When a drug is tested in a mouse carrying a human tumor graft, the two kinds of measurement we can make are almost perfectly complementary in what they cost and what they reveal. The cheapest one — calling tumor size with calipers, or imaging light emitted by luciferase-tagged tumor cells — is non-invasive and can be taken on the same animal every few days for weeks. The richest one — single-cell RNA sequencing, which reads the activity of thousands of genes in individual cells — requires killing the animal and removing the tumor, so it can only ever give isolated snapshots from different individuals. Current computational methods stitch such snapshots into continuous "trajectories" of cell-state change, but every one of them either assumes the total number of tumor cells stays constant, or guesses how the population grows and shrinks. That guess is exactly where these methods fall apart under chemotherapy, where tumors collapse and resistant clones expand. Meanwhile the caliper curve in the lab notebook already measures that growth and death directly — and no one uses it.
This proposal argues that the measured tumor volume curve should not be an afterthought but a hard physical constraint on trajectory inference. We will build and test an unbalanced-transport trajectory model in which the population growth term is not learned but pinned to the non-invasive mass curve from the same experiment, so that the sequencing snapshots only need to explain how cells change state, not how many there are. The scientific payoff is a question that static endpoint assays cannot answer: during a treated tumor's regression and relapse in a living host, which cell states appear, when do they appear relative to the visible relapse, and can we detect the resistant state before the tumor starts growing back.
Core Claim
The claim is that a trajectory inference model constrained by a measured per-capita growth rate will reconstruct and forecast drug-treated tumor evolution in vivo more faithfully than methods that infer growth from sequencing snapshots alone, and that this constraint converts an unidentifiable problem into a well-posed one. The mechanism is a single equation. Let $\rho_t$ be the distribution of tumor-cell states in gene-expression space, $v_t$ the drift field we want to infer, and $\Phi(t)$ the bioluminescence flux or caliper-derived volume measured non-invasively on living animals. We impose the unbalanced continuity equation
$$\frac{\partial \rho_t}{\partial t} + \nabla\!\cdot(\rho_t v_t) \;=\; \gamma(t)\,\rho_t, \qquad \gamma(t) = \frac{d}{dt}\log \Phi(t),$$
where $\gamma(t)$, the per-capita net growth rate, is read off the measured curve rather than fit to the data. Inference then reduces to estimating $v_t$ — how states move — subject to matching the observed snapshot distributions $\hat\rho_{t_k}$ at harvest times $t_k$:
$$\min_{v}\ \int_0^T\!\!\int \|v_t(x)\|^2\, d\rho_t(x)\, dt \quad\text{s.t.}\quad \rho_{t_k} \approx \hat\rho_{t_k} \text{ for all } k,\ \text{and the growth equation above holds.}$$
The motivation for believing a partial, low-dimensional observation can discipline a high-dimensional trajectory comes from Partially Observed Trajectory Inference using Optimal Transport and a Dynamics Prior, which shows that unobserved latent states can be recovered from partial observations when a dynamics prior is available, reporting a 2-Wasserstein distance of 1.01–1.07 on the Wikipedia traffic benchmark versus 1.61–1.91 for its fully-observed counterpart. Our partial observation is different in kind — a scalar mass channel rather than a subset of coordinates — but the logic transfers: the mass channel carries information the snapshots cannot, because it is continuous and comes from living animals whose trajectories we never otherwise observe.
The gap we fill is concrete. The global convex framework of Towards a mathematical theory of trajectory inference handles cell birth and death only through a heuristic operator-splitting scheme, with identifiability proofs limited to the mass-conservative case; Manifold Interpolating Optimal-Transport Flows for Trajectory Inference, Trajectory Inference with Smooth Schrödinger Bridges, PACE: Geometry-Aware Bridge Transport for Single-Cell Trajectory Inference, and Trajectory inference via Acceleration Matching all assume closed populations outright, which their authors explicitly flag as disqualifying for tumors under cytotoxic therapy. None of them has been run on in vivo xenograft data, and no reported result found on Lacuna for any of these methods on an in vivo or mass-varying tumor benchmark. Our proposal differs from all of them in one specific way: the growth term is externally measured, not assumed away and not inferred.
Research Program
The program has three parts: the model, a synthetic and benchmark validation phase, and a dedicated in vivo experiment designed so that the physical channel and the sequencing channel come from the same system.
