What Does RNA–Protein Discordance Mean?
A gene’s RNA abundance and its encoded protein abundance are related, but they are not interchangeable measurements. RNA is a snapshot of transcript abundance; protein reflects translation, degradation, trafficking, and other processes that may operate on different time scales. In CITE-seq, RNA counts and antibody-derived tag (ADT) counts also have different measurement processes. Therefore, a difference in their apparent patterns can be biologically informative, technical, or both.
The key is to define the comparison before interpreting it. A cell with detectable RNA and undetectable surface protein is not, by itself, evidence of unusual regulation. Nor is a negative correlation between RNA and protein sufficient to establish a biological mechanism. Interpretation depends on the gene and protein being matched correctly, the population and time point being compared, the measurement sensitivity of each modality, and the scale on which each result is summarized.
A useful first distinction is between within-cell discordance and between-condition discordance. The first asks whether cells with relatively more RNA also tend to have relatively more protein. The second asks whether a condition changes RNA and protein in the same direction or by similar amounts. These questions need not have the same answer: a condition can shift both modalities on average while cell-to-cell RNA and protein values are only weakly associated.
The measurements also need not be on a common scale. A log-normalized RNA value and a centered log-ratio (CLR)-normalized ADT value are not quantities that can be subtracted to obtain a biologically meaningful “RNA minus protein” effect. CLR values express relative abundance against a set of measured tags, whereas RNA normalization has a different reference and measurement process. Compare directions and patterns within each modality, and use a contrast defined on a scale appropriate to that modality.
Biological and Technical Explanations
Several biological processes can produce genuine discordance. Transcription may change before enough protein has been translated to alter abundance. Existing protein may persist after RNA declines. Translation efficiency can vary, and proteins may be processed, degraded, or transported to the cell surface at rates that differ from transcript production. For a surface marker, intracellular protein abundance and surface ADT signal are not identical: trafficking can affect how much protein is accessible to the antibody.
Technical explanations must be considered with equal care. RNA and ADT measurements differ in sensitivity and dynamic range; low transcript counts can be missed even when protein is readily detected, or a protein can be difficult to measure despite RNA being present. ADTs can include nonspecific or ambient background, while low signal may be compressed by background correction. Library depth, sample handling, antibody performance, and technical batch can change apparent detection. The earlier material on background components in CITE-seq protein counts is relevant when assessing whether low-level ADT signal is specific.
Before interpreting a mismatch, establish that the RNA feature and antibody target really correspond. Check gene and protein naming, antibody specificity, whether the antibody recognizes the relevant isoform or epitope, and whether the assay measures surface-accessible protein. A mismatch in target definition cannot be repaired by statistical analysis.
Worked Example: A Delayed Surface-Protein Response
Hypothetical data. In one cell population, investigators measure a transcript and its surface protein at 0, 2, 4, 8, and 12 hours after a stimulus. Within each modality, values are expressed as log2 fold changes from that modality’s own baseline.
| Hours after stimulus | RNA log2 fold change | Surface-protein log2 fold change |
|---|---|---|
| 0 | 0.0 | 0.0 |
| 2 | 1.0 | 0.2 |
| 4 | 2.0 | 0.8 |
| 8 | 1.0 | 1.5 |
| 12 | 0.0 | 1.1 |
The RNA response reaches its observed maximum at 4 hours: its value rises from 0.0 to 2.0 log2 fold-change units, equivalent to a fourfold change on the RNA scale. The surface-protein response reaches its observed maximum at 8 hours, at 1.5 log2 fold-change units on the protein scale. The later protein peak is consistent with a delayed protein response following transcription.
This pattern is suggestive, not proof of a specific mechanism. The scales are separately referenced to each modality’s baseline, so the values 2.0 and 1.5 should not be subtracted or interpreted as a quantitative RNA–protein gap. The time points are also only two hours apart around the RNA peak and four hours apart around the protein peak; the true peak times could lie between observations. A denser time course, consistent cell-population definitions, and replication would help assess whether the lag is reproducible rather than a sampling fluctuation or batch effect.
