When high-throughput sequencing instruments analyze a biological sample, they do not read entire RNA transcripts from end to end. Instead, they generate millions of short fragments called "reads." The foundational task of transcriptomics is to map these short reads back to their coordinates on a reference genome or transcriptome to quantify gene expression.
However, this mapping process is frequently ambiguous. A single short read can often align perfectly to multiple different locations or transcript variants. This phenomenon, known as multi-mapping, stems from two fundamental biological realities:
- Alternative Splicing: A single gene can produce multiple distinct mRNA isoforms by mixing and matching exons. Because different isoforms share identical genomic segments, a short read originating from a shared exon cannot be uniquely assigned to a specific variant (Shai et al., 2004; Bhattacharjee et al., 2012).
- Genomic Repetitiveness: Large portions of genomes consist of highly repetitive sequences, such as transposable elements (Navarro et al., 2019; Groza et al., 2024) or low-complexity terminal untranslated regions (UTRs) (Derr et al., 2016). Reads from these regions match multiple distinct genes or loci with equal probability.
Du et al. framed this challenge as an inherent property of the "transcriptome information landscape": because sequencing reads are significantly shorter than the full-length transcripts they represent, unique "mappability" drops in regions of high sequence similarity (Du et al., 2012; Hamaguchi et al., 2021).
The Cost of Naive Solutions
Traditional sequencing pipelines typically handle multi-mapped reads in one of two ways: they either discard them entirely or assign them randomly to one of their potential targets. As Kahles et al. note, both approaches introduce severe errors (Kahles et al., 2015). Discarding ambiguous reads drastically reduces statistical power—destroying up to 90% of available data in specialized applications like ribosome profiling (Cui et al., 2019). Conversely, random assignment introduces stochastic noise, leading to false positives and skewed expression profiles.
To unlock accurate quantification, researchers treat read mapping not as a simple string-matching problem, but as a statistical deconvolution problem.
The Deconvolution Intuition
Imagine trying to determine the individual volumes of two overlapping water sources flowing into a single pipe. You cannot measure them directly from the combined flow. However, if you know the unique chemical signature of each source, and you can measure the flow at points where they do not overlap, you can mathematically infer the missing proportions.
In transcriptomics, unique exons or sequence variations act as these unblended reference points. If an ambiguous read maps to both Isoform A and Isoform B, algorithms look at the regions unique to each isoform to resolve the tie. If the unique regions of Isoform A show high read coverage while the unique regions of Isoform B have none, the ambiguous read is statistically allocated to Isoform A.
Isoform A: [=== Exon 1 ===][=== Exon 2 ===]
Isoform B: [=== Exon 2 ===][=== Exon 3 ===]
^
Ambiguous Read maps here
Expectation-Maximization (EM)
This iterative logic is elegantly captured by the Expectation-Maximization (EM) algorithm, a cornerstone of modern isoform quantification pipelines (Deng et al., 2011).
- Expectation Step (E-step): The algorithm uses current estimates of transcript abundances to calculate the probability that an ambiguous read originated from each candidate transcript. The read is then fractionally divided among those transcripts based on those probabilities.
- Maximization Step (M-step): The algorithm updates the global transcript abundance estimates by summing up all the uniquely mapped reads and the fractionally assigned ambiguous reads.
This cycle repeats until the abundance estimates stabilize, effectively using the global context of the dataset to resolve local mapping uncertainties.
Optimization and Structural Constraints
Beyond EM, researchers use constrained optimization frameworks to enforce biological realism. For example, Bohnert et al. developed rQuant, which applies quadratic programming to model non-uniform sequencing coverage biases across genomic positions, preventing systematic experimental artifacts from distorting the deconvolution (Transcript quantification with RNA-Seq data).
When sequencing depth is low, the deconvolution problem can become mathematically ill-posed, meaning multiple distinct combinations of transcript abundances could explain the observed reads. To solve this, Bernard et al. introduced a convex formulation that utilizes a "group-lasso" penalty (A convex formulation for joint RNA isoform detection and quantification from multiple RNA-seq samples). This mathematical constraint leverages the biological intuition that related samples (such as replicates or tissues from the same organism) typically express the same core subset of isoforms, allowing the algorithm to borrow statistical strength across samples to resolve ambiguities.
Leveraging Local and Structural Context
Other approaches bypass complex global optimization by exploiting local physical properties of sequencing data.
Coverage Uniformity
Because fragmentation during sequencing is largely stochastic, reads originating from a true transcript should ideally be distributed uniformly across its sequence. The MMR tool, developed by Kahles et al., exploits this property by evaluating all potential alignment locations for a multi-mapped read and assigning it to the locus that yields the most uniform local coverage profile (MMR: A Tool for Read Multi-Mapper Resolution).
Graph Genomes
Standard linear reference genomes struggle to represent highly repetitive structural variations, such as transposable element insertions, leading to systemic misalignments. Groza et al. demonstrated that by replacing linear references with graph genomes—which represent sequence variations as alternative paths in a network—algorithms can map reads through complex, repetitive regions with significantly higher topological precision (A unified framework to analyze transposable element insertion polymorphisms using graph genomes).
Propagating Uncertainty Downstream
A crucial paradigm shift in this field is the realization that statistical deconvolution rarely yields absolute certainty. Even the most sophisticated EM or Bayesian model cannot completely eliminate the ambiguity inherent in short-read data.
Historically, pipelines would arrive at a "best guess" for transcript abundance and pass that single value to downstream analyses, such as differential expression (DE) testing. This practice masks the underlying mapping uncertainty, leading to high false-positive rates for genes with complex alternative splicing patterns.
To prevent this, frameworks like BDSeq, developed by Zhang et al., explicitly calculate the measurement uncertainty during the quantification phase and propagate it directly into the DE testing framework (Detecting differential expression from RNA-seq data with expression measurement uncertainty). By treating transcript abundance as a probability distribution rather than a fixed point estimate, downstream statistical tests can automatically discount highly ambiguous signals, ensuring that discovered biological insights are grounded in robust, uniquely identifiable sequence data.
Key Foundational Work
Transcript quantification with RNA-Seq data (2009) — contributed 1 concept elements to this direction.
Probabilistic Inference of Alternative Splicing Events in Microarray Data (2004) — contributed 1 concept elements to this direction.