The model is a global unbalanced trajectory inference built on the gWOT-style variational formulation, extended in two ways. First, $\gamma(t)$ is fixed by the measured mass curve, estimated with a hierarchical wrapper: all animals contribute volume or bioluminescence curves, the population-level $\gamma(t)$ is shared, and per-animal deviations are modeled as noise, which matters because snapshots come from sacrificed cohorts while curves come from survivors. Second, the drift is parameterized as a smooth neural vector field on a low-dimensional embedding (5–10 dimensional principal component space, the convention of the embryoid-body benchmarks and the regime where Trajectory Inference with Smooth Schrödinger Bridges remains tractable), with a Riemannian penalty on motion normal to the data manifold borrowed from PACE: Geometry-Aware Bridge Transport for Single-Cell Trajectory Inference, since xenograft harvests are sparser than in vitro time courses and corner-cutting is a documented failure mode there.
The mixed-species complication is handled by assignment, not alignment. Human and mouse reads are separable by genome of origin, so host stromal cells are partitioned out of the tumor trajectory rather than merged into it. We deliberately avoid adversarial removal of the host compartment: Integrating single-cell RNA-seq datasets with substantial batch effects shows that strong adversarial discriminators force biologically distinct cell types to over-mix and lose identity, and their cycle-consistent alternative is a static harmonization tool with no dynamics. Instead, the murine stromal compartment is retained as a second, coarser population whose trajectory is inferred jointly, so that host-response dynamics (of the kind documented in D-xylose suppresses hepatocellular carcinoma progression by regulating dihydrodiol dehydrogenase and remodeling the immune microenvironment) can be read alongside the human compartment without contaminating it.
The in vivo experiment uses A549 lung carcinoma cells, chosen deliberately: A scalable gene network model of regulatory dynamics in single cells already modeled A549 drug response over 72 hours in vitro, so extending the same line into an animal gives a direct in vitro-to-in vivo comparison of inferred dynamics. Luciferase-expressing A549 cells are grafted subcutaneously into immunodeficient mice; all animals are measured by caliper twice weekly and by bioluminescence imaging, following the non-invasive longitudinal imaging logic of BKM120 sensitizes glioblastoma to the PARP inhibitor rucaparib by suppressing homologous recombination repair. Treatment follows a regression-relapse design with a cytotoxic regimen of the type shown in vivo to produce deep regressions, such as the albumin-encapsulated SN-38 nanoparticles that Albumin-based nanoparticles encapsulating SN-38 demonstrate superior antitumor efficacy compared to irinotecan reports outperforming irinotecan. Roughly eight harvest time points span baseline, early response, maximal regression, and relapse, with a cohort of three animals sacrificed per point and single-cell RNA sequencing of each tumor split by species. One additional day-14 cohort is sacrificed while the remaining animals are followed to individual relapse, so that the inferred day-14 population composition can be tested against the observed distribution of relapse times.
Proposed Experiments
Validation proceeds from synthetic to benchmark to in vivo. First, we construct mass-varying versions of existing synthetic benchmarks — the Petal and Dyngen branching systems used by Manifold Interpolating Optimal-Transport Flows for Trajectory Inference and the Lotka-Volterra dynamics used by Trajectory inference via Acceleration Matching — by prescribing a growth curve $\gamma(t)$ and reweighting the snapshots accordingly. Because the true trajectories are known, we can measure whether pinning $\gamma$ to the true curve (as the caliper would supply it) recovers drift better than free-inference competitors. Second, we run the standard leave-one-time-point-out protocol on the conservative embryoid body, CITE-seq, and reprogramming benchmarks (EB, CITE50, Schiebinger2019) to confirm we lose nothing when mass is genuinely conserved, using the same protocol under which PACE: Geometry-Aware Bridge Transport for Single-Cell Trajectory Inference reports an average 23.7% reduction in MMD and Wasserstein error across nine held-out experiments and a 15.4% improvement in RNA-velocity alignment, and under which Trajectory inference via Acceleration Matching reports a holdout 2-Wasserstein error of 0.376 on Lotka-Volterra versus 0.427–1.116 for first-order flow matching, and a 1-Wasserstein error of 49.8 on CITE50 versus 200.9 for the vanilla first-order baseline.
