Sequence the Right Animals: Phenotype-Guided RNA-seq Allocation in Small Physiology Trials

Stress and nutrition experiments in plants, fish and livestock routinely measure a dozen or more physiological traits on every individual. These include fast-responding biochemical markers such as antioxidant enzyme activity and lipid damage, and slow structural outcomes such as height, biomass or condition factor. RNA sequencing, which reads out the activity of roughly 20,000 genes at once, is then run on only three to six individuals per treatment, usually picked at random. The transcriptome is analysed only as treated versus control. That design can say which genes respond to the treatment. It is nearly powerless to say which genes track the physiology, because with a handful of random replicates the individual variation that would link a gene to an enzyme level or a growth outcome is small and confounded with the treatment.

We propose to treat the cheap physiological panel, which is measured on everyone, as the information that decides who gets sequenced. Within each treatment group, we choose individuals whose leading (fast biochemical) and trailing (slow structural) profiles span the widest, least correlated range. We then model each gene's expression as a response to those two physiological axes, with treatment held fixed. Because the selection depends only on variables that enter the model as predictors, the gene-on-physiology slopes stay unbiased. Their variance falls with the spread of the predictors, so the same sequencing budget buys more power. The claim is concrete and testable: at a fixed number of sequenced samples, phenotype-guided allocation with a two-axis joint test recovers more gene–physiology links at calibrated false discovery rate than random replicates with standard per-trait analysis.

Core Claim

The power of a regression slope depends on how spread out the predictor is. For a gene $g$ regressed on a physiological score $x$ with treatment as a covariate, the slope's sampling variance is approximately

$$\mathrm{Var}(\hat\beta_g) \approx \frac{\sigma_g^2}{n\,\mathrm{Var}_w(x)},$$

where $\sigma_g^2$ is the gene's residual variance, $n$ is the number of sequenced samples, and $\mathrm{Var}_w(x)$ is the within-treatment variance of $x$ among the sequenced individuals. This is the ordinary least-squares variance formula. Random selection fixes $\mathrm{Var}_w(x)$ at its population value. Choosing individuals from the tails of the within-treatment distribution can roughly double or triple it, which is equivalent to sequencing two to three times as many random animals.

With two predictors, a leading score $L$ and a trailing score $T$, the relevant quantity is the determinant of their within-treatment covariance among the sequenced individuals. So we select by a D-optimal criterion, which is the classical experimental-design rule that maximizes that determinant. The criterion also deliberately breaks the usual correlation between biochemical and structural traits. Fitting expression on both axes then lets one joint two-degree-of-freedom test detect genes linked to either axis. The separate coefficients still say whether a gene tracks the immediate biochemical response, the delayed structural one, or both.

The statistically important point is the direction of the regression. Expression is the outcome and physiology is the predictor, and selection depends only on the predictors. Missingness of transcriptomes is therefore ignorable given the measured physiology, and no selection correction is needed for the slopes. The joint copula approach used for body fat traits Identification of novel genes whose expression in adipose tissue affects body fat mass and distribution models the phenotypes as outcomes given each gene. That direction would be biased under phenotype-based selection, so the two approaches are complementary rather than interchangeable.

Relation to Existing Work

Three kinds of prior work shape this proposal.

The first motivates it. Integrated physiology-plus-RNA-seq studies run the two analyses in parallel and never link them statistically. The tomato salt-acclimation study Integrated Physiological and Transcriptomic Analyses Reveal the Mechanism of Salt Acclimation-Induced Salinity Tolerance in Tomato Seedlings measured growth, photosynthesis, ion ratios, membrane damage and antioxidant enzymes, then ran ANOVAs on those traits and separate differential-expression analysis on root transcriptomes. The gastric injury study Aspirin Eugenol Ester Alleviates Gastric Injury by Inhibiting Ferroptosis and Oxidative Stress measured physiology in vivo but sequenced cell lines, so no individual carries both readouts. The largemouth bass trial Dietary Supplementation of an Organic Acid-Based Feed Attractant in Juvenile Largemouth Bass found significant shifts in liver antioxidant capacity and condition factor with no change in body weight, on 270 fish. That is exactly the setting where leading and trailing indicators diverge, and where a random handful of sequenced fish would waste most of the panel's information.

The second line supplies the methods. Identification of novel genes whose expression in adipose tissue affects body fat mass and distribution is the closest existing work. It paired adipose RNA-seq with MRI-derived fat mass and fat distribution in 160 people and used joint copula models (the CJAMP package) to gain power over separate regressions, reporting 625 adiposity-associated genes. It also showed the two traits map to different biology: immune processes for fat mass and metabolic processes for distribution. We reuse its central lesson that jointly modelling correlated physiological traits beats per-trait testing, and its finding that distinct traits carry distinct gene sets. That finding is why we keep two axes instead of one composite. We differ in three ways. Its cohort was observational and sequenced everyone, whereas we address designed trials where the sequencing budget is the binding constraint. It used two traits, whereas we compress a large battery into a pre-specified leading/trailing pair. Its outcome-side copula would be biased by the selection we propose.

