Cancer and disease genomics studies routinely run a three-step pipeline. First, they fit a survival model to RNA sequencing data and get a ranked list of "prognostic" genes. Second, they re-measure a handful of those genes with quantitative real-time PCR (qRT-PCR), a targeted assay that counts how many amplification cycles it takes for a gene's signal to appear. Third, they pick one or two genes for expensive knockdown or knockout experiments. This proposal argues that the first step quietly biases the third. Expression values are noisy, and the noise differs from gene to gene. Low-abundance transcripts are measured far less precisely than abundant ones. Noise in a covariate shrinks its estimated effect toward zero, so the ranking that decides which genes reach the bench is tilted toward well-measured genes rather than important ones.
The opportunity is that step two already produces the missing ingredient. If qRT-PCR technical replicates and paired RNA-seq/qRT-PCR measurements are collected on the same samples, they directly estimate each gene's measurement-error variance. That estimate can then be fed into an existing exact correction for survival models with noisy covariates. We propose to turn the validation step from a yes/no confirmation into a calibration substudy, and to re-rank candidate genes for functional validation using bias-corrected, uncertainty-aware hazard ratios. A ranking that changes after correction is a concrete, testable claim about which knockdowns are worth paying for.
Core Claim
The proportional hazards model (the Cox model) relates each covariate to the instantaneous risk of an event, such as death or relapse, through a hazard ratio. It handles patients whose event times are unobserved because follow-up ended (censoring). Suppose gene $j$'s true log-expression $X_j$ is observed as $W_j = X_j + U_j$, where $U_j$ is measurement noise with gene-specific variance $\sigma_j^2$. Then the naive Cox coefficient is approximately attenuated by a reliability factor:
$$\beta_j^{\text{naive}} \approx \lambda_j \beta_j, \qquad \lambda_j = \frac{\mathrm{Var}(X_j)}{\mathrm{Var}(X_j) + \sigma_j^2}.$$
Here $\beta_j$ is the true log hazard ratio. The form is the classical first-order attenuation result for additive error; it is exact for linear models and approximate for Cox. Because $\lambda_j$ varies across genes, the naive ranking by $|\beta_j^{\text{naive}}|$ or by p-value is not a ranking by $|\beta_j|$.
An Exact Corrected Log-Likelihood Function for Cox's Proportional Hazards Model under Measurement Error supplies the method. It derives an exact corrected likelihood for Cox regression under additive normal error, including the heteroscedastic case, by dividing exponential terms by the error's moment generating function. That paper also states the obstacle: the correction requires knowing the error distribution, which is rarely available for omics data. Our claim is that the qRT-PCR validation step, if designed for it, provides exactly that knowledge. The corrected, sparse estimates then give a different and better prioritization of genes for functional validation.
Why This Matters for the Functional Validation Gap
The research file documents a large literature that nominates candidate genes from omics screens and then stops. Examples include multi-omics nomination of freezing-resistance pathways in tea Hui Wu et al. and immune-metabolic candidates in gestational diabetes Niankun Chen et al.. The file also records that transcript changes often reflect downstream injury rather than causal drivers Kyongsok So et al., and that wet-lab validation scales so slowly that researchers can only test a small fraction of candidates.
These papers motivate the proposal: they show that the bottleneck is choosing which few genes to perturb. Our contribution does not remove the need for knockouts or RNA interference. It changes the input to that decision, and it measures whether the change raises the hit rate of downstream perturbation experiments.
Closest Existing Work
The obvious version of this intersection is "build a Cox prognostic signature from RNA-seq, confirm the genes by qRT-PCR, run a cell viability assay." This is already done. Hui Guan et al. constructed prognostic models from phase-separation-related genes in colorectal cancer and performed cell viability assays. The section of the research file that would have listed further overlapping work was truncated, so we cannot rule out other close matches.
The difference from Guan et al. and similar studies is specific. They treat qRT-PCR as an error-free confirmation, and they rank genes from the uncorrected model. We treat qRT-PCR replicates as a calibration instrument for per-gene error variance. We propagate that variance into the hazard model, and we evaluate the result by whether it changes and improves which genes are perturbed.