The in vivo evaluation has three readouts. The first is distributional accuracy: hold out one harvest time point's snapshot entirely and predict it from the remaining snapshots plus the full mass curve, scored by Wasserstein distance and maximum-mean discrepancy in the shared embedding. The second is biological lead time: we score every inferred tumor state with the 189-gene Stem Cell Gene Set and PC1 Stemness Index that A gene expression profile of stem cell pluripotentiality and differentiation is conserved across diverse solid and hematopoietic cancers derived from 3,209 Concordia microarrays and validated on mouse embryonic stem cell differentiation from day 0 to day 14, and we test whether a rise in inferred stemness or a resistant-state fraction at the day-14 cohort precedes detectable volumetric relapse in the tracked animals. The premise that molecular change precedes visible phenotype is exactly what temporal profiling studies keep finding — Zhihua Liu et al. show myocardial infarction shifting from acute inflammatory to chronic metabolic programs over time, and epigenetic priming precedes morphological commitment in pancreatic differentiation (pancreatic differentiation) — but it has never been quantified against a per-animal relapse clock. The third readout is the discrepancy between the caliper-derived and bioluminescence-derived $\gamma(t)$, which indexes stromal and necrotic dilution of the bulk volume and is itself a check that the constraint channel is measuring tumor cells rather than lump mass.
Baselines
The naive reference point is the standard preclinical endpoint assay: terminal excision and treated-versus-control differential expression at the final time point, the design used throughout the xenograft literature represented in the file, for example in Integrating network pharmacology, molecular docking and experimental verification to reveal the mechanism of artesunate in inhibiting choroidal melanoma and Knockdown of Livin inhibits growth and invasion of gastric cancer cells through blockade of the MAPK pathway in vitro and in vivo. For the interpolation task the corresponding numerical null is a carry-forward prediction, where the held-out snapshot is predicted to equal the previous one. Beating the endpoint baseline means recovering a transient state (for example a resistant or high-stemness population visible mid-treatment) that the day-28 comparison cannot see, and beating the null means lower held-out distributional error.
The strongest existing competitor is the operator-splitting branching extension of the global trajectory framework in Towards a mathematical theory of trajectory inference, because it is the only prior method with any mechanism for birth and death; it will be run with the same snapshots, the same embedding dimensionality, and the same tuning budget, with its growth operator inferred freely from the snapshots. No reported result found on Lacuna for that method's quantitative error on any in vivo or mass-varying benchmark, so the comparison numbers will be established on our synthetic mass-varying benchmarks first, where truth is known. Alongside it we run the best mass-conservative methods — PACE: Geometry-Aware Bridge Transport for Single-Cell Trajectory Inference and Trajectory inference via Acceleration Matching — with snapshot masses renormalized, and the perturbation predictor scPPDM: A Diffusion Model for Single-Cell Drug-Response Prediction applied zero-shot to predict treated states, noting that it predicts a single post-treatment endpoint rather than a trajectory and was trained on the in vitro Tahoe-100M atlas. All methods receive the same data except the mass curve, which only our model and the designated ablations can ingest; that asymmetry is the claim, not an accident of setup. A method beats the field if it achieves lower held-out Wasserstein and MMD error on the mass-varying synthetic benchmarks and on the in vivo interpolation task, with confidence intervals over the three animals per held-out time point.
The key ablation isolates the one ingredient we are adding: the identical model with $\gamma(t)$ learned freely from snapshots instead of pinned to the measured curve. A second ablation gives the gWOT operator-splitting baseline the same pinned $\gamma(t)$, so that any remaining gap isolates our continuous global formulation rather than the constraint itself. A third swaps the bioluminescence-derived $\gamma$ for the caliper-derived one, testing which physical channel is the better constraint.
Why not just…?
Why not just use in vitro time courses or organotypic slices instead of animals? They are cheaper and avoid cross-species mixture, and A Feasibility Study of Co-Established Patient-Derived Subcutaneous Xenograft and Organotypic Slice Cultures in Hormone-Naive Primary Prostate Cancer Preclinical Modeling: A Single-Institution Experience shows slice cultures can be faster to establish than xenografts. But the proposal's whole input signal — a measured mass curve from a living, perfused, drug-clearing host — does not exist in a dish, and the in vitro–in vivo divergence is documented throughout the file: drugs that work on plastic fail in animals for pharmacokinetic and microenvironmental reasons (MicroRNA-218 Inhibits Cell Cycle Progression and Promotes Apoptosis in Colon Cancer by Downregulating BMI1 Polycomb Ring Finger Oncogene; In Vitro and in Vivo Anticancer Effects of American Ginseng Berry: Exploring Representative Compounds). We do use the in vitro benchmarks first; the animals are where the new information channel lives.
