# Multi-Mapping Reads in RNA-seq

Multi-mapping reads are sequencing reads that align equally well to more than one location in the reference genome or transcriptome, while unique reads align to exactly one location. The rule of thumb for choosing a handling strategy is straightforward: discard multi-mappers or use a merged-gene counter when your question is differential expression of well-annotated, mostly unique protein-coding genes, and use a probabilistic or hierarchical allocation method when your question involves transcript isoforms, duplicated genes, allele-specific expression, or non-coding RNAs, where multi-mappers carry most of the signal.

This guide explains what multi-mapping reads in RNA-seq are, why they appear, the main options for handling them with their pros and cons, and how to pick the right approach for your experiment. It also covers practical commands and a troubleshooting section for the most common mapping errors.

## What Multi-Mapping Reads Are and Why They Happen

A multi-mapping read is a read whose best alignment score is tied, or nearly tied, at two or more loci. The aligner cannot decide which locus produced the read, so it reports the read as ambiguous. Most aligners report this as a multimapping count in their summary log, and most either keep a limited number of alignments or discard the read entirely.

Multi-mapping arises from several mechanisms that duplicate sequence across a genome. Segmental duplications, whole-genome duplication, retrotransposition, and recombination all create regions where two or more genes share long stretches of identical or near-identical sequence [1]. Paralogs that descended from a common ancestor often retain shared exons or shared protein domains. Some genes sit inside other genes, so a read from the embedded gene can also align to the host gene. Transcripts from different RNA biotypes can overlap in the same genomic region, and some transcripts, such as ribosomal RNAs, are present in the genome in high copy numbers [1].

The extent of multi-mapping depends strongly on biotype. Long non-coding RNAs and messenger RNAs share less sequence similarity with other genes than biotypes that encode shorter RNAs, which means short RNA genes and repetitive RNA families are far more prone to ambiguous alignment [1]. In a diploid transcriptome that incorporates known genetic variants, multi-mapping reads aligning equally well to multiple genomic locations, isoforms, or alleles can comprise more than 85% of reads [2]. In a standard reference-genome alignment of a poly(A)-selected human library, the multi-mapping fraction is usually far lower, but the exact number depends on the genome, the read length, and the library type.

The core problem is one of information. A read drawn from a duplicated gene is indistinguishable at the sequence level from a read drawn from its paralog. Every downstream decision about that read is a statistical assumption, and different assumptions produce different expression estimates for the same data.

## Why the Handling Choice Changes Your Results

Three broad strategies have historically been used for duplicated genes: drop the duplicated genes, distribute the reads uniformly across their mapping locations, or estimate expression with a statistical model. Each of these produces biased results in different ways.

Discarding multi-mappers is the most common default. It is simple and conservative, but it systematically undercounts any gene with duplicated sequence, and it discards real information. In allele-specific expression analysis, where reads must be assigned between the two parental alleles of the same transcript, discarding multi-mappers can remove most of the usable data and introduces bias [2]. Weighted allocation methods that assign multi-mappers in proportion to estimated abundance outperform methods that discard them, and hierarchical allocation improves the estimates further [2].

[Uniform distribution](/blog/data-analysis/uniform-distribution), in which each multi-mapping read is split equally across all of its alignments, keeps every read but assumes that all loci sharing a read express it at the same level. That assumption fails whenever paralogs are differentially regulated, which is exactly the situation in which duplicated genes are scientifically interesting [3].

Statistical allocation methods, usually based on expectation-maximization, use the pattern of reads across loci to infer which locus each read most likely came from. These methods are unbiased under their model assumptions, but they depend on a good transcriptome reference and on enough informative reads to distinguish the loci.

The practical consequence is that two pipelines processing the same FASTQ files can produce different count matrices for the same genes, and downstream differential expression results can differ. The choice of multi-mapping strategy is therefore not a technical detail. It is part of your experimental design.

