Dynamic Fidelity: Benchmarking Cell Lines by Time-Resolved Response to RNA Metabolic Therapeutics

RNA drugs that rewire metabolism — a small interfering RNA (siRNA) that silences a glycolytic gatekeeper, say — are discovered and validated almost exclusively in immortalized cell lines, and those lines are chosen by a single number: how similar their baseline gene-expression profile is to the corresponding primary tissue. But the perturbation being modeled is not static. Metabolic reprogramming shifts a cell's membrane composition, endocytic activity, and energy economy over hours, which means the very machinery that decides whether a delivery vehicle gets in is itself changed by the therapy. Nobody has asked whether a cell line that "looks like" tissue at baseline also behaves like tissue in the first 72 hours after an RNA metabolic intervention — what we call dynamic fidelity — and nobody has measured how the metabolic shift caused by a first dose feeds back into the uptake of a second one.

The core hypothesis is that static similarity does not predict dynamic response fidelity, and that delivery efficiency is a function of the cell's current metabolic state rather than its nominal identity. If true, this reorders how RNA therapeutic candidates should be screened: by time-resolved response trajectories against primary references, not baseline correlation, and with dosing scheduled against the metabolic state the therapy itself induces. Three strands of existing work motivate this without answering it. The TCGA-110 panel selected the "best" cell lines per cancer type using only baseline Spearman correlation over the top 5,000 variable genes, and its own authors flag that experimental validation of whether high-fidelity lines better predict drug responses is still missing (Comprehensive transcriptomic analysis of cell lines as models of primary tumors across 22 tumor types). The hepatic drug-metabolizing-enzyme panel shows the same limitation in liver: HepaRG correlates with native liver at $\rho = 0.67$ while workhorse HepG2 sits at $0.26$, yet all rankings rest on untreated baseline profiles (Drug-Metabolizing Gene Expression Identity: Comparison Across Liver Tissues and Model Cell Lines). And FLeCS, a learnable ordinary-differential-equation model of gene regulation, can already predict 72-hour single-cell trajectories under perturbation in K562 and A549 cells — but it treats the perturbation as an instantaneous, perfectly delivered switch and ignores metabolism entirely (A scalable gene network model of regulatory dynamics in single cells). The gap sits exactly at the intersection: no study pairs a targeted RNA perturbation with time-resolved profiling to ask whether the model system's trajectory matches the primary reference, and no study couples the perturbation's metabolic consequences back into delivery kinetics.

Core Claim

We claim two things, one descriptive and one causal. Descriptively: across a panel of cell lines spanning a known range of static fidelity, static baseline correlation with primary tissue is a poor predictor of dynamic fidelity — the agreement between the line's time-resolved transcriptional and metabolic response to a standard RNA metabolic knockdown and the primary tissue's response. Causally: knocking down a metabolic enzyme (for example a pyruvate dehydrogenase kinase, which dams pyruvate out of the mitochondrial TCA cycle and forces lactate production, as characterized in glioblastoma by Refining the Role of Pyruvate Dehydrogenase Kinases in Glioblastoma Development) changes the cell's internalization of RNA delivery vehicles over the following hours, so a second dose delivered at a model-predicted metabolic timepoint outperforms the same dose delivered on a fixed schedule. Both claims are testable with existing benchmarks and reagents; neither is addressed by any study in this file.

Research Program

The program has three connected parts built around one standard perturbation: a fluorescently labeled siRNA (or splice-switching oligonucleotide) targeting a metabolic gatekeeper, delivered by a peptide-guided or lipid vehicle, applied identically to every model in a panel.

The first part is the dynamic fidelity benchmark. We take the hepatic system as the anchor because it has the best-characterized static fidelity ladder in the file: HepaRG ($\rho = 0.67$ versus liver, $0.90$ versus primary hepatocytes), HepG2 ($0.26$), Huh7 ($0.24$), and non-parenchymal lines near zero, with primary human hepatocytes and intact liver tissue as references (Drug-Metabolizing Gene Expression Identity). We extend the panel with the top-ranked lines from TCGA-110 for a solid-tumor indication where dynamic references exist. Each model receives the same siRNA knockdown, and we profile transcriptomes at 0, 6, 12, 24, 48, and 72 hours, inferring metabolic fluxes at each timepoint with GEFMAP, which predicts flux vectors from single-cell expression while enforcing mass conservation in the stoichiometric null space ($Sv = 0$) and reports flux prediction Pearson correlation of $0.928 \pm 0.016$ on synthetic data and $0.841 \pm 0.014$ on an augmented E. coli set (Inferring Metabolic States from Single Cell Transcriptomic Data via Geometric Deep Learning). Dynamic fidelity is then a trajectory-level agreement score between each line and the primary reference over the full time course. The decisive statistic is simple: the rank correlation between static $\rho$ and dynamic fidelity across the panel. If it is near zero, static selection is failing at exactly the job it is used for.