Compare Like With Like
For condition comparisons, a defensible approach is to estimate a condition contrast separately for RNA and protein, using the same biological units where possible. In a multi-donor study, for example, compute a donor-level contrast for each modality and compare the signs and uncertainty of those contrasts. Donors—not the many cells from one donor—are the independent biological replicates for a population-level condition effect. Paired conditions should be paired within donor.
A paired \(t\) interval for donor-level differences assumes that the donors are independent of one another and that the differences are approximately normally distributed in the target population. With very few donors, the normality assumption is difficult to assess and the interval can be sensitive to individual values. The contrast must also be defined on a scale with a clear interpretation. A protein CLR contrast is a difference in relative, composition-dependent abundance; it is not an absolute change in molecules per cell.
Worked Example: Contrasting Paired RNA and Protein Changes
Hypothetical data. Six independent donors each contribute cells under control and treatment conditions. For one matched gene–protein pair, the donor-level RNA contrasts are treatment minus control in log2 normalized RNA units: 0.6, 0.8, 0.7, 1.0, 0.9, and 0.8. The corresponding protein contrasts are treatment minus control in CLR units for the same ADT: −0.20, −0.10, −0.30, 0.00, −0.20, and −0.10. The conditions are paired within donor; the six donors are assumed independent. For this illustrative calculation, assume the donor-level differences in each modality are approximately normally distributed.
For RNA, the mean paired contrast is \( \bar{x}_{RNA}=(0.6+0.8+0.7+1.0+0.9+0.8)/6=0.800 \) log2 units. The deviations from the mean are −0.2, 0, −0.1, 0.2, 0.1, and 0, whose squared values sum to 0.10. Thus \(s_{RNA}=\sqrt{0.10/5}=0.1414\) and \(SE_{RNA}=0.1414/\sqrt{6}=0.05774\) log2 units. With \(df=5\), \(t^*=2.571\) for a two-sided 95% interval, giving \(0.800\pm2.571(0.05774)=[0.652,0.948]\) log2 units.
For protein, the mean paired contrast is \( \bar{x}_{P}=(-0.20-0.10-0.30+0.00-0.20-0.10)/6=-0.150 \) CLR units. The deviations are −0.05, 0.05, −0.15, 0.15, −0.05, and 0.05. Their squared values sum to \(0.0025+0.0025+0.0225+0.0225+0.0025+0.0025=0.055\). Therefore, \(s_P=\sqrt{0.055/5}=0.1049\) CLR units and \(SE_P=0.1049/\sqrt{6}=0.0428\) CLR units. The 95% interval is \(-0.150\pm2.571(0.0428)=[-0.260,-0.040]\) CLR units.
Both intervals exclude zero under the stated assumptions: RNA increases on its log2 normalized scale, while the ADT has a negative treatment contrast on its CLR scale. This is evidence of opposite estimated condition responses in these data, not evidence that the treatment caused a particular molecular mechanism. Because the protein contrast is CLR-based, its interpretation is relative to the measured tag composition. The RNA and protein intervals are not directly comparable in magnitude, and six donors provide limited information about the distribution of donor effects.
Diagnostics and Failure Modes
A discordant pattern is more credible when it persists under reasonable checks and is not concentrated in a technical subset. Examine modality-specific detection rates and distributions by donor, condition, and batch; check whether low counts or high background drive the apparent mismatch; and verify that the same cell population is being compared in both modalities. Where possible, inspect summaries across biological replicates rather than relying on a large number of cells from a single sample. A per-cell pattern can be dominated by differences in cell state or sample composition, even when the question concerns a condition effect.
