that raw read totals are not directly comparable across samples; normalization and variance modeling are required for differential testing. Biological replication, design balance, sample metadata, and batch-aware analysis are more important than small differences among otherwise defensible software tools. Transcript-level estimates require special caution because short reads often cannot distinguish isoforms; uncertainty-aware methods and transcript-group reporting are active attempts to make the resolution of the result match the resolution of the data.

For targeted and clinical assays, the consensus is that restricted panels can be scientifically and clinically stronger than whole-transcriptome assays when the question is restricted, the assay is validated, and the blind spots are disclosed. The same assay should not be used to support claims outside its reportable range. In oncology, targeted RNA-seq is particularly useful for expressed fusions and splice alterations, but it complements rather than replaces DNA sequencing, copy-number analysis, methylation profiling, histopathology, and clinical assessment.

## Open Questions, Controversies, Deprecated Models, and Common Misconceptions

Controversies:

- One unsettled area is how best to integrate absolute molecule standards, spike-ins, and compositional normalization. Spike-ins can help diagnose technical variation, but they can behave differently from endogenous RNA and may not solve global RNA composition changes. Another open area is transcript-level differential testing. Tools increasingly propagate quantification uncertainty, but there is no universal answer for when to report transcript, transcript-group, exon, event, or gene-level changes. The statistically appropriate feature may not match the biological story a researcher hoped to tell.
- Another active area is bias correction. Fragment sequence, GC content, positional coverage, and priming models improve quantification, but correction can introduce its own assumptions. A correction model trained on one organism, protocol, or read geometry may not generalize. Clinical targeted panels have parallel issues: probe-level normalization, amplicon dropout, fusion false positives, and degraded-sample artifacts must be evaluated within the validated assay.

Common misconceptions:

- "More reads can fix a badly matched library design." Sequencing a poly(A)-selected bacterial sample deeper does not efficiently recover the bacterial mRNA population; the library design must match the RNA class.
- "A fold change without uncertainty is a differential-expression result." Differential-expression claims require uncertainty estimates, replication, model assumptions, and appropriate multiple-testing control.
- "A small adjusted p-value proves biological importance." Statistical significance can reflect precision, sample size, or confounding; biological importance requires effect size, context, and mechanism.
- "Absence from a targeted panel means absence from the sample." Targeted panels only report features within their design and sensitivity range.
- "Batch correction cannot rescue a fully confounded design." Sixth, bulk RNA-seq measures mixtures; it should not be overread as cell-intrinsic regulation without evidence that cell composition is stable or modeled.