The second part tests the causal claim. Delivery of the labeled vehicle is measured under deliberately shifted metabolic states — glucose versus galactose medium, hypoxia versus normoxia — because tumor bioenergetics are plastic states that transition between glycolysis and oxidative phosphorylation depending on microenvironmental pressure (Absolute dynamic and relative static: the relationship of glycolysis and OXPHOS in cancer development), and because chronic intermittent hypoxia provably rewires metabolic gene expression on 24-hour cycles in liver, peaking at ZT6 (Chronic intermittent hypoxia reshapes circadian metabolic architecture in a model of sleep apnea). If vehicle uptake tracks the inferred flux state rather than the line's name, the right conditioning signal for delivery design is a continuous metabolic descriptor, not a categorical label. This directly extends the soft-prompt peptide generator of Parameter-Efficient Adaptation of a Pretrained Language Model via Soft Prompt Tuning Enables Hit-Enriched Conditional Cell Line-Specific Generation of Cell-Penetrating Peptides, which conditions generation on static line labels and itself calls for moving to continuous biological descriptors as an unaddressed future requirement; we test its cross-line transfer when the conditioning vector is a GEFMAP flux snapshot instead of a label.

The third part is the coupled model. We extend the FLeCS dynamics (available as a codebase, FLeCS), which evolve each gene's expression as production minus decay,

$$\frac{dx_i}{dt} = p_i(\text{parents}(i)) - d_i x_i + u_i(t), \qquad u_i(t) = \eta(v(t))\, c(t),$$

where $u_i(t)$ is the effective knockdown rate on target gene $i$, written as the extracellular cargo concentration $c(t)$ times a state-dependent efficiency $\eta$ that is a learned function of the inferred flux vector $v(t)$. This one term is the entire conceptual addition: delivery stops being an idealized instantaneous switch and becomes a metabolic-state-gated process, which is precisely what FLeCS omits. The model is trained on the first part's time courses and asked to predict held-out 72-hour trajectories and, critically, the measured uptake of a second dose.

Baselines

The naive reference point is the vehicle-only control plus a single end-point readout: qPCR or one bulk RNA-seq timepoint at 24 or 48 hours, which is how the field currently reads out knockdown, as in the COX-2 siRNA chemosensitivity study in gastric cancer lines (Targeted inhibition of COX-2 expression by RNA interference suppresses tumor growth and potentiates chemosensitivity to cisplatin in human gastric cancer cells). This baseline costs a fraction of a time course; the proposal beats it only if the time course reveals response structure that changes the conclusion — for example, if lines ranked equivalently at 24 hours diverge by 72 hours.

The strongest existing competitor for line selection is the TCGA-110 static panel from Comprehensive transcriptomic analysis of cell lines as models of primary tumors across 22 tumor types, which reports median Spearman correlation of 0.75 for liquid tumors and roughly 0.50 for difficult solid tumors such as cholangiocarcinoma, and the hepatic DME ladder of Drug-Metabolizing Gene Expression Identity, which reports HepaRG at $\rho = 0.67$ versus liver and $0.90$ versus primary hepatocytes against HepG2's $0.26$. We will run these baselines on exactly our panel with the same sequencing depth, the same batch correction, and the same gene sets, so the only difference is static versus dynamic scoring. Beating them means one of two results: a line that static scoring ranks high but dynamic scoring exposes as unrepresentative (or vice versa), or a demonstrated prediction advantage — dynamic profiles from one perturbation predicting a second perturbation's response in primary tissue better than static $\rho$ does.

The strongest computational competitor is unmodified FLeCS itself, which resolved known regulatory mechanisms in 11 of 12 CRISPRi knockdowns in K562 and outmatched OT-CFM and scTour baselines (A scalable gene network model of regulatory dynamics in single cells). We retrain it on our data with its original instantaneous-perturbation assumption and give both models the same compute, the same prior regulatory network, and the same training timepoints; our model wins only if the $\eta(v(t))$ term improves held-out trajectory prediction and second-dose uptake prediction over FLeCS with the term ablated to a constant. That ablation — state-gated versus fixed delivery efficiency — isolates the proposal's key ingredient.

Why not just…?