## The Core Comparison: Every Option Side by Side

![Decision path asking what the experiment measures, leading to discard, merged-gene counting, or hierarchical allocation.](/article-images/illustrations/multi-mapping-reads-rna-seq-strategy-triage.webp)

The table below is the decision core of this article. Each option is explained on its own terms in the sections that follow, and the table summarizes when each one is the right choice.

| Option | How it works | Best for | Main risk | Representative tools |
|--|--|--|--|--|
| Discard multi-mappers | Count only uniquely aligned reads | Differential expression of unique protein-coding genes | Undercounts duplicated genes, loses information, biases allele-specific analysis | htseq-count, featureCounts default mode |
| Uniform fractional assignment | Split each multi-mapping read equally across its alignments | Quick estimates, simple pipelines | Assumes equal expression at all loci, which is false for differentially regulated paralogs | featureCounts fractional mode, custom scripts |
| Expectation-maximization allocation | Iteratively assign reads in proportion to current abundance estimates | Transcript-level, isoform, allele-specific, and duplicated-gene quantification | Depends on a good transcriptome reference, converges to a local solution | MMSEQ, EMASE, Salmon, kallisto |
| Merged-gene counting | Merge duplicated genes into a single feature and count ambiguous reads against it | Gene-level differential expression including duplicated gene families | Loses per-copy resolution | mmquant |
| Hierarchical allocation | Resolve ambiguities in order, genes first, then isoforms, then alleles | Allele-specific expression, total RNA-seq with overlapping biotypes | More complex to configure and interpret | EMASE, MGcount |
| Quasi-mapping and pseudoalignment | Report potential loci of origin without base-to-base alignment | Fast quantification of large transcriptomes | No alignment details, so mismatch-driven errors can be missed | RapMap, Salmon, kallisto |
| Report all alignments | Keep every alignment in the BAM and decide downstream | Exploratory analysis, custom pipelines | Large files, downstream tools must handle ambiguity explicitly | STAR with a raised multimapping limit, HISAT2 |

## Option 1: Discard Multi-Mappers

This is the default behavior of most counting tools. A read that aligns to more than one locus is simply not counted, and only uniquely aligned reads contribute to gene counts.

**Pros.** The counts are unambiguous and easy to interpret. Every count corresponds to a single genomic locus. The approach is fast, works with any counting tool, and requires no modeling. For protein-coding genes with unique exons, the loss of information is usually modest, and the resulting counts are directly comparable across samples processed the same way.

**Cons.** Any gene with duplicated sequence is undercounted, sometimes to zero. The bias is not random, it is systematic, so it can distort differential expression results in exactly the gene families where duplication matters. In allele-specific expression, discarding multi-mappers causes bias and substantial loss of information [2]. In a diploid transcriptome with variants, where multi-mappers can exceed 85% of reads, discarding them is rarely acceptable [2].

**When to use it.** Your biological question concerns protein-coding genes with unique sequence, your library is poly(A)-selected, and duplicated genes are not part of your hypothesis. This is the standard choice for routine bulk differential expression in well-annotated organisms such as human and mouse.

## Option 2: Uniform Fractional Assignment

Each multi-mapping read is split equally among all of its alignments. A read that aligns to three loci contributes one-third of a count to each.

**Pros.** No reads are discarded, so total counts are preserved. The approach is simple to implement and easy to explain. It is a reasonable default when you have no information about relative expression among the loci and you want a conservative, assumption-light estimate.

**Cons.** The equal-split assumption is wrong whenever paralogs are expressed at different levels. If gene A is highly expressed and its paralog gene B is silent, uniform splitting inflates B and deflates A. This bias is systematic and can create false differential expression signals. The mmquant paper identifies uniform distribution as one of the strategies that provides biased results [3].

**When to use it.** You need a quick estimate, you are working with an organism whose annotation is too poor for EM-based methods, or you want a baseline against which to compare a more sophisticated method.

## Option 3: Expectation-Maximization Allocation

EM-based methods treat each multi-mapping read as a probabilistic assignment. The algorithm starts with an initial guess of transcript abundances, computes the probability that each read came from each of its possible loci given those abundances, reallocates reads accordingly, and repeats until the estimates converge. Reads that are informative, meaning they align to only one of a set of candidate loci, effectively pin down the abundances that the ambiguous reads are then distributed toward.

MMSEQ is an early and influential example. It simultaneously estimates isoform expression and allelic imbalance in diploid organisms by modeling the expression of haplotype-specific isoforms. If the two parental isoform sequences are unknown, MMSEQ can reconstruct them individually. The method deconvolves the mapping of reads to multiple transcripts, takes non-uniform read generation into account, and works with paired-end reads [4].

