When a tumor cell senses a change in its environment — a nutrient shortfall, a hormone surge, a drug — some molecule makes first contact and flips the first switch. Everything downstream, from kinase cascades to genome-wide changes in gene expression, is mapped in fine detail; the first contact almost never is, because it is fast, often post-translational, and invisible to standard steady-state assays. At the same time, the one setting where sensing really matters — a living animal with systemic metabolism, vasculature, and pharmacokinetics — is the setting where transcriptomics is used least ambitiously: in vivo xenograft studies (human tumors grown in mice) typically read out only tumor size and a few protein stains, while the RNA sequencing is imported from static public databases of patient tissue. We propose to close both gaps at once.
The concrete proposal: run dense, multi-time-point RNA sequencing directly on xenograft tumors while a defined upstream perturbation is applied systemically to the host, then fit a causal kinetic model — an ordinary-differential-equation model of gene production and decay rates over a regulatory network — to separate the primary sensing nodes and their direct targets from the hundreds of secondary transcriptional ripples. The testbed is the best-characterized metabolic sensing circuit in the corpus: Caveolin-1 (CAV1) sensing the ketone body sodium 3-hydroxybutyrate in pancreatic cancer. The proposal's deliverable is a reusable recipe — perturb, sample the in vivo transcriptome over time, fit the kinetic model, validate the inferred top node — transferable to other sensors and other tumor types.
Core Claim
No published study has measured the genome-wide transcriptional cascade triggered inside a xenograft tumor by a controlled manipulation of an upstream sensor. The closest work, the CAV1 study in pancreatic adenocarcinoma Integrative insights into the role of CAV1 in ketogenic diet and ferroptosis in pancreatic cancer, identified CAV1 as the upstream mediator of ketogenic-diet sensing using public static RNA-seq (TCGA, GTEx, GSE141017) plus in vitro cell lines, and then validated in vivo only with caliper growth curves and multiplex immunofluorescence at endpoint — the xenografts themselves were never transcriptomically profiled, and the AMPK/NRF2 → SLC7A11/SLC40A1 cascade was never shown to operate in vivo at the RNA level. Our claim is that a longitudinally sampled in vivo RNA-seq time-course, decoded by a causal kinetic model adapted from FLeCS A scalable gene network model of regulatory dynamics in single cells, will (a) reveal whether the in vitro-derived cascade is reproduced, re-wired, or absent in the host context, and (b) rank candidate upstream sensory nodes in a way that endpoint differential expression cannot.
Research Program
The experimental core is a ketogenic-diet versus normal-chow experiment in pancreatic cancer xenografts (MIA PaCa-2 and PANC1 lines, matching the CAV1 study, plus at least two PDX models), with serial early sacrifices — for example days 0, 1, 3, 7, 14, and humane endpoint — and bulk RNA-seq on each tumor, separating human from mouse reads to exclude host stromal contamination. A parallel arm genetically perturbs the sensor itself (CAV1 knockdown and overexpression), so the model can be fit to both an environmental perturbation and a node-specific one, which is what lets the kinetic model assign causality rather than correlation.
Computationally, we adapt the public FLeCS codebase Bertinus/FLeCS — validated in vitro where it recovered known direct targets in 11 out of 12 K562 Perturb-seq knockdowns, outperforming standard differential expression — to bulk in vivo data. Two adaptations are substantive and are where the methodological contribution lives. First, sampling times are censored by sacrifice: RNA is harvested at whatever time the tumor crosses the humane volume limit, and in a real mouse clinical trial dataset 88% of dropouts were exactly such planned sacrifices, a Missing-At-Random mechanism that biased a naive mixed model's growth slope to 0.24 log(mm³)/week where a joint survival-longitudinal model gave −0.02 On the Choice of Longitudinal Models for the Analysis of Antitumor Efficacy in Mouse Clinical Trials of Patient-derived Xenograft Models. Formally, tumor $i$ is sampled at $t_i = \min(T_i^{\text{sac}}, t_{\text{schedule}})$, where $T_i^{\text{sac}}$ is the (tumor-volume-dependent) sacrifice time; the kinetic model's likelihood must integrate over this censoring rather than treat $t_i$ as fixed. Second, because PDX models retain only 50–70% of donor mutational load and frequently undergo whole-genome doubling in the host Whole genomes define concordance of matched primary, xenograft, and organoid models of pancreas cancer, we passage-match all conditions and carry ploidy-aware copy-number states as covariates so that drift is not misread as sensing.
Validation follows the playbooks already proven in the corpus: co-immunoprecipitation and Western blot for predicted protein-protein edges (the methods that established CAV1–SLC7A11 binding), with SLC7A11 and SLC40A1 transcriptional trajectories as the primary positive-control readout. As a second, harder testbed with independent ground truth, we apply the same pipeline to glioblastoma xenograft lines, where network modeling on phosphoproteomics previously unmasked MEK1 as a universal hidden hub — a hub the raw mass spectrometry missed entirely because the assay covered only tyrosine phosphorylation, confirmed afterward by Western blot in 7 of 8 lines Network Modeling Identifies Patient-specific Pathways in Glioblastoma. If RNA-based kinetic modeling can recover MEK1 as a top connector in those same lines, that is strong evidence the method finds sensors through the transcriptional back door. This also differs from the Steiner-tree work that linked EGFRvIII signaling to transcription entirely in cell culture Linking Proteomic and Transcriptional Data through the Interactome and Epigenome Reveals a Map of Oncogene-induced Signaling: we bring that receptor-to-transcriptome question into a living host, with time as an axis neither Steiner-tree study had.