COMBI Combining Multiple Hypothesis Testing with Machine Learning Increases the Statistical Power of Genome-wide Association Studies contributes the idea of reducing the testing burden before calibrated testing, with permutation-based error control that we borrow for calibration. On the WTCCC data it raised replication precision from 32% to 61% relative to raw thresholding.

The third line constrains the design. Instrument-based causal methods such as PMR-Egger Testing and controlling for horizontal pleiotropy with probabilistic Mendelian randomization in transcriptome-wide association studies, MOSTWAS MOSTWAS: Multi-Omic Strategies for Transcriptome-Wide Association Studies and ti-scMR ti-scMR: trajectory-inference-based dynamic single-cell Mendelian randomization need genotype panels of hundreds to hundreds of thousands of individuals. They cannot run on 24 sequenced seedlings of one cultivar or a clonal fish cohort. We therefore do not claim causal identification. We claim the weaker but achievable result that the gene tracks physiology across individuals beyond the shared treatment effect. This partly addresses the "bystander gene" problem without genetic instruments, and we state that limit plainly.

Research Program

The first component is the selection rule. From the full-cohort panel, we build the two axes as pre-registered sums of standardized traits within each treatment. The leading score combines enzyme activities, the inverse of membrane-damage markers and total antioxidant capacity. The trailing score combines biomass, height or condition factor, and growth rate. The grouping is fixed by biology before any data are seen, not by principal components. Within each treatment, we pick $m$ individuals maximizing the determinant of the $(L,T)$ covariance. A greedy exchange algorithm suffices at these sizes. We add a small random reserve so that the full-range distribution remains estimable for diagnostics.

The second component is the analysis. We use a standard count-based linear model on expression with precision weights, for example the limma-voom approach. Treatment, $L$ and $T$ are covariates. Genes are ranked by the joint F-test on the $L$ and $T$ coefficients, and false discovery rate is controlled at 10%. As a robustness layer, we calibrate with a within-treatment permutation of the physiology labels, in the spirit of COMBI's permutation step. This guards against heavy-tailed counts at small $n$. Each discovered gene is then labelled leading-dominant, trailing-dominant or shared by its coefficient pattern.

The third component is a power calculator. For a planned trial, it takes a pilot panel and assumed gene-level effect and variance distributions, borrowed from public RNA-seq of the same tissue. It returns expected discoveries under random versus guided allocation for each candidate $m$. This is the practical deliverable for experimental physiologists.

Baselines

The naive reference is current practice. We sequence $m$ random individuals per treatment, run differential expression on treatment alone, and then correlate the significant genes with each physiological trait one at a time under Benjamini-Hochberg correction across genes and traits. This is what the tomato and gastric studies effectively do. Beating it requires more validated gene–physiology links at equal false discovery rate.

The strongest existing competitor is the joint copula analysis of Identification of novel genes whose expression in adipose tissue affects body fat mass and distribution, run with CJAMP on the same randomly chosen individuals. We give it the same two axes as its bivariate phenotype pair and the same sequencing count. That paper reports 625 associated genes at $N=160$. We found no reported result on Lacuna quantifying CJAMP's power gain over univariate regression at fixed sample size, so we will measure that gain ourselves in the subsampling experiments below. Beating it means higher recovery of the reference gene set at matched empirical false discovery rate when both methods see the same number of sequenced samples.

The composite-score baseline regresses expression on a single first principal component of the full battery. The grey relational grade used in Seed-Soaking with Melatonin for the Improvement of Seed Germination, Seedling Growth, and the Antioxidant Defense System under Flooding Stress is the analogous single-score ranking. We run this baseline with both random and guided selection, so that it competes on the axis-count question alone.

Two ablations isolate the ingredients. Guided selection with per-trait analysis tests whether the design alone carries the gain. Random selection with the two-axis joint test tests whether the model alone carries it. A third ablation selects on the trailing axis only. If leading-only genes vanish under it, that shows the leading/trailing decomposition matters rather than merely adding variance. All methods use identical counts, normalization, covariates and false discovery target, with no method-specific tuning beyond what is pre-registered.

Why not just…?

Why not just sequence more random replicates? Under the variance formula, guided selection is equivalent to multiplying $n$ by the ratio of selected to population predictor variance. The question is therefore empirical: how large is that ratio in real panels? The power calculator reports it directly. If the gain is below about 1.5-fold for typical trials, the honest recommendation is to buy more replicates, and the subsampling experiment will show which case holds. Budget is the reason this matters. The trials in the file sequence few of the 60 to 270 animals they phenotype, for example Effects of Dietary 25-Hydroxycholecalciferol Alone or in Combination with Vitamin D3 on Growth Performance with 60 piglets.