EMASE applies the same idea to allele-specific expression with a hierarchical structure, described in its own section below. Salmon and kallisto apply EM-style inference to transcript-level quantification, using equivalence classes of reads that share the same set of candidate transcripts.

**Pros.** These methods use all of the reads, including multi-mappers, and they are statistically principled under their model. They can recover expression for duplicated genes and distinguish isoforms and alleles. They handle non-uniform read distribution and paired-end information [4].

**Cons.** They require a complete and accurate transcriptome reference. If a paralog is missing from the reference, reads from it will be misassigned to the paralogs that are present. EM converges to a local solution, and with very few informative reads the estimates can be unstable. Interpretation is also subtler, because a count is an estimate rather than an observation.

**When to use it.** Transcript-level quantification, isoform usage, allele-specific expression, duplicated gene families, or any setting where the multi-mapping fraction is high and the reference is trustworthy.

## Option 4: Merged-Gene Counting

Merged-gene counting takes a different approach to the duplicated-gene problem. Instead of trying to assign reads to individual copies, it detects that the genes are duplicated, merges them into a single feature, and counts ambiguous reads against the merged gene.

mmquant implements this idea. The tool detects that a read maps to different positions, concludes that the corresponding genes are duplicated, merges those genes, and computes counts of ambiguous reads based on both the input genes and the merged genes. It is designed as a drop-in replacement for htseq-count and featureCounts, which means it accepts the same kind of inputs and produces a count matrix that downstream differential expression tools can use directly [3].

**Pros.** The counts are unbiased for duplicated genes in the sense that the merged feature reflects the combined expression of the copies, without discarding reads or making an equal-split assumption. The workflow change is minimal, because the tool replaces an existing counter rather than adding a new analysis stage. Output is compatible with standard differential expression pipelines.

**Cons.** You lose per-copy resolution. If you need to know which copy of a duplicated gene is responding to treatment, merged-gene counting cannot answer that. The merged gene is also a new feature, so annotation-based summaries and gene-set analyses need care.

**When to use it.** Gene-level differential expression where duplicated gene families are present and matter, and you want to keep the pipeline simple. This is often the best compromise between discarding and full EM allocation for gene-level questions.

## Option 5: Hierarchical Allocation

Multi-mapping ambiguity has structure. A read can be ambiguous between genes, between isoforms of the same gene, or between alleles of the same isoform. Treating all of these ambiguities identically wastes that structure.

EMASE resolves ambiguities in a fixed order. It first resolves ambiguities among genes, then among isoforms, and lastly between alleles. This hierarchical allocation estimates total [gene expression](/blog/guides/gene-expression), isoform usage, and allele-specific expression in one integrated model. In simulations, EMASE outperforms standard allele-specific estimation methods, and it has been applied to F1 hybrid mice, where it revealed widespread allele-specific expression associated with cis-acting polymorphisms and a small number of parent-of-origin effects at known imprinted genes [5]. The published evaluation also found that methods aligning RNA-seq reads to a diploid transcriptome incorporating known genetic variants improve estimates of allele-specific expression and total gene expression compared with methods that use reference genome alignments [2].

MGcount applies hierarchical thinking to total RNA-seq, where coding and non-coding transcripts overlap in complicated ways. It first assigns reads hierarchically to small-RNA and long-RNA features to account for length disparity when transcripts overlap in the same genomic position. It then aggregates RNA products with similar sequences, where reads systematically multi-map, using a graph-based approach. The output is a transcriptomic count matrix compatible with downstream bulk and single-cell pipelines, and the software runs as a Python module or a single-file executable [6].

**Pros.** Hierarchical methods respect the biology of ambiguity, they use all reads, and they produce integrated estimates of gene, isoform, and allele expression. They are the most accurate option in simulations for allele-specific expression [2].

**Cons.** They require careful configuration, a variant-aware reference for allele-level work, and more computational and interpretive effort than simple counting. The output structure is richer, which means downstream analysis must be planned accordingly.

**When to use it.** Allele-specific expression in F1 hybrids or outbred populations, total RNA-seq with overlapping biotypes, and any analysis where the hierarchy of ambiguity matters.