Two methods papers constrain the design. Jianqing Fan, Yan Feng, and Yichao Wu extend sure independence screening to Cox models, iterating it and combining it with the SCAD penalty, a sparsity penalty that leaves large coefficients unbiased. Michael Komodromos et al. give sparse variational Bayes for high-dimensional Cox models, which returns posterior inclusion probabilities. Neither models covariate noise. Komodromos et al. note that their mean-field approximation underestimates posterior variance, which matters when we use inclusion probabilities to rank genes. DeepSurv trains a neural network on the Cox partial likelihood to capture non-linear effects; it also assumes inputs are measured without error.
Research Program
The work has three parts.
Part 1: a calibration design embedded in validation. For a cohort with RNA-seq and survival follow-up, a subset of samples is re-assayed by qRT-PCR for a candidate panel of roughly 50–100 genes, which is within the throughput limit the file describes. Each sample is run in technical triplicate against at least three reference genes. Replicate spread estimates within-assay noise. Disagreement between the two platforms on the same RNA estimates the RNA-seq error variance $\sigma_j^2$, under the assumption that the two platforms' errors are independent; this is the standard method-comparison logic.
Reference-gene instability is known to skew relative-expression estimates. We model it as an extra, sample-level error component, estimated from the spread across alternative reference-gene normalizations, rather than assuming a single reference gene is stable.
Part 2: estimation. The estimated heteroscedastic error variances enter the exact corrected Cox likelihood. We add a sparsity prior in the spirit of Komodromos et al.; where calibration is affordable, we replace mean-field inference with sampling (MCMC) on the reduced candidate panel. We account for uncertainty in the estimated $\sigma_j^2$ by resampling the calibration subset, because plugging in a noisy error estimate as if it were known would understate uncertainty a second time.
Part 3: prioritization. Genes are ranked by the posterior probability that the corrected hazard ratio exceeds a clinically meaningful threshold. We report which genes enter or leave the top-$k$ relative to the naive ranking. Those movers are the substance of the scientific claim.
Baselines
The do-nothing reference is the pipeline the field uses today. A univariate-screened, penalized Cox model is fit on RNA-seq log-expression, genes are ranked by coefficient magnitude, and qRT-PCR is used only to report agreement in direction. Every other method is compared against this ranking on the same cohort, same candidate panel and same train/test splits.
The strongest high-dimensional competitor is iterative sure independence screening with SCAD Fan, Feng and Wu, alongside sparse variational Bayes Cox Komodromos et al.. These are the best-established sparse Cox estimators in the file. Each receives the same tuning budget for penalty or prior hyperparameters, chosen by the same cross-validation folds. No reported result was found on Lacuna for a concordance index or variable-selection accuracy of either method on any cohort we plan to use, so the numbers we compare against will come from our own runs.
The non-linear competitor is DeepSurv, trained on the same features with an architecture and learning-rate search capped at the same compute budget. It tests whether flexibility alone recovers what error correction recovers. No reported result was found on Lacuna for DeepSurv on these data.
The simplest error correction is regression calibration. It replaces each noisy value with its predicted true value given the estimated error variance, then fits an ordinary penalized Cox model. This uses exactly the same calibration data as our method, so it isolates the choice of estimator.
The key ablation runs our full method with one shared error variance for all genes. This removes heteroscedasticity, the ingredient that should reorder genes. If the shared-variance version produces the same ranking as the gene-specific one, the central mechanism is wrong.
Beating the baselines means three things. In simulation, the method must have higher recovery of truly non-null genes in the top $k$ and lower hazard-ratio bias. In real data, it must give better transfer of a signature trained on RNA-seq to qRT-PCR-measured held-out samples, scored by concordance index. In the wet-lab arm, it must give a higher rate of knockdowns with a measurable phenotype.
Why not just…?
