When a drug is applied to cells in a screen, what the cell actually experiences is not the pipetted dose. The compound must cross membranes, survive metabolic enzymes, and accumulate intracellularly before it can shift a single gene's expression. Yet the models we use to predict how individual cells respond to drugs — an autoencoder that disentangles a cell's basal state from its drug, dose, and cell-type context — represent each molecule only as a generic structural fingerprint, with no notion of whether that molecule can physically get into a hepatocyte or survive a chondrocyte's metabolic machinery. In parallel, the drug-likeness community has built molecular representations that explicitly encode this pharmacokinetic viability: a drug-likeness model trained sequentially on absorption, then distribution, then metabolism, then excretion tasks learns embeddings that carry systemic-viability context and improve downstream classification by up to +18.2% across molecular foundation models (ADME-drug-likeness: enriching molecular foundation models via pharmacokinetics-guided multi-task learning for drug-likeness prediction). These two lines of work have never touched. The result is that single-cell perturbation models systematically confuse "the cell was not responsive to this drug" with "this drug never reached the cell."
This proposal builds a perturbation predictor for single-cell RNA sequencing data — transcriptomes measured one cell at a time, which reveal rare subpopulations and continuous activation gradients that bulk sequencing averages away (Resolving Cellular Heterogeneity via Single Cell Transcriptomics) — whose chemical encoder is a pharmacokinetics-aware molecular representation, and whose training and evaluation explicitly include functional pathway terms: instead of scoring a prediction gene by gene, we ask whether the predicted expression shift activates the same coordinated biological programs (Gene Ontology and KEGG gene sets) that the measured shift activates, in each cell type separately. The central hypothesis is that pharmacokinetic structure is a major, currently unused source of signal for predicting which cellular subpopulations respond to which compounds.
Core Claim
The core mechanism has two coupled parts. First, a learned bioavailability correction: we replace the applied dose $d$ in the response model with an effective intracellular dose $d_{\text{eff}} = d \cdot g(z_{\text{PK}})$, where $z_{\text{PK}}$ is a vector of predicted pharmacokinetic properties for the compound (Caco-2 permeability, microsomal stability, solubility, and related absorption–distribution–metabolism–excretion endpoints, predicted by models trained on the 22 ADMET datasets in Therapeutics Data Commons) and $g$ is a learned, cell-type-conditioned monotone scalar. This encodes the physical fact that uptake and local metabolism govern the exposure a cell actually sees, and it makes dose–response extrapolation pharmacologically grounded rather than purely statistical. Second, a pharmacokinetics-enriched encoder: the drug embedding fed to the response decoder is the representation from a model trained sequentially along the pharmacokinetic lifecycle, in the style of ADME-drug-likeness: enriching molecular foundation models via pharmacokinetics-guided multi-task learning for drug-likeness prediction, which showed that sequential $A \to D \to M \to E$ training with gradient-conflict mitigation captures upstream-to-downstream biological flow that independent or simultaneous multi-task training does not.
The response model itself follows the compositional perturbation autoencoder design of Predicting cellular responses to complex perturbations in high-throughput screens, whose chemCPA variant already accepts chemical structure encoders for unseen drugs — our change is precisely what that encoder carries and how dose enters it. The evaluation is new in kind. Existing chemical perturbation models are scored on gene-level reconstruction ($R^2$ over differentially expressed genes; chemCPA reports 0.62 for drugs entirely excluded from training). We add a functional concordance metric: run overrepresentation-style enrichment (the standard technique that asks whether a gene list overlaps a known pathway more than chance would predict, formalized by Fisher's exact test) on the predicted and measured response gene lists per cell type, and measure agreement between the two enriched Gene Ontology / KEGG program profiles. This choice is motivated by the benchmark of Multimodal Benchmarking of Foundation Model Representations for Cellular Perturbation Response Prediction, which found that models struggling at individual-gene expression prediction can still capture higher-order functional gene-set activation, and that curated biological networks outperform raw expression models on Gene Ontology prediction — gene-level and pathway-level competence genuinely dissociate, so measuring only the former risks accepting models that are biologically incoherent and rejecting models that are biologically right but gene-level noisy.