## Option 6: Quasi-Mapping and Pseudoalignment

A separate family of tools avoids full base-to-base alignment altogether. RapMap introduced quasi-mapping, an approach that attempts only to report the potential loci of origin of a sequencing read, not the base-to-base alignment by which it derives from the reference. By using efficient data structures that exploit the special structure of shared sequence in transcriptomes, quasi-mapping provides highly accurate mapping information substantially faster than existing alignment tools [7]. RapMap underlies the mapping engine in Salmon.

Pseudoalignment, used by kallisto, similarly determines which transcripts a read is compatible with without computing a full alignment. RNACache takes a related approach, using MinHashing to map RNA-seq reads to transcriptomes quickly [8].

**Pros.** These methods are fast, which matters for large transcriptomes and large sample numbers. They naturally report all compatible loci, which feeds directly into EM-based quantification. They handle the multi-mapping burden of transcriptome alignment efficiently, a burden that can slow existing aligners considerably [7].

**Cons.** Because no base-to-base alignment is computed, mismatch-driven errors, such as those from nucleotide conversion methods, can be missed. Genome-level features, splicing junctions, and intronic reads are not handled by transcriptome-only mapping. For applications that need alignment details, a genome aligner is still required.

**When to use it.** Transcript-level quantification with a trusted reference annotation, or de novo settings where speed matters. Combine with a genome aligner if you also need intronic or junction-level information.

## Option 7: Report All Alignments and Decide Later

Some pipelines keep every alignment in the BAM file and defer the multi-mapping decision to the counting step. STAR controls this behavior with the parameter, outFilterMultimapNmax, which sets the maximum number of loci a read may map to and still be reported. The default value is 10, and raising it retains more multi-mapping reads in the output. This is standard STAR behavior described in its official documentation.

**Pros.** You retain maximum flexibility. The same BAM can support unique-only counting, fractional counting, and custom allocation schemes without re-running the alignment. For exploratory work, this is the most informative output.

**Cons.** BAM files grow larger, and downstream tools must handle ambiguity explicitly. A counting tool that silently ignores multi-mappers will produce the same result as Option 1, so the flexibility is only real if you use it.

**When to use it.** When you plan to try several handling strategies, or when you need to audit multi-mapping behavior before committing to a method.

## Worked Scenarios

### Scenario A: Routine Differential Expression of Protein-Coding Genes

You are comparing treated and control human poly(A)-selected libraries, and your hypothesis concerns protein-coding genes with unique exons. Align with STAR using default multimapping settings, then count with featureCounts in its default mode, which counts only uniquely aligned reads. Check the multimapping fraction in the STAR alignment log. If it is low, as it typically is for this library type, the unique-only counts are reliable. If a duplicated gene family appears in your results, re-run the count with mmquant to check whether the merged-gene count changes your conclusion [3].

### Scenario B: A Duplicated Gene Family Under Selection

You are studying a gene family that expanded by retrotransposition, and you need expression of the family as a whole. Individual copies cannot be distinguished by short reads, so per-copy estimates would be unreliable. Use mmquant, which merges the duplicated genes and counts ambiguous reads against the merged feature, giving an unbiased family-level count [3]. If you need per-copy resolution, longer reads are the better tool, since [long-read sequencing](/knowledge/bioinformatics/long-read-sequencing-technologies-pacbio-and-oxford-nanopore) outperforms short reads for complex transcript structures [9].

### Scenario C: Allele-Specific Expression in F1 Hybrids

You are measuring paternal versus maternal expression in hybrid mice. Build a diploid transcriptome that incorporates the known parental variants, align reads to it, and analyze with EMASE. Multi-mapping reads can exceed 85% of reads in this setting, so discarding them is not viable [2]. EMASE resolves gene, isoform, and allele ambiguity hierarchically, and this hierarchical allocation improves the accuracy of allele-specific estimates compared with methods that treat all multi-reads equivalently [2]. In F1 hybrid mice, this approach has revealed widespread allele-specific expression associated with cis-acting polymorphisms and parent-of-origin effects at imprinted genes [5].