Why not just fit the Cox model directly on the qRT-PCR measurements? qRT-PCR is more precise per gene, but it is affordable only on a subset of samples and a panel of genes. A model fit on it alone has far fewer events and cannot screen the transcriptome. qRT-PCR is also not error-free: reference-gene instability skews its fold changes. We include a qRT-PCR-only Cox fit on the calibration subset as an extra comparison. If it matches our method's ranking accuracy in simulation at realistic subset sizes, the joint model is unnecessary. We expect it to lose because of lower power, but that is the experiment that decides.
Why not just use regression calibration instead of the exact corrected likelihood? This is a real threat. Regression calibration is simple, widely used, and often adequate when effects are moderate. It is approximate for Cox models and degrades for large hazard ratios and heavy censoring, which is the regime the exact correction was derived to handle. We do not argue the point in advance: regression calibration is a full baseline. If it matches the exact correction, we report that the contribution is the calibration design and the error-aware ranking, not the estimator, and we recommend the simpler estimator.
Why not just collect more samples or use a better sequencing depth? More samples shrink sampling variance but leave attenuation bias unchanged; the reliability factor $\lambda_j$ does not depend on cohort size. Deeper sequencing reduces counting noise for low-abundance genes, but it does not touch biological-replicate and normalization error, and it costs more than a targeted qRT-PCR panel. The simulation arm varies cohort size and per-gene noise separately to show whether ranking errors persist as $n$ grows.
Why not just use a flexible model like DeepSurv? Flexibility addresses non-linearity, not noisy inputs. An attenuated signal stays attenuated inside a neural network. DeepSurv is run as a baseline. Its failure to move the low-abundance genes that correction moves would support the claim.
Proposed Experiments
Simulation. The first experiment is a simulation with known ground truth. We generate survival times from a Cox model with sparse true effects, using realistic expression covariances from a real cohort. We add heteroscedastic noise whose variance falls with expression level, and simulate qRT-PCR replicates with reference-gene drift. We vary censoring fraction, number of events, calibration-subset size and noise heterogeneity. Metrics are top-$k$ recall of true non-null genes, hazard-ratio bias, and credible-interval coverage. Coverage is reported explicitly because Komodromos et al. warn that mean-field inference undercovers.
Real-data platform transfer. The second experiment uses a cancer cohort with RNA-seq and survival, plus a laboratory qRT-PCR calibration substudy on archived RNA. We train on RNA-seq and evaluate concordance on held-out samples measured by qRT-PCR. We also report how many top-ranked genes change identity after correction, and whether the movers are disproportionately low-abundance, as the attenuation formula predicts.
Wet-lab arm (decisive check). The decisive check is a small, pre-registered functional arm in a relevant cell line, following the knockdown-plus-viability design of Guan et al.. We knock down an equal number of genes from three groups: those ranked top only by the naive model, those ranked top only by the corrected model, and those ranked top by both. qRT-PCR confirms knockdown efficiency, and viability and migration assays score the phenotype.
The proposal succeeds if corrected-only genes yield phenotypes at a rate at least comparable to shared genes and higher than naive-only genes. It fails if the correction reorders genes but the reordering carries no functional signal. Either outcome tells the field whether its validation step should be redesigned as a calibration step.
References
- An Exact Corrected Log‐Likelihood Function for Cox's Proportional Hazards Model under Measurement Error and Some Extensions
- Transcriptome and Metabolome Analysis Revealed the Freezing Resistance Mechanism in 60-Year-Old Overwintering Camellia sinensis
- Integrative multi-omics analysis reveals cellular and molecular insights into gestational diabetes mellitus
- Transcriptome profiling revealed salt stress-responsive genes in Lilium pumilum bulbs
- Liquid-liquid phase separation-driven molecular subtyping and prognostic modeling in colorectal cancer
- High-dimensional variable selection for Cox's proportional hazards model
- Variational Bayes for high-dimensional proportional hazards models with applications within gene expression
- Deep Survival: A Deep Cox Proportional Hazards Network