Herbal formulas are said to work through "division of labor": several compounds each repair a different part of a disease, and the whole formula outperforms any part. The evidence for this comes almost entirely from bulk measurements of whole tissue. Such measurements average over every cell type, so they cannot say whether two components cooperate inside the same cell, act separately on different cells, or simply change how much of each other reaches the blood. We propose a full factorial experiment on a formula whose division of labor is already documented. Every subset of its four components is given to diseased animals, and the heart is profiled at single-cell resolution. A statistical model then splits the formula's effect, cell type by cell type, into the separate contribution of each component plus whatever is left over (the interaction).
The key test separates two kinds of interaction. If one component boosts the others by raising their absorption (a pharmacokinetic interaction), it should scale the other components' effects by roughly the same factor in every cell type. If components cooperate on the cell's own machinery (a pharmacodynamic synergy), the extra effect should appear only in particular cell types and gene programs. Bulk assays cannot tell these apart. Single-cell profiling of all subsets can.
Core Claim
The formula is QiShenYiQi Pills. Yuanyuan Chen et al. studied it in rats with heart enlargement caused by narrowing of the ascending aorta. They report that astragaloside IV and notoginsenoside R1 restore fatty-acid energy metabolism, that 3,4-dihydroxy-phenyl lactic acid counters oxidative stress, and that Dalbergia odorifera acts as an adjuvant: it has little activity alone, but the full formula beats every single or partial combination. That study used gel-based proteomics of whole heart tissue, so it cannot say which cells each component reaches.
We claim three things can be resolved per cell type: the main effect of each component, the interactions among components, and changes in how common each cell type is. We further claim that the adjuvant's contribution will take the form of a single, cell-type-independent gain on the other components' effects, which is the signature of an absorption effect. This direction is suggested by Linmei Pan et al., who showed that co-ingredients increase intestinal transport of active compounds, and by Dan Wu et al., who argue that the gut and first-pass metabolism act as gatekeepers for which compounds reach tissue. Any synergy that remains after accounting for this gain is a candidate for true cooperation inside cells. The model localizes it to specific cell types.
Relation to prior work
The closest existing study is the morusin aortic dissection study. It predicted hub targets with network pharmacology, then checked which cell types express those targets in an untreated human single-cell atlas (GEO GSE213740). The drug was never sequenced, and it was a single compound rather than a mixture. We sequence treated tissue and decompose a mixture, which is a different question: what the drug does to each cell type, not where its predicted targets happen to be expressed.
Combination models supply part of the method and also serve as competitors. CPA represents each drug as a vector in a learned compressed space and models a combination by adding the vectors. chemCPA adds a molecular-structure encoder and pretraining on the large L1000 bulk drug-response dataset. Both were tested only on single drugs or pairs of synthetic drugs. We test whether additive composition holds for a four-component formula, and whether it fails in a specific pattern.
CellOT learns a map that moves each untreated cell's profile to where a treated cell would be (an optimal-transport map). It handles responses where cells split into responders and non-responders, but it treats each drug as a separate condition and has no notion of components. Our decomposition can sit on top of any of these response models.
Two resources supply evaluation tools. scPerturb provides the E-distance, a measure of how far apart two clouds of cells are, along with a permutation test. It also sets a design floor: at least 300 cells per condition and 1,000 counts per cell are needed to tell a perturbation from noise. We apply that floor per cell type. The OP3 benchmark showed that models do much worse on genes that actually change than on genes that stay flat. We therefore report every score on the changed genes.
Network pharmacology studies such as Renyuan Zheng et al. motivate the work. They produce hub-target lists from static databases with no cell resolution, and we score such lists against measured cell-type responses.
Research Program
The experiment has 16 arms: every subset of the four components, from vehicle alone to the full formula, all given to rats with surgically narrowed aortas, plus sham-operated controls. Doses match Chen et al., and treatment lasts one month. Each arm contains at least four animals so that animal-to-animal variation can be estimated, and cardiac function is measured by echocardiography as in the original study.