Why not just run endpoint assays? The cheapest version of this study is a 48-hour knockdown readout per line, and a reviewer is right that this would answer "did the siRNA work." The file's evidence says it would not answer ours. In cadmium kidney toxicity, the initial insult is early metabolic disruption that only later becomes cell-cycle arrest, so a late end-point obscures the primary molecular target (Temporal transcriptomic analysis of cadmium-induced kidney toxicity reveals early metabolic disruption prior to structural damage); in myocardial infarction, peripheral blood shows an acute inflammatory burst that later resolves into long-term metabolic and neurosecretory remodeling, invisible at a single timepoint (Dynamic transcriptomic landscape of peripheral blood mononuclear cells during acute myocardial infarction and recovery); and in the intermittent-hypoxia liver study, a sample taken at ZT18 would conclude hypoxia suppresses metabolic genes while a sample at ZT6 shows the opposite (Chronic intermittent hypoxia). Still, the endpoint design is a real threat to the budget, and the decisive check is built in: if static and dynamic fidelity rankings turn out to agree, the endpoint studies were sufficient and the expensive layer adds nothing. We state that openly as the study's falsification condition.

Why not just use the standard workhorse lines? HepG2, Huh7, U87, and U251 are robust, published everywhere, and would make the wet-lab work easier. The file is unambiguous about why this fails for metabolic questions: HepG2 and Huh7 correlate with native liver at only $\rho = 0.26$ and $0.24$, with cytochrome P450 drug-metabolizing enzymes near zero TPM (Drug-Metabolizing Gene Expression Identity), and three widely used colorectal lines in CCLE turned out to be overgrown fibroblasts rather than epithelial cancer cells at all (Comprehensive transcriptomic analysis of cell lines as models of primary tumors across 22 tumor types). Our design does not discard these lines — it deliberately includes them as low-static-fidelity points on the ladder, because the question is precisely whether that label predicts their dynamic behavior. If it does, the workhorse lines are salvageable and we will say so.

Why not just use Cell Painting morphology instead of sequencing time courses? High-throughput imaging is non-destructive, cheap, and already coupled to transcriptomics by PhenoSeq, a conditional diffusion model that generates single-cell expression embeddings from morphology (Cell Painting Generates Single-Cell Transcriptomics via Conditional Diffusion). But that same paper quantifies a morphology ceiling: compounds like Simvastatin that leave U2-OS cells morphologically unchanged cannot be mapped to their transcriptomic effect at all, and adding synthetic transcriptomics recovered only about 29% of the performance gap, with multi-profile accuracy of 0.271 versus 0.157 for imaging alone in data-sparse conditions. A metabolic knockdown that leaves shape intact while rewiring flux is exactly the failure mode. We take the threat seriously and include a morphology-only arm at every timepoint: if imaging alone ranks the panel's dynamic fidelity identically to the sequenced time course, morphology wins on cost and we report that.

Proposed Experiments and Decisive Checks

The experiment set is, in order: first, the hepatic dynamic-fidelity ladder (HepaRG, HepG2, Huh7, non-parenchymal lines, primary human hepatocytes, and liver tissue reference data from the compiled GEO datasets of the DME panel), one siRNA target, six timepoints, with GEFMAP flux inference at each point and PhenoSeq-compatible imaging in parallel; second, the state-shift delivery assay, measuring labeled-vehicle uptake in glucose versus galactose and hypoxia versus normoxia, with GEFMAP-inferred flux as the candidate predictor; third, the model comparison, training extended FLeCS against unmodified FLeCS on the same time courses with the flux-gated efficiency term ablated; and fourth, the scheduled second dose, comparing model-timed versus fixed-schedule redosing on knockdown durability.

Three checks decide the proposal. The static-versus-dynamic rank correlation across the panel must be demonstrably low (near zero) for the central claim to stand; no reported result found on Lacuna for any existing time-resolved cell-line-versus-primary-tissue response comparison, so this number does not exist yet anywhere. The flux-gated model must beat unmodified FLeCS on held-out 72-hour trajectories and on predicting measured second-dose uptake — FLeCS's reported 11 of 12 knockdowns in K562 is the reference point for what a strong dynamics model achieves without metabolic gating (A scalable gene network model of regulatory dynamics in single cells). And the state-shift assay must show uptake varying with metabolic state within the same line by more than assay noise, which is the minimum evidence that conditioning delivery design on continuous metabolic descriptors — rather than the static cell-line labels used by the soft-prompt CPP generator, whose unconditioned base model hit only 3.4% functional enrichment versus 46.0% with label conditioning (Parameter-Efficient Adaptation of a Pretrained Language Model via Soft Prompt Tuning) — is worth the added complexity. If all three hold, the output is a reusable dynamic-fidelity benchmark, a state-gated trajectory model, and the first direct evidence that RNA metabolic therapeutics should be timed against the metabolic state they create.