Related Work and What It Does Not Do
Three existing works sit closest and define the gap. The chemCPA extension of Predicting cellular responses to complex perturbations in high-throughput screens encodes chemical structure for out-of-distribution drug prediction, but uses basic chemoinformatic fingerprints or generic graph encoders, treats dose as a raw covariate, and evaluates only on global gene reconstruction error — no pharmacokinetic content, no pathway-level evaluation. The conjoint DDR2 study Conjoint analysis of single-cell sequencing and high-throughput virtual screening regarding DDR2 in osteoarthritis disease models and the morusin network-pharmacology study Integrated Network Pharmacology, Single‐Cell Transcriptomics Unveil the Mechanistic Role of Morusin in Aortic Dissection both connect single-cell sequencing with drug-likeness filtering, but strictly sequentially and one-way: single-cell data identifies an overexpressed target, then a generic virtual-screening pipeline with static ADMET heuristics (TOPKAT, Discovery Studio, Lipinski rules) filters compounds; the single-cell data never informs the drug-likeness criteria and the drug's effect on the cellular landscape is never modeled computationally. Transcriptomics-Based Drug Repurposing Approach Identifies Novel Drugs against Sorafenib-Resistant Hepatocellular Carcinoma connects perturbation signatures with pathway analysis but only in bulk, losing subpopulation heterogeneity, and treats drug-likeness as a discrete FDA-approval filter. We differ from all three in direction and mechanism: single-cell response prediction is the object, pharmacokinetics is a learned, continuous input rather than a post-hoc binary filter, and functional enrichment is a training and evaluation objective rather than a descriptive afterthought. One prior result constrains rather than motivates us: Multimodal Benchmarking of Foundation Model Representations for Cellular Perturbation Response Prediction shows large expression foundation models often fail to beat a "No Change" baseline, so we treat heavy generative machinery as an open empirical question, not a default.
Research Program
The work proceeds in three stages. Stage one builds the pharmacokinetic encoder: pre-train a molecular model on the 22 ADMET datasets of Therapeutics Data Commons — the shared benchmark for absorption, permeability, metabolic, and toxicity endpoints, which also supplies the scaffold and temporal splits needed to avoid the generalization gap that random splits create (Therapeutics Data Commons: Machine Learning Datasets and Tasks for Drug Discovery and Development) — then train sequentially along $A \to D \to M \to E$ with gradient-conflict mitigation, following ADME-drug-likeness: enriching molecular foundation models via pharmacokinetics-guided multi-task learning for drug-likeness prediction. For reference on this pre-training task, SMILES-Mamba — a state-space sequence model processing chemical strings with linear cost — pre-trained on 250,000 unlabeled ZINC molecules and achieved the highest score in 14 of 22 ADMET tasks and top-two in 17 (SMILES-Mamba: Chemical Mamba Foundation Models for Drug ADMET Prediction); we will use it as the ADMET-pre-training backbone where its linear-time scaling helps, since the encoder must be applied to hundreds of screened compounds cheaply.
Stage two couples the encoder to the perturbation model. We fine-tune the compositional autoencoder of Predicting cellular responses to complex perturbations in high‐throughput screens on sci-Plex3 — a massively multiplexed single-cell chemical screen profiling 188 compounds across three cancer cell lines and four doses — with three encoder variants trained identically: the pharmacokinetics-sequential embedding with the bioavailability-corrected dose, the same embedding without the dose correction (isolating the encoder's contribution), and the original chemCPA fingerprint encoder (the incumbent). Where the distribution of responses matters — drug responses are often bimodal across subpopulations, a shape that mean-shift models miss — we compare against the neural optimal-transport approach of Learning single-cell perturbation responses using neural optimal transport, which learns a map between untreated and treated cell distributions precisely because single-cell assays destroy the cell they measure and so pre- and post-treatment cells can never be paired.
Stage three is the functional layer. For each held-out drug and cell type, we compute enrichment over Gene Ontology and KEGG gene sets — the curated hierarchies of biological processes and wiring diagrams of signaling and metabolic cascades that are the substrate of functional enrichment analysis (Interpreting Gene Expression via Functional Enrichment Analysis) — on both predicted and measured response gene lists, and score concordance as the overlap of significant programs weighted by enrichment-score correlation. This also becomes a training signal: a differentiable rank-based enrichment surrogate added to the reconstruction loss, so the model is penalized for predicting gene shifts that, taken together, activate no coherent program. Finally, we test the pay-off for screening: on the disease setting of the osteoarthritis study (Conjoint analysis of single-cell sequencing and high-throughput virtual screening regarding DDR2 in osteoarthritis disease models, whose GSE200843 dataset profiles chondrocyte subpopulations), we re-rank its screened ZINC15 compounds by predicted chondrocyte-subpopulation pathway response rather than docking energy alone, and check whether the in-vitro-validated hit stays on top — a cheap retrospective test of whether cell-type-resolved functional prediction changes which compounds a screening pipeline would advance.