Plasma levels of astragaloside IV, R1 and 3,4-dihydroxy-phenyl lactic acid are measured by liquid chromatography–mass spectrometry in every arm. This gives an independent readout of absorption interactions.
Hearts are profiled by single-nucleus sequencing, because adult heart muscle cells are too large for droplet capture of whole cells. We pool nuclei from several animals in one run and later assign each nucleus back to its animal using genetic or barcode labels (sample multiplexing). Sequencing depth is set so that each major cell type has at least 300 nuclei per arm: muscle cells, fibroblasts, endothelial cells, macrophages and others.
For cell type $c$, gene $g$ and component subset $s \subseteq \{A, R, D, O\}$ (astragaloside IV, R1, the lactic-acid compound, Dalbergia), we first compute per-animal average expression within that cell type (pseudobulk). We then fit a sparse factorial model:
$$y_{cg}(s) = \mu_{cg} + \sum_{k \in s} \beta_{cgk} + \sum_{\{k,l\} \subseteq s} \gamma_{cgkl} + \varepsilon,$$
Here $\mu_{cg}$ is the vehicle-treated disease level, $\beta_{cgk}$ is the main effect of component $k$, $\gamma_{cgkl}$ is the pairwise interaction, and $\varepsilon$ captures animal-level noise. Higher-order terms are kept as a residual. Changes in cell-type proportions are fitted separately, with the same factorial structure, as a compositional model on cell counts. This prevents a change in how many cells of a type exist from being mistaken for a change in what those cells express.
The pharmacokinetic hypothesis is a restricted version of this model, in which the adjuvant only multiplies the other components' effects by one shared factor $\alpha$:
$$y_{cg}(s) = \mu_{cg} + \big(1 + \alpha\, \mathbb{1}[O \in s]\big) \sum_{k \in s \setminus O} \beta_{cgk}.$$
We compare the free model and the gain model by held-out likelihood across animals. A single $\alpha$ that fits all cell types, and that agrees with the measured rise in plasma levels, supports the absorption explanation. A cell-type-specific $\alpha_c$, or a large leftover $\gamma$, marks pharmacodynamic synergy. That synergy is then localized by ranking cell types by the E-distance between the observed full-formula cells and the cells predicted by the additive model.
Before spending the animal budget, we validate the estimator on Combosciplex. That dataset contains single-cell profiles of drugs given alone and in pairs, so the interaction term can be checked on combinations whose behavior is already measured.
Baselines
The do-nothing reference predicts the full-formula response as the vehicle-treated disease state. A slightly stronger version predicts it as the best single component. These set the floor and show how much of the response any decomposition must explain.
The additive cell-type baseline is our model with all interaction terms fixed at zero. Comparing it with the full model is the ablation that isolates the central question, which is whether interaction terms carry real signal and in which cell types.
The strongest existing method is CPA, which composes components by adding their vectors in a learned space. We train it on the 15 non-full arms and ask it to predict the held-out full formula. It gets the same cells, the same cell-type labels, and the same hyperparameter-search budget as our model, with tuning done by leave-one-animal-out validation. The strongest reported number for this kind of task is CPA's coefficient of determination of 0.81 on the unseen Panobinostat + Alvespimycin pair in Combosciplex (CPA). For structure-aware prediction, chemCPA reports 0.79 on changed genes for sci-Plex3 at 10 µM (chemCPA). chemCPA can only be applied to the three purified components, because Dalbergia is a crude extract with no single chemical structure. That limitation is itself informative. No reported result was found on Lacuna for any model predicting a multi-component herbal formula. Beating CPA means a higher coefficient of determination on held-out full-formula changed genes, averaged across cell types, with no cell type in which CPA wins by a significant margin across animals.