### Scenario D: Total RNA-seq with Overlapping Biotypes

You are sequencing total RNA, which captures messenger RNAs, long non-coding RNAs, and small RNAs from the same library. Transcripts from different biotypes overlap in the same genomic regions, and some are present in high copy numbers, so ambiguous alignments are common [1]. Use MGcount, which assigns reads hierarchically to small-RNA and long-RNA features and then aggregates multi-mapping products with a graph-based approach, producing a count matrix compatible with bulk and single-cell downstream pipelines [6]. Note that library preparation itself changes the multi-mapping burden. Targeted enrichment of eukaryotic total RNA libraries increased multi-mapping reads to rRNAs compared with poly(A)-enriched libraries, so poly(A) selection is the cleaner choice for eukaryotic captures when rRNA ambiguity is a concern [10].

### Scenario E: Single-Cell RNA-seq

You are processing [10x Genomics](/blog/guides/10x-genomics) data and need gene-level counts per cell. STARsolo, built into STAR, quantifies single-cell and single-nucleus RNA-seq data and accounts for multi-gene reads, which are necessary to detect certain classes of biologically important genes. On simulated data that closely resembles realistic single-cell RNA-seq, STARsolo is highly accurate and significantly outperforms pseudoalignment-to-transcriptome tools, and it replicates the results of CellRanger while running considerably faster [11].

### Scenario F: Dual RNA-seq and Multi-Species Experiments

![Flowchart showing adapter trimming, pathogen-first mapping, then host mapping of remaining reads.](/article-images/illustrations/multi-mapping-reads-rna-seq-dual-rna-seq-order.webp)

You are sequencing a host-pathogen system. Mapping order matters. When adapter-trimmed reads are first mapped to the pathogen genome, more reads align to the pathogen genome than when unmapped reads are derived from the traditional host-first mapping approach, because pathogen-first mapping prevents the misalignment of shorter pathogen reads to the host genome [12]. For multi-species pooling, a concatenated multi-species index can work well. In a three-species plant pooling experiment using a concatenated coding-sequence index, every pure library was at least 99.8% correctly assigned and cross-species mis-mapping was at most 0.2%, showing that pooling does not compromise species-level quantification when the index is built carefully [13].

## Practical Commands and Parameters

The commands below are minimal working examples. Adjust paths and parameters to your data.

**Alignment with STAR, retaining multi-mappers for downstream decisions:**

```
STAR --runThreadN 8 \
     --genomeDir star_index \
     --readFilesIn sample_R1.fastq.gz sample_R2.fastq.gz \
     --readFilesCommand zcat \
     --outFileNamePrefix sample_ \
     --outSAMtype BAM SortedByCoordinate \
     --outFilterMultimapNmax 10
```

The default, outFilterMultimapNmax of 10 is standard STAR behavior. Inspect the file sample_Log.final.out for the multimapping read percentage before choosing a counting strategy.

**Unique-only counting with featureCounts:**

```
featureCounts -p --countReadPairs \
     -T 8 -t exon -g gene_id \
     -a annotation.gtf \
     -o sample_counts.txt \
     sample_Aligned.sortedByCoord.out.bam
```

In default mode, featureCounts does not count multi-mapping reads. Adding the -M flag changes this so that every alignment of a multi-mapping read is counted, which is a form of over-counting rather than allocation. Choose one mode deliberately and document it.

**Merged-gene counting with mmquant:**

mmquant is a drop-in replacement for htseq-count and featureCounts, so it accepts the same kind of annotation and alignment inputs and produces a count matrix for the same downstream tools [3]. Run it with your GTF annotation and your BAM file, and it will merge duplicated genes automatically and count ambiguous reads against the merged features.

**Transcript-level quantification with Salmon:**

```
salmon index -t transcripts.fa -i salmon_index
salmon quant -i salmon_index -l A \
     -1 sample_R1.fastq.gz -2 sample_R2.fastq.gz \
     -p 8 --validateMappings -o sample_salmon
```

Salmon uses quasi-mapping, introduced with RapMap, to report potential loci of origin without base-to-base alignment, which is what makes it fast on large transcriptomes [7].

**Allele-specific expression with EMASE:**

The EMASE workflow aligns reads to a diploid transcriptome built from the two parental haplotypes with known variants, then applies hierarchical expectation-maximization to allocate reads among genes, isoforms, and alleles in that order [2]. The software is freely available under a GNU license at the churchill-lab repository on GitHub [5].