Why not just use co-expression modules correlated with traits? Grouping genes into co-expressed modules and correlating module summaries with traits cuts the number of tests and is widely used. It is the most serious threat to the analysis half of the proposal. It is not a threat to the design half, because a module analysis run on randomly chosen samples suffers from the same low predictor spread. We will run module-trait correlation under both random and guided selection. If modules plus guided selection match or beat gene-level testing, we adopt modules as the analysis layer and keep the selection rule as the contribution.

Why not just collapse the panel into one score? The bass trial shows antioxidant gains without weight gain. The adiposity study shows fat mass and fat distribution map to different processes. Both indicate that one composite averages away axis-specific genes. The trailing-only and principal-component baselines test this directly: the two-axis model should recover leading-dominant genes that a single score misses.

Why not just use Mendelian randomization for causality? It needs genetic variation and cohorts far larger than designed trials. PMR-Egger's calibration was demonstrated on biobank-scale data. We give up causal claims in exchange for applicability, and say so.

Proposed Experiments

The first experiment is retrospective subsampling on EPIC-Potsdam-style data, using paired individual-level adipose RNA-seq and MRI traits for 160 people. We treat the full-cohort analysis as the reference set. Within sex-by-age strata, which stand in for treatment, we repeatedly draw subsets of 24 and 48 individuals, either at random or by the D-optimal rule. We score each method by recall of the full-cohort reference genes and by the fraction of discoveries that fall outside the reference. A secondary readout is how often the six genes replicated in the UK Biobank (DBNDD1, PTPRU, ERAP1, ANKDD1A, LINC02798, MCC1) are recovered. Where access allows, we repeat this on the postmenopausal bone biopsy cohort Identification of Transcripts with Shared Roles in the Pathogenesis of Postmenopausal Osteoporosis and Cardiovascular Disease with $n=84$. Its clinical measures supply the physiological axes.

The second experiment is a prospective tomato salt-acclimation trial following the protocol of the tomato study (14 mL of 7.5 g/L NaCl for 9 days before severe stress). We use 96 seedlings across control, acclimation and stress groups, with the full physiological panel on every plant, including SOD, POD, MDA, electrolyte leakage, Na⁺/K⁺ ratio, height, stem diameter and leaf area. Root tissue is sampled at the time of the enzyme assays, which addresses the timing mismatch between fast and slow markers. Every plant is sequenced, which gives ground truth; the comparison is then run by subsampling 8 per group. Phenylpropanoid and peroxidase genes reported by the original study serve as positive controls, and RT-qPCR on held-out plants serves as validation.

The decisive checks are three. First, guided selection must raise reference-set recall over random selection at matched empirical false discovery rate in both datasets. Second, the gain must be at least the variance ratio predicted by the formula, which would confirm the mechanism rather than chance. Third, null permutations must hold the false discovery rate at its nominal 10% under guided selection, confirming the claim that selection on predictors leaves the slopes unbiased. If the first check fails while the third holds, the design adds nothing and the proposal reduces to recommending the joint two-axis model on random samples.

References

  1. Identification of novel genes whose expression in adipose tissue affects body fat mass and distribution: an RNA-Seq and Mendelian Randomization study
  2. Integrated Physiological and Transcriptomic Analyses Reveal the Mechanism of Salt Acclimation-Induced Salinity Tolerance in Tomato Seedlings
  3. Aspirin Eugenol Ester Alleviates Gastric Injury by Inhibiting Ferroptosis and Oxidative Stress
  4. Dietary Supplementation of an Organic Acid-Based Feed Attractant in Juvenile Largemouth Bass (Micropterus salmoides): Effects on Growth, Morphohistology, and Oxidative Stress
  5. Combining Multiple Hypothesis Testing with Machine Learning Increases the Statistical Power of Genome-wide Association Studies
  6. Testing and controlling for horizontal pleiotropy with probabilistic Mendelian randomization in transcriptome-wide association studies
  7. MOSTWAS: Multi-Omic Strategies for Transcriptome-Wide Association Studies
  8. ti-scMR: trajectory-inference-based dynamic single-cell Mendelian randomization identifies causal genes underlying phenotypic differences
  9. Seed-Soaking with Melatonin for the Improvement of Seed Germination, Seedling Growth, and the Antioxidant Defense System under Flooding Stress
  10. Effects of Dietary 25-Hydroxycholecalciferol Alone or in Combination with Vitamin D3 on Growth Performance, Blood Vitamin D Status, Immune Response, Bone Integrity, and Antioxidant Capacity of Nursery Pigs
  11. Identification of Transcripts with Shared Roles in the Pathogenesis of Postmenopausal Osteoporosis and Cardiovascular Disease