A pseudobulk baseline fits the same factorial model to whole-tissue averages built from our own data. It tests whether cell resolution adds anything, using identical animals.
The network-pharmacology baseline follows the morusin pipeline. It predicts hub targets for the formula and assigns them to cell types using the untreated atlas. We score it by how well its predicted targeted cell types match the cell types where we measure main effects.
Why not just…?
Why not just run bulk RNA-seq on the 16 arms? This is the most serious alternative, because it is far cheaper. Bulk data cannot separate a change in cell proportions from a change in expression within a cell type. Tangjuan Zhang et al. showed that proportion changes in fibrotic tissue run in opposite directions across closely related subtypes. The pseudobulk baseline settles the question directly on our data. If it recovers the same interaction structure, and the same evidence for a single gain factor, as the cell-resolved model, then bulk is enough. In that case we would recommend bulk factorial designs in place of single-cell ones.
Why not just run CPA or chemCPA on the formula? Those models assume combinations add up in their learned space, which is the very hypothesis under test. Running them unchanged cannot tell us where additivity fails. CPA is included as the main competitor. If it predicts the full formula as well as our model does, the honest conclusion is that this formula behaves additively, which is also a publishable answer.
Why not just profile the full formula against vehicle? A two-arm experiment shows what the formula does but cannot attribute any effect to a component. Division of labor is a claim about attribution, so it needs the subsets.
Why not rely on network pharmacology and docking? These methods assume every compound reaches tissue in usable amounts, and the files' pharmacokinetic studies (Dan Wu et al., Linmei Pan et al.) argue against that. Their cell-type predictions are scored directly in our experiments.
Decisive Checks
On Combosciplex, the interaction estimator must flag the pairs whose measured combined response differs from the sum of the single-drug responses, while leaving the additive pairs at zero. Otherwise the method is not ready for the animal study.
In the rat study, the absorption hypothesis is confirmed if three things line up: the gain model fits held-out animals as well as the free model, the fitted $\alpha$ agrees with the plasma-level ratios measured with and without Dalbergia, and $\alpha$ does not vary by cell type. It is refuted if the adjuvant's effect is concentrated in a few cell types even after correcting for plasma levels.
Division of labor is supported if the main effects of astragaloside IV with R1 load mainly on energy-metabolism programs in heart muscle cells, while the lactic-acid compound's main effect loads on oxidative-stress programs in a different set of cells. If both act on the same cells with large interaction terms, the division-of-labor account in Chen et al. is wrong at the cellular level.
Finally, the full formula's advantage on echocardiography must be explained by the cell types the model flags. We check this by regressing per-animal ejection fraction on the per-cell-type response scores.
References
- QiShenYiQi Pills, a compound in Chinese medicine, protects against pressure overload-induced cardiac hypertrophy through a multi-component and multi-target mode
- Exploratory Pharmacokinetics of Geniposide in Rat Model of Cerebral Ischemia Orally Administered with or without Baicalin and/or Berberine
- UPLC-PDA determination of paeoniflorin in rat plasma following the oral administration of Radix Paeoniae Alba and its effects on rats with collagen-induced arthritis
- Integrated Network Pharmacology, Single‐Cell Transcriptomics Unveil the Mechanistic Role of Morusin in Aortic Dissection
- Predicting cellular responses to complex perturbations in high‐throughput screens
- Predicting Cellular Responses to Novel Drug Perturbations at a Single-Cell Resolution
- Learning single-cell perturbation responses using neural optimal transport
- scPerturb: Harmonized Single-Cell Perturbation Data
- A benchmark for prediction of transcriptomic responses to chemical perturbations across cell types
- Molecular mechanisms of Huanglian Jiedu decoction in treating Alzheimer’s disease by regulating microbiome via network pharmacology and molecular docking analysis
- Single Cell <scp>RNA</scp>‐Seq Identifies Cell Subpopulations Contributing to Idiopathic Pulmonary Fibrosis in Humans