**Total RNA-seq quantification with MGcount:**

MGcount runs as a Python module or a single-file executable program, takes the alignment and the annotation, and outputs a transcriptomic count matrix compatible with bulk and single-cell downstream pipelines, along with the graphs that model repeated transcript structures for different biotypes [6].

## Troubleshooting Multi-Mapping Problems

Multi-mapping issues are among the most common sources of an RNA seq error that quietly corrupts downstream results. The symptoms, causes, and fixes below cover the cases that matter most.

**Symptom: low unique mapping rate.** Check the aligner log first. A high multimapping fraction often means rRNA or other repetitive RNA contamination, which is common in total RNA libraries. Targeted enrichment of eukaryotic total RNA libraries increases multi-mapping reads to rRNAs compared with poly(A)-enriched libraries, so prefer poly(A) selection when rRNA ambiguity is a concern [10]. If contamination is not the cause, check that the reference genome matches your sample. Mapping reads to a non-native reference genome, for example mapping all strains to a common reference in a multi-strain bacterial comparison, produces false positives in differential expression analysis, and methods exist to identify the regions most affected by non-native alignments [14].

**Symptom: mismatch-containing reads fail to map or map incorrectly.** Nucleotide conversion RNA-seq methods, such as metabolic RNA labeling and RNA [bisulfite sequencing](/knowledge/molecular-biology/bisulfite-sequencing), interrogate chemical RNA modifications and produce reads containing systematic mismatches. Biases in mapping these reads remain poorly understood, and simulation frameworks such as splice_sim can measure mapping accuracies of spliced-read mappers under realistic conversion conditions and help derive strategies to prevent interpretation biases [15].

**Symptom: pathogen reads disappear in dual RNA-seq.** Mapping order is the usual culprit. Map adapter-trimmed reads to the pathogen genome first, then map the remainder to the host. This captures more pathogen reads and prevents the misalignment of shorter pathogen reads to the host genome [12].

**Symptom: inconsistent counts across pipelines.** Different tools handle multi-mappers differently, so counts for duplicated genes will differ between a unique-only counter, a fractional counter, and an EM-based quantifier. Pick one strategy, document it, and use it for every sample in the study. If duplicated genes matter, run mmquant as a cross-check [3].

**Symptom: small RNA results look noisy.** Small RNA analysis has its own mapping conventions. A benchmarked human small RNA pipeline recommends trimming with a lower length bound of 15 and an upper bound of read length minus 40% of the adapter length, mapping with bowtie allowing one mismatch, and filtering features by a mean threshold above 5 before differential expression analysis [16]. Small RNA genes are short and often multi-copy, so the biotype-specific multi-mapping behavior described in the review of duplicated sequences applies strongly here [1].

**Symptom: the aligner itself seems to be the problem.** Aligner choice matters, and no single aligner wins everywhere. A benchmark of 17 aligners on simulated and empirical data found that each tool was best in specific conditions. For Ion Torrent single-end RNA-seq samples, CLC and BWA-MEM reached the best results in efficiency, accuracy, duplication rate, saturation profile, and running time. For Illumina paired-end [transcriptomics](/knowledge/bioinformatics/modern-transcriptomics-bulk-single-cell-spatial) data, Novoalign excelled in accuracy and saturation analyses, alongside CLC [17]. If your mapping rate is unexpectedly low, testing a second aligner is a reasonable diagnostic step.

**Symptom: quantification of highly expressed genes looks off.** Short-read sequencing retains superior quantification accuracy for highly expressed genes and stronger concordance with microarray data, while long reads excel at complex [alternative splicing](/blog/guides/alternative-splicing) and novel isoforms [9]. If your question is quantification of abundant, well-annotated genes, short reads with a well-chosen multi-mapping strategy remain the safer choice.

## Common Mistakes and Limitations

**Mixing multi-mapping strategies within a study.** Counts produced under different assumptions are not comparable. If some samples were counted with unique-only mode and others with a fractional or EM method, the differential expression model is fitting a mixture of definitions. Standardize the strategy across all samples.

**Assuming the default is unbiased.** The default of most counting tools is to discard multi-mappers. That default is fine for unique protein-coding genes and wrong for duplicated genes, allele-specific analysis, and short RNA biotypes [1], [2]. Know what your default is doing.