Proposed Experiments
The primary evaluation uses sci-Plex3 with drug-holdout splits (entire compounds excluded from training, matching the unseen-drug protocol of chemCPA) and dose-holdout splits (train on low doses, test on high doses, where the bioavailability correction should matter most, since extracellular dose no longer approximates intracellular exposure). Metrics are, first, the incumbent: $R^2$ over differentially expressed genes for unseen drugs, where chemCPA reports 0.62 (Predicting cellular responses to complex perturbations in high‐throughput screens); and second, our functional concordance metric, for which no reported result exists anywhere on Lacuna for any chemical perturbation model — establishing that this evaluation is itself new ground. A secondary experiment transfers to the four-cell-line Perturb-seq benchmark of Multimodal Benchmarking of Foundation Model Representations for Cellular Perturbation Response Prediction to check whether pharmacokinetic embeddings also help on genetic perturbations — they should not, and that null result would confirm the mechanism is genuinely pharmacokinetic rather than a generic embedding-size effect. A third experiment validates the bioavailability mechanism directly: group held-out drugs by predicted permeability and metabolic stability and test whether the baseline's prediction error is systematically larger for low-bioavailability compounds, and whether our correction removes that gradient. ADMET pre-training quality is checked against the DrugMAP drug-likeness benchmarks, where the reference sequential-training method reports up to +2.4% over prior state-of-the-art classifiers and up to +18.2% improvement across molecular foundation models (ADME-drug-likeness: enriching molecular foundation models via pharmacokinetics-guided multi-task learning for drug-likeness prediction); for generalization sanity we adopt TDC's scaffold and temporal splits, since the same benchmark shows out-of-distribution drug–target interaction performance falling from roughly 0.7 to roughly 0.4 under temporal splits (Therapeutics Data Commons: Machine Learning Datasets and Tasks for Drug Discovery and Development).
Baselines
The do-nothing reference point is the "No Change" baseline: predict that the drug produces no transcriptional shift at all. This is not a strawman — in the Perturb-seq benchmark of Multimodal Benchmarking of Foundation Model Representations for Cellular Perturbation Response Prediction, large expression foundation models frequently failed to beat it, and no explicit numerical error value is reported on Lacuna for that comparison, so it must be run, not assumed. Any perturbation model we build must beat this baseline on both the gene-level $R^2$ and the functional concordance metric on the same held-out drugs, or the added machinery is unjustified.
The strongest existing competitor is chemCPA itself, the compositional autoencoder with a chemical structure encoder for unseen-drug prediction, trained and evaluated on the same sci-Plex3 data with the same drug-holdout protocol (Predicting cellular responses to complex perturbations in high‐throughput screens). It reports $R^2$ of 0.62 on differentially expressed genes for entirely excluded drugs, and its parent CPA reports $R^2 = 0.81$ on the Panobinostat + Alvespimycin combination; the 0.62 figure is the number to beat on the unseen-drug split. Fairness requires giving chemCPA the same base architecture capacity, the same tuning budget over hyperparameters, the same training data and splits, and the same early-stopping protocol as our model — the only difference should be the encoder and the dose re-parameterization. Beating it means a higher held-out $R^2$ on excluded drugs with a confidence interval excluding chance, plus higher functional concordance; if the gain appears only on one of the two metrics, we report that honestly as a partial win and analyze which regime it appears in.
Two further baselines isolate the proposal's ingredients. A linear baseline — ridge regression from chemical fingerprints to mean expression change — is required because the TDC benchmark shows expert-crafted descriptors like RDKit2D often match or outperform deep molecular models on ADMET tasks (Therapeutics Data Commons: Machine Learning Datasets and Tasks for Drug Discovery and Development), so simple chemical inputs must be ruled out as sufficient. And the key ablation runs our exact architecture with the pharmacokinetic embedding replaced by the standard chemCPA fingerprint while keeping the bioavailability-corrected dose, and separately with the pharmacokinetic embedding but raw dose: the first isolates the encoder's contribution, the second the dose correction's, and together they tell us whether the claimed mechanism — pharmacokinetic content, not just a differently-shaped embedding — is what carries any improvement.
Why not just…?
Why not just use bulk perturbation databases like LINCS L1000 instead of single-cell screens? LINCS already offers thousands of drug-induced transcriptomic signatures at far lower cost, and bulk connectivity mapping has delivered repurposing candidates, as in Transcriptomics-Based Drug Repurposing Approach Identifies Novel Drugs against Sorafenib-Resistant Hepatocellular Carcinoma. The problem is that bulk averaging destroys exactly what our hypothesis concerns: the same drug produces bimodal responses across subpopulations — the neural optimal-transport study documents MEK/PI3K pathway decoupling in melanoma, where resistant and sensitive subpopulations move in opposite directions and the average shows nothing (Learning single-cell perturbation responses using neural optimal transport) — and cell-type-specific uptake, the core of the bioavailability mechanism, is invisible in a bulk average. The L1000 route could still be a useful cheap pre-training signal, and we will test that as an auxiliary, but it cannot evaluate our claim.