A useful diagnostic is to ask what result would be expected if one modality were simply less sensitive. If RNA detection is rare but protein detection is common, the transcript may be near the RNA assay’s detection limit; apparent RNA absence does not establish biological absence. Conversely, broad ADT signal that tracks background or batch more strongly than cell identity may not represent specific protein abundance. Check whether the apparent disagreement changes when low-quality cells, problematic tags, or a technical batch are examined separately. Such checks help locate a problem but do not justify selectively removing observations solely because they disagree with the biological hypothesis.
Do not treat a correlation near zero as a complete discordance analysis. A mixture of cell states can obscure a within-state relationship, and a lagged biological response can weaken same-time correlation. Likewise, a positive correlation does not show that RNA quantitatively predicts protein or that one modality validates the other. Correlation summarizes association on the chosen scales; its meaning depends on population, time, measurement sensitivity, and design.
Worked Example: Unequal Detection Sensitivity
Hypothetical data. In 100 cells from a defined population, a transcript is detected in 18 cells, whereas the matching ADT is above the chosen positivity threshold in 82 cells. A shallow sequencing result might tempt an analyst to conclude that most protein-positive cells lack the transcript.
The observed detection proportions are \(18/100=0.18\) for RNA and \(82/100=0.82\) for ADT. Their difference is \(0.82-0.18=0.64\), or 64 percentage points. This arithmetic describes detection under the two assays; it does not estimate a 64-point biological difference in abundance, because the detection thresholds and sensitivities are not shared scales.
Suppose a sensitivity assessment using a more sensitive targeted RNA assay detects the transcript in 38 of 50 comparable cells, or \(38/50=0.76\). This makes a sensitivity explanation plausible, but the two detection proportions should not be treated as an unbiased before-and-after estimate unless the assays, cells, and sampling are appropriately matched. The targeted assay may use different cells or conditions. The conclusion is therefore limited: the original RNA non-detection rate cannot, on its own, establish true transcript absence. Further paired validation under comparable conditions is needed to distinguish assay sensitivity from biological regulation.
Choosing an Interpretation and a Follow-Up
The main alternatives are not competing statistical tests so much as competing explanations and summaries. Comparing raw counts is generally unhelpful because counts depend on modality-specific depth and capture. Comparing normalized patterns within each modality can reveal direction and timing, while donor-level condition contrasts address reproducibility at the biological-replicate level. A time course is especially informative for a proposed lag; a targeted assay or independent protein measurement can assess sensitivity or specificity. Each strategy answers a different question, and none turns an observational mismatch into causal evidence by itself.
An interpretation is stronger when the target correspondence is clear, the pattern recurs across independent donors, technical factors do not explain it, and the proposed explanation predicts additional observations. For example, a lag hypothesis predicts a reproducible ordering of RNA and protein responses over time. A post-transcriptional explanation may be plausible when RNA changes but protein does not, but it still needs evidence beyond that mismatch. Report the observed scales, time points, cell population, detection criteria, and uncertainty so readers can distinguish a biological interpretation from a measurement artifact.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- Why is it generally invalid to subtract a normalized RNA value from a CLR-normalized ADT value?
- A transcript peaks at 2 hours and its surface protein peaks at 8 hours after stimulation. What interpretation is plausible, and what would strengthen it?
- In a paired donor study, why should the condition contrast be computed within donor rather than treating all cells as independent replicates?
- An ADT is detected in many cells while its transcript is detected in few. Name two technical or biological possibilities that should be considered before concluding the transcript is absent.
Solutions. (1) The values have different references, units, and measurement processes; their difference has no direct biological interpretation. (2) A delayed protein response is plausible, but a denser, replicated time course with consistent cell populations would help establish reproducible timing. (3) Cells from one donor share biological context and are not independent donor replicates; within-donor contrasts preserve pairing and donor is the relevant unit for a population-level condition effect. (4) The RNA may be below the transcript assay’s detection sensitivity, or the ADT signal may reflect background or nonspecific binding; genuine persistence of surface protein despite low RNA is another biological possibility.