**Treating uniform splitting as a neutral choice.** Splitting a read equally across its loci is an assumption, not a neutral act. It is wrong whenever paralogs differ in expression, which is the common case.

**Using a reference genome from a different strain or species.** Non-native alignments create false positives in differential expression, and the affected regions can be identified systematically [14]. Match the reference to your sample, or build a sample-specific or variant-aware reference.

**Ignoring the multimapping fraction in the aligner log.** The STAR log reports the multimapping percentage for every run. A sudden change in this number between sample batches signals a library preparation or contamination problem, and it is one of the cheapest quality checks available.

**Forgetting that multi-mapping differs by biotype.** Messenger RNAs and long non-coding RNAs share less sequence similarity with other genes than biotypes encoding shorter RNAs [1]. A pipeline tuned for protein-coding genes can be badly mis-calibrated for small RNAs, and the reverse is also true.

**Skipping validation.** Simulation frameworks let you test your pipeline before committing to it. Splice_sim simulates nucleotide conversion RNA-seq datasets and measures mapping accuracy per genomic annotation [15], and AsaruSim simulates single-cell long-read Nanopore datasets, including PCR amplification and sequencing errors, to evaluate isoform detection methods [18]. Running your pipeline on simulated data with known truth is the most direct way to detect multi-mapping bias.

**One limitation worth stating plainly.** Every dataset is different, and the best strategy depends on your genome, your library type, your read length, and your biological question. Validate the chosen approach on your own data, ideally with spike-ins or simulations, before committing to it for a full study.

## Frequently Asked Questions

### What are multi-mapping reads in RNA-seq?

They are sequencing reads that align equally well to more than one location in the reference genome or transcriptome, usually because the loci share duplicated or repeated sequence. The aligner reports them as ambiguous, and every downstream tool must decide how to count them.

### Does discarding multi-mapping reads bias my results?

Yes, it systematically undercounts duplicated genes and biases allele-specific expression analysis. Weighted allocation methods outperform discarding, and hierarchical allocation improves the estimates further [2].

### Which aligner settings control multi-mapping in STAR?

The parameter, outFilterMultimapNmax sets the maximum number of loci a read may map to and still be reported, with a default of 10 in standard STAR documentation. Raising it retains more multi-mapping reads in the BAM.

### Should I use EM quantification or simple counting for gene-level differential expression?

For unique protein-coding genes, simple counting is fine. For duplicated gene families, use a merged-gene counter such as mmquant, which is a drop-in replacement for htseq-count and featureCounts [3]. For transcript-level or allele-level questions, use an EM-based method.

### How do multi-mapping reads affect allele-specific expression?

They can dominate the analysis. In a diploid transcriptome with known variants, multi-mapping reads can comprise more than 85% of reads, so discarding them causes bias and substantial loss of information, while hierarchical allocation resolves gene, isoform, and allele ambiguity in order and improves accuracy [2].

### Why do my reads multi-map more in total RNA-seq than in poly(A) RNA-seq?

Total RNA-seq captures short RNA genes and repetitive RNA families that poly(A) selection largely excludes, and these biotypes share more sequence similarity and are present in high copy numbers [1]. Targeted enrichment of eukaryotic total RNA libraries also increases multi-mapping reads to rRNAs compared with poly(A)-enriched libraries [10].