Baselines
The naive reference point is the standard in vivo arm as actually practiced: normal-chow growth curves and endpoint histology, with no perturbation transcriptomics at all — this is precisely what the CAV1 study's xenograft arm delivered Integrative insights into the role of CAV1 in ketogenic diet and ferroptosis in pancreatic cancer, so it is both the field's default and the floor to beat. The strongest existing-method competitor is standard differential expression plus GO/KEGG enrichment on a single treated-versus-control endpoint, run on our own time-course data collapsed to its final time point; this is what the CAV1 study effectively did with public data, and its known weakness is documented — in the FLeCS benchmark, standard differential expression was confounded by indirect compensatory changes, while the kinetic model matched known biology in 11 of 12 knockdowns A scalable gene network model of regulatory dynamics in single cells. We run both baselines on identical RNA-seq samples, with the same gene set, compute, and tuning budget, and the same DESeq2-style preprocessing. A third comparator is a static-response predictor in the spirit of MM-Net, the multimodal model that predicts PDX drug response from baseline expression and histology with a Matthews Correlation Coefficient of 0.3102 versus 0.2958 for expression alone Data augmentation and multimodal learning for predicting drug response in patient-derived xenografts from gene expressions and histology images; we retrain it on baseline xenograft expression to predict ketogenic-diet response, testing whether static baselines capture what dynamics do. Beating the baselines means three concrete things: the kinetic model ranks CAV1 within the top inferred upstream nodes for the diet perturbation where endpoint differential expression does not; predicted direct targets (led by SLC7A11, SLC40A1) separate from secondary genes with a reproducible time-lag structure; and held-out time points are predicted better than by any baseline. Two ablations isolate our key ingredients: collapsing the time-course to a single endpoint (isolating longitudinal sampling) and removing the censoring-aware likelihood (isolating the sacrifice correction, where the −0.02 versus 0.24 slope flip in the mouse-trial data shows the order of magnitude at stake).
Why not just…?
Why not just do it in vitro or in organoids? Because the perturbation is systemic. A ketogenic diet works through whole-body hepatic metabolism, circulating ketone levels, and native vascularization — none of which exist in a dish, which is why the CAV1 study needed living hosts for its diet arm Integrative insights into the role of CAV1 in ketogenic diet and ferroptosis in pancreatic cancer. Organoids are a partial answer at best: matched primary–PDX–organoid trio sequencing showed organoids inherit the PDX bottleneck plus culture-specific selection rather than the patient context Whole genomes define concordance of matched primary, xenograft, and organoid models of pancreas cancer. We do use the in vitro cascade as prior knowledge and validation target, not as the measurement itself.
Why not just use phosphoproteomics instead of RNA-seq? This is a real threat, since mRNA abundance tracks protein abundance only 40–60% of the time A high-stringency blueprint of the human proteome, and sensing is often post-translational. But mass spectrometry has its own documented blind spots: tyrosine-focused phosphoproteomics missed MEK1 entirely in the glioblastoma xenograft study, and only network reconstruction over the noisy data found it Network Modeling Identifies Patient-specific Pathways in Glioblastoma. Our design concedes the point partially: every model-predicted edge gets targeted protein-level validation (Co-IP, Western blot, phospho-assays), and if the RNA-only model fails to recover CAV1 while a phospho time-course would, that itself is a decisive, publishable negative result about the limits of transcription-only sensor inference.
Why not just run standard DEG analysis? Because the corpus already benchmarks this head-to-head and DEG loses: in the K562 Perturb-seq comparison, standard differential expression was swamped by secondary ripples while the kinetic model recovered direct targets in 11 of 12 knockdowns A scalable gene network model of regulatory dynamics in single cells, and simple correlation between omics layers failed where Steiner-tree optimization found druggable hubs Linking Proteomic and Transcriptional Data through the Interactome and Epigenome Reveals a Map of Oncogene-induced Signaling. Still, we run DEG as our primary baseline on identical data, so the comparison is settled empirically in the in vivo setting rather than by citation.
Decisive Checks
Three results would make or break the proposal. First, does the in vivo ketogenic-diet transcriptome reproduce the in vitro AMPK/NRF2 program — early SLC7A11 and SLC40A1 repression with the expected lag behind CAV1 loss — or reveal a re-wired cascade? Either answer is informative; no reported result exists on Lacuna for in vivo transcriptomics of this perturbation. Second, does the kinetic model recover MEK1 as a top hidden connector in the glioblastoma lines, where the gold standard is Western-blot confirmation in 7 of 8 lines Network Modeling Identifies Patient-specific Pathways in Glioblastoma? Third, does the censoring-aware likelihood change the inferred direct-target set relative to the naive fit by an amount explainable through the documented sacrifice mechanism (88% of dropouts in the mouse-trial dataset)? For trajectory prediction on held-out time points there is no reported in vivo benchmark on Lacuna; we report the first numbers ourselves, which is part of the point.
References
- Integrative insights into the role of CAV1 in ketogenic diet and ferroptosis in pancreatic cancer
- A scalable gene network model of regulatory dynamics in single cells
- On the Choice of Longitudinal Models for the Analysis of Antitumor Efficacy in Mouse Clinical Trials of Patient-derived Xenograft Models
- Whole genomes define concordance of matched primary, xenograft, and organoid models of pancreas cancer
- Network Modeling Identifies Patient-specific Pathways in Glioblastoma
- Linking Proteomic and Transcriptional Data through the Interactome and Epigenome Reveals a Map of Oncogene-induced Signaling
- Data augmentation and multimodal learning for predicting drug response in patient-derived xenografts from gene expressions and histology images
- A high-stringency blueprint of the human proteome