When biologists follow a non-model organism over time, such as a coral through a bleaching season or an insect through development, they usually have no reference genome. They assemble a transcript catalogue from the reads themselves, count reads per assembled fragment at each sampling visit, and look for genes whose levels change. Many of these organisms have extreme base composition: some insect mitochondrial genomes exceed 78–81% A+T. Sequencing libraries capture AT-rich and GC-rich fragments unevenly, and that unevenness differs from one library-preparation batch to the next. In a time series, batches line up with calendar time. A change in how strongly composition distorts the counts therefore looks exactly like a gene going up or down in a season.
This note proposes a way to tell the two apart. The key observation is that composition affects expression in two ways. One is a time-invariant association, which is real biology. The other is a visit-to-visit change in how composition distorts measurement, which is technical. Only the second should be removed. The second can be measured inside each assembled transcript, independently of biology, by comparing how read coverage rises and falls across stretches of different local GC content at each visit.
Core Claim
In reference-free longitudinal transcriptomics of compositionally extreme organisms, part of the apparent temporal signal is visit-specific composition bias. Standard normalization by sequencing depth and length cannot remove it, and naively regressing out GC content would erase genuine biology along with it. We claim the technical part can be identified and removed. The model estimates a visit-specific bias curve from within-transcript coverage profiles, pools those estimates across visits under a subject-level mixed model, and constrains the curves to average to zero over visits. The time-invariant composition–expression relationship is thereby left intact.
Let $y_{c,v}$ be the read count for assembled contig $c$ at visit $v$, $N_v$ the library size, and $g_c$ the contig's composition profile (GC by compartment where open reading frames can be called). We fit
$$\log \mathbb{E}[y_{c,v}] = \log N_v + \alpha_c + h_v(g_c) + x_{c,v}, \qquad \sum_v h_v(\cdot) = 0,$$
where the terms are as follows. The baseline $\alpha_c$ absorbs any stable dependence of expression on composition. The smooth visit-specific bias $h_v$ is estimated from coverage variation across GC windows within contigs. The biological deviation $x_{c,v}$ is then modeled with subject-level random effects, as in the iPOP analyses.
Within-contig coverage pins down $h_v$ separately from biology because a biological change in a transcript scales all of its windows together, while technical bias reshapes coverage window by window. The main caveat is changes in which splice variant (isoform) is expressed. We handle it by restricting calibration to contigs with no detected alternative structure.
Why this intersection, and what prior work does
Three strands of prior work motivate the gap.
Impact of GC content on gene expression pattern in chicken shows that composition is genuine biology. GC content explains about 10% of expression variance across 19 tissues. 5′ UTR GC correlates positively with expression level (r = 0.2328), while coding and intronic GC correlate negatively. This result constrains the design: any correction that simply regresses GC out of expression would delete a real effect. That is why our model keeps the time-invariant term.
Transcriptome deep-sequencing and clustering of expressed isoforms from Favia corals shows the resource-scarce regime. A de novo coral transcriptome could be annotated for only about 20% of contigs against Nematostella, and adding Acropora gained just 1–3%. Any longitudinal analysis must therefore count unannotated contigs too, and it cannot rely on curated gene models to check for bias. The mitogenome studies quantify the composition extremes these assemblies face, for example 78.6% A+T in The mitochondrial genome of Asyndetus clavipes.
The iPOP studies supply the longitudinal analysis template that we reuse:
- Deep longitudinal multiomics profiling reveals two biological seasonal patterns in California reports 1,133 seasonal analytes across 902 visits.
- Longitudinal multi-omics of host–microbe dynamics in prediabetes shows that individuals track their own baselines and that infection can be distinguished from health with 88.6% accuracy.
Both operate on high-quality human references and never examine composition-dependent measurement drift. We do not assume their seasonal findings are artifacts. We use their setting to test how much a reference-free pipeline would distort such conclusions.
Multi-Omics Factor Analysis is relevant as the downstream integration step. It discovers latent factors and tolerates missing assays. It would likely recover a composition-drift factor, but it would have no way to label that factor as technical. Our model supplies that label before integration.
The closest existing combination of composition and scarcity, Metagenomic binning with assembly graph embeddings, treats k-mer composition as a useful signal for grouping contigs in single snapshots. We treat composition as a nuisance whose effect changes over time.
Research Program
We would build a quantification layer that sits between de novo assembly and longitudinal statistics, in three parts.
The first part is assembly-aware unit construction. AT-rich, low-coverage transcripts fragment more often, so the unit being counted itself depends on composition. We cluster contig fragments that share reads, using the approach of the Favia pipeline's CAP3 redundancy reduction but across visits, and record each cluster's window-level GC profile.
The second part estimates the visit-specific bias curves $h_v$ jointly across all visits. We use smooth splines over local GC. Information is shared across visits processed in the same batch, and the zero-sum constraint is enforced.
The third part passes the corrected counts to the same mixed-effects and seasonal models used in the iPOP work, and optionally into MOFA. This lets us check whether composition-driven factors vanish while the biological factors remain.