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "FAQPage",
  "mainEntity": [
    {
      "@type": "Question",
      "name": "What are multi-mapping reads in RNA-seq?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "They are sequencing reads that align equally well to more than one location in the reference genome or transcriptome, usually because the loci share duplicated or repeated sequence. The aligner reports them as ambiguous, and every downstream tool must decide how to count them."
      }
    },
    {
      "@type": "Question",
      "name": "Does discarding multi-mapping reads bias my results?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Yes, it systematically undercounts duplicated genes and biases allele-specific expression analysis. Weighted allocation methods outperform discarding, and hierarchical allocation improves the estimates further."
      }
    },
    {
      "@type": "Question",
      "name": "Which aligner settings control multi-mapping in STAR?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "The parameter, outFilterMultimapNmax sets the maximum number of loci a read may map to and still be reported, with a default of 10 in standard STAR documentation. Raising it retains more multi-mapping reads in the BAM."
      }
    },
    {
      "@type": "Question",
      "name": "Should I use EM quantification or simple counting for gene-level differential expression?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "For unique protein-coding genes, simple counting is fine. For duplicated gene families, use a merged-gene counter such as mmquant, which is a drop-in replacement for htseq-count and featureCounts. For transcript-level or allele-level questions, use an EM-based method."
      }
    },
    {
      "@type": "Question",
      "name": "How do multi-mapping reads affect allele-specific expression?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "They can dominate the analysis. In a diploid transcriptome with known variants, multi-mapping reads can comprise more than 85% of reads, so discarding them causes bias and substantial loss of information, while hierarchical allocation resolves gene, isoform, and allele ambiguity in order and improves accuracy."
      }
    },
    {
      "@type": "Question",
      "name": "Why do my reads multi-map more in total RNA-seq than in poly(A) RNA-seq?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Total RNA-seq captures short RNA genes and repetitive RNA families that poly(A) selection largely excludes, and these biotypes share more sequence similarity and are present in high copy numbers. Targeted enrichment of eukaryotic total RNA libraries also increases multi-mapping reads to rRNAs compared with poly(A)-enriched libraries."
      }
    }
  ]
}
</script>

## Sources

1. [Handling multi-mapped reads in RNA-seq](https://pubmed.ncbi.nlm.nih.gov/32637053/)
2. [Hierarchical analysis of RNA-seq reads improves the accuracy of allele-specific expression.](https://pubmed.ncbi.nlm.nih.gov/29444201/)
3. [mmquant: how to count multi-mapping reads?](https://pubmed.ncbi.nlm.nih.gov/28915787/)
4. [Haplotype and isoform specific expression estimation using multi-mapping RNA-seq reads.](https://pubmed.ncbi.nlm.nih.gov/21310039/)
5. [Hierarchical Analysis of Multi-mapping RNA-Seq Reads Improves the Accuracy of Allele-specific Expression](https://doi.org/10.1101/166900)
6. [MGcount: a total RNA-seq quantification tool to address multi-mapping and multi-overlapping alignments ambiguity in non-coding transcripts.](https://pubmed.ncbi.nlm.nih.gov/35030988/)
7. [RapMap: a rapid, sensitive and accurate tool for mapping RNA-seq reads to transcriptomes.](https://pubmed.ncbi.nlm.nih.gov/27307617/)
8. [RNACache: Fast Mapping of RNA-Seq Reads to Transcriptomes Using MinHashing](https://doi.org/10.1007/978-3-030-77961-0_31)
9. [Systematic evaluation of long- and short-read RNA-seq for human peripheral blood](https://pubmed.ncbi.nlm.nih.gov/41635783/)
10. [Targeted enrichment outperforms other enrichment techniques and enables more multi-species RNA-Seq analyses.](https://pubmed.ncbi.nlm.nih.gov/30190541/)
11. [STARsolo: accurate, fast and versatile mapping/quantification of single-cell and single-nucleus RNA-seq data](https://doi.org/10.1101/2021.05.05.442755)
12. [Optimal Dual RNA-Seq Mapping for Accurate Pathogen Detection in Complex Eukaryotic Hosts](https://pubmed.ncbi.nlm.nih.gov/39959292/)
13. [A biological spike-in enables cost-effective multi-species RNA-seq and reveals global transcriptional collapse under dark stress](https://doi.org/10.64898/2026.06.16.732598)
14. [The quantitative impact of read mapping to non-native reference genomes in comparative RNA-Seq studies](https://pubmed.ncbi.nlm.nih.gov/28700635/)
15. [Splice_sim: a nucleotide conversion-enabled RNA-seq simulation and evaluation framework](https://pubmed.ncbi.nlm.nih.gov/38918865/)
16. [Approaches for sRNA Analysis of Human RNA-Seq Data: Comparison, Benchmarking](https://pubmed.ncbi.nlm.nih.gov/36835604/)
17. [New evaluation methods of read mapping by 17 aligners on simulated and empirical NGS data: an updated comparison of DNA- and RNA-Seq data from Illumina and Ion Torrent technologies](https://pubmed.ncbi.nlm.nih.gov/34155424/)
18. [AsaruSim: a single-cell and spatial RNA-Seq Nanopore long-reads simulation workflow](https://pubmed.ncbi.nlm.nih.gov/39985444/)