Why not just do static endpoint differential expression? Because it is precisely the thing this proposal must beat, and there is direct evidence it is weaker than dynamic modeling: A scalable gene network model of regulatory dynamics in single cells reports that its dynamic model aligned more closely with known biology in 11 out of 12 CRISPRi knockdowns than differential expression did, and temporal studies repeatedly show that the informative programs are transient stage-specific ones (Zhihua Liu et al.). Still, if endpoint differential expression at day 28 identifies the same resistance population our trajectory finds, and with the same lead-time value of zero, the dynamic machinery is not paying for itself — and the endpoint baseline in our design will test exactly that.
Why not just use gWOT's existing birth-death operator splitting? This is the real threat, and we treat it as the primary ablation rather than arguing around it. The honest position: that method has a mechanism for growth, and if its freely inferred growth matches the measured curve and its held-out interpolation error matches ours, our central claim of identifiability is wrong. Our expectation is the opposite — with sparse snapshots and a relapse-driven non-monotonic $\gamma(t)$, free inference of an unrestricted growth function is underdetermined in a way the caliper is not — but the pinned-versus-free ablation on synthetic benchmarks with known truth is designed to settle it, and we will report it either way.
Why not just remove mouse reads with adversarial domain adaptation? Because adversarial alignment is documented to over-mix biologically distinct cell types when the discriminator is strong (Integrating single-cell RNA-seq datasets with substantial batch effects). In a xenograft the problem is easier at the read level — human and mouse transcripts are separable by genome — so we assign rather than align, and we keep the host compartment as a modeled population instead of deleting it, since host stromal remodeling is part of the biology we want to observe, not a contaminant to hide.
Decisive Checks
The proposal stands or falls on three numbers. On synthetic mass-varying benchmarks with known drift and known $\gamma(t)$, the pinned model must beat free-growth gWOT and renormalized PACE and Acceleration Matching on held-out distributional error; if it does not, the constraint adds nothing and the idea is dead. On the in vivo interpolation task, it must beat the carry-forward null and every competitor on held-out snapshots from real tumors. And on the lead-time question, the day-14 inferred resistant-state or stemness fraction — scored with the 189-gene index of A gene expression profile of stem cell pluripotentiality and differentiation is conserved across diverse solid and hematopoietic cancers — must predict the observed relapse times of tracked animals better than chance, with a lead over caliper-only detection measured in days. That last check is the one no endpoint assay, no in vitro screen, and no mass-conservative trajectory model can even attempt, and it is the reason the two directions in this proposal belong together.
References
- Partially Observed Trajectory Inference using Optimal Transport and a Dynamics Prior
- Towards a mathematical theory of trajectory inference
- Manifold Interpolating Optimal-Transport Flows for Trajectory Inference
- Trajectory Inference with Smooth Schrödinger Bridges
- PACE: Geometry-Aware Bridge Transport for Single-Cell Trajectory Inference
- Trajectory inference via Acceleration Matching
- Integrating single-cell RNA-seq datasets with substantial batch effects
- D-xylose suppresses hepatocellular carcinoma progression by regulating dihydrodiol dehydrogenase and remodeling the immune microenvironment
- A scalable gene network model of regulatory dynamics in single cells
- BKM120 sensitizes glioblastoma to the PARP inhibitor rucaparib by suppressing homologous recombination repair
- Albumin-based nanoparticles encapsulating SN-38 demonstrate superior antitumor efficacy compared to irinotecan
- A gene expression profile of stem cell pluripotentiality and differentiation is conserved across diverse solid and hematopoietic cancers
- Potential biomarkers of acute myocardial infarction based on weighted gene co-expression network analysis
- Transcriptional changes and the role of ONECUT1 in hPSC pancreatic differentiation
- Integrating network pharmacology, molecular docking and experimental verification to reveal the mechanism of artesunate in inhibiting choroidal melanoma
- Knockdown of Livin inhibits growth and invasion of gastric cancer cells through blockade of the MAPK pathway in vitro and in vivo
- scPPDM: A Diffusion Model for Single-Cell Drug–Response Prediction
- A Feasibility Study of Co-Established Patient-Derived Subcutaneous Xenograft and Organotypic Slice Cultures in Hormone-Naive Primary Prostate Cancer Preclinical Modeling: A Single-Institution Experience
- MicroRNA-218 Inhibits Cell Cycle Progression and Promotes Apoptosis in Colon Cancer by Downregulating BMI1 Polycomb Ring Finger Oncogene
- In Vitro and in Vivo Anticancer Effects of American Ginseng Berry: Exploring Representative Compounds