RNA Sequencing Tables: Raw Counts and Transcript Abundance

What an RNA-Sequencing Table Records

RNA sequencing (RNA-seq) produces reads from RNA molecules in a sample. After processing and comparison with a reference, reads or fragments can be assigned to genes or transcripts. A resulting table is often displayed with molecular features as rows and samples as columns. The entry at their intersection summarizes evidence for that feature in that sample; its precise meaning depends on how the table was made.

The most important distinction is between a count table and an abundance table. A raw count records how many reads or fragments were assigned to a feature, or an estimate of that number. An abundance estimate adjusts counts to represent the feature's estimated share of sequenced RNA, often accounting for feature length as well as the number of reads or fragments sequenced. Neither is simply a direct count of molecules in the original tissue.

A read is a sequence read produced by the instrument. In paired-end sequencing, two reads can come from opposite ends of one sequenced fragment. Consequently, a pipeline may report read counts or fragment counts. These are not interchangeable units: the same set of paired-end fragments may yield twice as many reads as fragments. Always check which unit a particular count table uses.

Gene-level and transcript-level tables answer questions at different levels of detail. A gene may produce multiple RNA transcripts. A gene-level entry summarizes evidence assigned to the gene, whereas a transcript-level entry distinguishes among the gene's individual transcripts. Which reads can be assigned to which feature depends on the reference annotation and on how well the reads distinguish similar sequences.

Reading Raw Counts

Consider a hypothetical raw gene-count table. The values are fragment counts; the samples are separate libraries. The gene identifiers are illustrative.

For Gene G1 in Sample A, the entry 120 means that 120 fragments were assigned to that gene under the specified processing and assignment rules. It does not mean that exactly 120 RNA molecules were present in the sample: one molecule can generate multiple sequenced fragments, and many molecules may not be sampled. The zero for Gene G3 in Sample A means no fragments were assigned there; it does not establish that the gene was biologically absent.

Raw counts also depend on how many fragments were sequenced. For example, the same count of 100 fragments can represent different proportions of a library if one library contains 10,000 assigned fragments and another contains 100,000. This is one reason a raw count is not itself a length- and library-adjusted abundance estimate. Later analysis may account for these differences, but the count table's entries remain counts.

Gene Counts and Transcript Counts

At transcript level, reads can be ambiguous. A read that matches sequence shared by two transcripts may provide evidence for both without identifying which transcript produced it. Quantification methods use the data and a model of read assignment to estimate transcript-level counts. These estimated counts can be fractional, even though an individual read or fragment cannot be divided into a fraction in the laboratory.

When transcript counts are available, a gene-level count may be formed by summing counts for its annotated transcripts, provided that those transcript-level values are defined on a compatible basis. This aggregation removes the distinction among the gene's transcripts. It does not recover that distinction later: two samples can have the same gene total but different transcript-level patterns.

Abundance Estimates: Length and Library Scaling

Longer transcripts provide more possible locations for reads to originate. For a given number of sequenced fragments, a long transcript can therefore receive more fragments than a short transcript simply because it offers more sequence to sample. Abundance estimates such as FPKM and TPM address this length effect and scale by the amount sequenced. They are not raw counts and should be interpreted using their stated units and definition.

A related measure, TPM (transcripts per million), first divides each feature's count by its length and then rescales those length-adjusted rates to sum to one million within the sample. Here “transcripts per million” names a relative abundance scale; it is not a direct count of one million physical RNA molecules.

This definition presumes that the counts and lengths refer to compatible features and that the count unit is consistent across the sum. Transcript quantification commonly uses an effective length: an adjustment for the portion of a transcript that can produce an observable fragment, given the sequencing and quantification setup. If a table reports TPM based on effective lengths, use those estimates rather than silently substituting a different length. The file or analysis documentation should identify the measure and its feature-length convention.

Why the Definitions Matter

The length adjustment and rescaling in TPM explain its within-sample total. Write \(r_i=c_i/L_i\) for each feature's length-adjusted rate and \(R=\sum_j r_j\) for their total. If \(R>0\), then:

Thus, subject to rounding and the feature set included, TPM values sum to one million. FPKM uses a different order of scaling: it adjusts each count by feature length and the total number of mapped reads or fragments. Its values are not defined to sum to one million. The two measures can therefore give different numerical values and should not be treated as interchangeable.

The table to use depends on the question. Raw counts preserve the count scale and are the appropriate input for methods that explicitly model counts; an abundance table is useful when the question calls for length-adjusted relative abundance estimates. A sample's TPM values describe relative shares within its quantified transcript pool. Because those shares depend on the other transcripts contributing to the total, TPM is not an absolute measure of RNA molecules per cell. Neither raw counts nor TPM alone establishes a causal effect or a biological difference between groups.

Diagnostics and Common Pitfalls

Before interpreting an RNA-seq table, check its row and column labels, feature level, and units. Confirm whether entries are read counts, fragment counts, estimated counts, TPM, FPKM, or another quantity; whether a reported length is transcript length or effective length; and which reference annotation was used. These details determine what a row and its value actually mean.

• Assuming every count must be an integer. Direct assignment counts are generally integer-valued, but quantification can report fractional estimated counts. Check the method before deciding that decimals are an error.
• Reading a count as a molecule number. A count is evidence sampled and assigned from a library, not a census of the original RNA molecules. Library preparation, sequencing depth, mapping, and assignment all affect it.
• Comparing raw counts as if they were abundance. A larger raw count can reflect a larger library or a longer feature as well as more relative RNA. Identify the scale before making a comparison.
• Calling TPM absolute abundance. TPM is rescaled to a fixed total within each sample. It describes relative abundance on that scale, not molecules per cell or tissue mass.
• Mixing feature definitions. A transcript-level value and a gene-level value do not describe the same feature. Different annotations or length conventions can also change estimates.

A table with unexpectedly many zeroes, implausible identifiers, or inconsistent units calls for checking the data-processing record before biological interpretation. A sum of TPM values far from one million may indicate that the table is rounded heavily, only a subset of features is shown, or the entries are not TPM as assumed. These checks flag possible mismatches; they do not by themselves identify the cause.