Baselines
The naive reference point is the current default for non-model time series: per-visit depth normalization with median-of-ratios or TPM on de novo contigs, followed by linear mixed-effects and seasonal models as in Sailani et al.. Every method receives the same assembly, the same reads and the same downstream model.
The strongest existing competitor is per-sample fragment-level GC bias correction, as implemented in standard transcript quantifiers that model fragment GC. It is the most direct existing attempt to remove exactly this artifact. We would run it with its default and tuned settings on the identical contig set. No reported result was found on Lacuna for this method on de novo or AT-extreme longitudinal data. The comparison is therefore entirely ours to establish.
A second competitor is GC-as-covariate regression, which adds contig GC as a fixed effect per visit. It is simple and widely used. We expect it to remove real biology.
The oracle upper bound is reference-based quantification on human data, where a reference exists. Its downstream target is the reported 88.6% infection-versus-health accuracy from Snyder et al..
Two ablations isolate the key ingredients:
- estimating $h_v$ per visit without pooling across visits;
- dropping the zero-sum constraint.
Beating the field means two things together. First, the method must produce fewer false temporal calls than per-sample correction on semi-synthetic AT-extreme data at matched sensitivity. Second, on the human re-analysis, it must recover a larger share of the reference pipeline's seasonal and infection calls than any baseline, while preserving the published 5′ UTR GC correlation of r = 0.2328 on chicken data.
Why not just…?
Why not just use existing per-sample GC bias correction? This is the real threat. On moderate-GC human data it may perform as well as our method, and we would report that. Its weaknesses for this setting are twofold. It estimates bias one sample at a time, with no constraint linking visits. It also assumes the transcript models are correct, which de novo contigs in AT-rich organisms violate through composition-dependent fragmentation. The semi-synthetic experiment decides the question. If per-sample correction matches us at 75–80% A+T, the contribution shrinks to the assembly-unit step, and we would narrow the claim accordingly.
Why not just map reads to the closest model organism's genome? Distant references capture little of the expressed repertoire. About 20% of Favia contigs were annotatable against Nematostella, and a second coral proteome added only 1–3% (Mehr et al.). Third-position divergence under AT drift further degrades nucleotide-level mapping. Discarding about 80% of transcripts is not a neutral choice in a time series.
Why not just randomize library batches across time points? For new studies this is correct, and we will recommend it. However, field and clinical sampling often forces processing in the order samples arrive, and existing archives cannot be redesigned. Randomization also removes the confound with time without removing the added variance. Our semi-synthetic runs will include a balanced-batch arm to quantify how much of the problem good design alone solves.
Why not just add spike-in controls? Commercial spike-ins span a modest GC range and number only in the dozens, so they cannot calibrate the AT-extreme range. We will instead use custom AT-rich spike-ins as independent validation of the fitted curves, not as the correction itself.
Proposed Experiments
The first experiment is a semi-synthetic benchmark with known ground truth. We take real AT-rich sequences, from the Asyndetus clavipes mitogenome and the deposited Favia contigs, and simulate reads over 20–40 visits. Known temporal trajectories, including seasonal ones, are planted in a subset of transcripts. We then inject visit-specific GC bias curves whose shape drifts with batch, and batch is confounded with time. The pipeline runs fully de novo. Metrics are the false discovery rate of temporal calls, recovery of true trajectories (correlation of estimated and planted log-fold changes), and recovery of the planted curves $h_v$. The genome-wide A+T fraction is varied from human-like levels up to 80% to locate where per-sample correction breaks down.
The second experiment is a reference-degradation test on iPOP-style longitudinal human RNA-seq. We would use the iPOP data if accessible, or otherwise a comparable public time series. The reference is withheld, transcripts are assembled de novo, and each pipeline reproduces the downstream analyses. We ask what fraction of reference-based seasonal transcript calls each pipeline recovers, how many spurious seasonal calls it adds, and whether infection-versus-health classification stays near the reported 88.6%. Separately, we test whether the visit-level bias curves fitted on the reference-based data show seasonal structure. If they do, part of any composition-enriched seasonal signal deserves scrutiny even with a reference. If they do not, that is reassuring and worth reporting.
The third experiment is a preservation check on the chicken multi-tissue data, treating tissues as "visits." Corrected data should keep the 5′ UTR versus coding GC dualism reported by Rao et al.. The GC-covariate baseline and the unconstrained ablation should erase it.
The decisive kill criterion comes from the real data. If the fitted visit-to-visit variation in $h_v$ is indistinguishable from bootstrap noise in both the human data and a real AT-rich series, the artifact is negligible in practice and the proposal should stop there. The project is designed so that this negative answer is also a publishable, useful result for anyone planning non-model time-series studies.
References
- Impact of GC content on gene expression pattern in chicken
- Transcriptome deep-sequencing and clustering of expressed isoforms from Favia corals
- The mitochondrial genome of Asyndetus clavipes (Diptera: Dolichopodidae)
- Deep longitudinal multiomics profiling reveals two biological seasonal patterns in California
- Longitudinal multi-omics of host–microbe dynamics in prediabetes
- Multi‐Omics Factor Analysis—a framework for unsupervised integration of multi‐omics data sets
- Metagenomic binning with assembly graph embeddings