Why not just apply Lipinski or TDC ADMET filters as a post-processing step on candidates? This is how every existing pipeline in the file does it — the DDR2 osteoarthritis study applies Discovery Studio ADMET and TOPKAT filters after docking (Conjoint analysis of single-cell sequencing and high-throughput virtual screening regarding DDR2 in osteoarthritis disease models), and the morusin study relies on static web-server predictions (Integrated Network Pharmacology, Single‐Cell Transcriptomics Unveil the Mechanistic Role of Morusin in Aortic Dissection). A binary filter discards the continuous, interdependent pharmacokinetic structure that sequential training demonstrably encodes: the ADME-DL result of up to +18.2% embedding improvement shows the information exists and is lost by thresholding (ADME-drug-likeness: enriching molecular foundation models via pharmacokinetics-guided multi-task learning for drug-likeness prediction). More decisively, the filter-then-predict pipeline cannot re-rank two compounds with identical docking profiles but different uptake, and our dose-extrapolation experiment is designed to show a regime (high applied dose, low bioavailability) where the filtered pipeline and ours make measurably different predictions. That said, this alternative is a real threat to the simple version of our claim: if our experiments show that a binary ADMET pre-filter on candidate compounds captures all the achievable gain, the learned integration is not worth its cost, and we will report that if it happens.
Why not just evaluate on $R^2$ like everyone else? Because the file's own evidence says gene-level and function-level competence dissociate: models that struggle to beat "No Change" on expression prediction still capture higher-order Gene Ontology activations, and curated-network models beat expression models on function while not on genes (Multimodal Benchmarking of Foundation Model Representations for Cellular Perturbation Response Prediction). Single-cell transcriptomics is dominated by technical dropout and stochastic noise gene by gene, while the phenotype lives in coordinated programs (Interpreting Gene Expression via Functional Enrichment Analysis), so a model can win on $R^2$ with biologically incoherent predictions. We keep $R^2$ as a co-primary metric precisely so the proposal can fail conventionally; the functional metric is an addition, not a replacement, and if the two metrics rank models identically on sci-Plex3, that itself is a publishable negative finding about what pathway evaluation adds for chemical perturbations.
Decisive Checks
The proposal stands or falls on three outcomes. If the pharmacokinetics-sequential encoder with bioavailability-corrected dose beats chemCPA's held-out 0.62 $R^2$ on excluded drugs and simultaneously improves functional concordance, and if the improvement concentrates where the mechanism predicts it (low-bioavailability compounds, dose extrapolation, cell types whose responses diverge in bulk averages), the hypothesis is supported. If the gain vanishes in the ablation that keeps the architecture but removes pharmacokinetic content, the mechanism claim is refuted even if raw numbers improve. And if single-cell foundation models again fail to beat simple baselines on our tasks — as they did in the Perturb-seq benchmark — that result, paired with the pharmacokinetic ingredient that the benchmark never tested, would sharpen rather than dissolve the open question of what representation actually carries drug response signal.
References
- ADME-drug-likeness: enriching molecular foundation models via pharmacokinetics-guided multi-task learning for drug-likeness prediction
- Deep sequencing reveals cell-type-specific patterns of single-cell transcriptome variation
- Predicting cellular responses to complex perturbations in high‐throughput screens
- Multimodal Benchmarking of Foundation Model Representations for Cellular Perturbation Response Prediction
- Conjoint analysis of single-cell sequencing and high-throughput virtual screening regarding DDR2 in osteoarthritis disease models
- Integrated Network Pharmacology, Single‐Cell Transcriptomics Unveil the Mechanistic Role of Morusin in Aortic Dissection
- Transcriptomics-Based Drug Repurposing Approach Identifies Novel Drugs against Sorafenib-Resistant Hepatocellular Carcinoma
- Therapeutics Data Commons: Machine Learning Datasets and Tasks for Drug Discovery and Development
- SMILES-Mamba: Chemical Mamba Foundation Models for Drug ADMET Prediction
- Learning single-cell perturbation responses using neural optimal transport
- Tissue-Specific Transcriptome Analysis Reveals Multiple Responses to Salt Stress in Populus euphratica Seedlings