Drugs that kill cancer cells by tipping the balance of the Bcl-2 protein family — pro-survival proteins (Bcl-2, Bcl-xL, Mcl-1) that act as sponges for pro-death proteins (Bax, Bak, Bim) — work beautifully in plastic dishes and then disappoint in animals and patients. One reason is measurement: we predict response from bulk RNA sequencing of cell lines, which tells us how much mRNA a population averages over, while the actual decision to die is made cell by cell from protein counts that mRNA levels only loosely track. Another reason is environment: hypoxic tumor regions actively push the Bcl-2 family balance toward survival, resetting the death threshold that in vitro assays never saw (Modulating the Stoichiometric Balance of Bcl-2 Family Proteins).
The proposal here is to treat the single-cell distribution of Bcl-2 family protein ratios as a deliberately chosen measurement layer for translation — one that is more stable and more domain-invariant across the in vitro–in vivo divide than transcriptomes — and to test that claim head-on. Concretely: build a quantitated intracellular flow cytometry panel that measures Bcl-2, Bcl-xL, Mcl-1, Bax, and Bim as molecules-per-cell in thousands of individual cells, pair it with RNA sequencing of the same samples, and ask whether a model trained on in vitro single-cell ratios predicts in vivo apoptotic response better than the two incumbent approaches: bulk mechanistic modeling of protein ratios, and transcriptome-based transfer learning.
Core Claim
The claim has two parts, one methodological and one empirical. Methodologically, no work in the corpus measures the simultaneously stained, single-cell stoichiometric ratios of the Bcl-2 family: bulk studies use Western blotting of populations, and the flow cytometry deployed on apoptosis is single-channel Annexin V viability gating (Systems analysis of the "weights" of Bcl-2 and Mcl-1 in mitochondrial apoptosis pathway establishes a predictor for best drug combination ratio), while high-parameter flow is reserved for surface lineage markers (A Conserved Landscape of Chemokine Receptor Co-expression Defines the Functional States of CD8+ T Cells in Melanoma). Empirically, we hypothesize that the fraction of cells past the apoptotic threshold — not the population mean — is the quantity that survives the in vitro-to-in vivo transition, because protein measurements are robust where mRNA is fragile (Optimal conditions for quantification of IRF4 mRNA and protein as a correlative biomarker in clinical cancer studies) and because rare resistant sub-populations are exactly what bulk averages erase.
A minimal formalization: for cell $i$, define the free-effector surplus
$$D_i = \alpha\,\mathrm{Bax}_i + \beta\,\mathrm{Bim}_i - \gamma\,\mathrm{Bcl2}_i - \delta\,\mathrm{Bcl\text{-}xL}_i - \epsilon\,\mathrm{Mcl1}_i,$$
where the Greek letters are affinity-like weights fit from data (following the ODE "weights" formulation of Sheng et al., which showed Bcl-2 and Mcl-1 levels and Mcl-1's synthesis/degradation rates are the controlling parameters). The predicted population kill is $K = \Pr_i[D_i > \theta]$, the fraction of single cells whose surplus exceeds threshold $\theta$. The single-cell distribution, not the mean of $D_i$, drives $K$ — and the distribution is what bulk measurements cannot see.
Research Program
Three components, in increasing order of risk.
First, the measurement layer. A spectral flow cytometry panel with intracellular staining for the five Bcl-2 family members plus cleaved caspase-3 and Annexin V as the MOMP-execution readout, calibrated with fluorescence standards so each channel reports molecules per cell rather than arbitrary signal. This is validated against Western blotting of the same lysates, so the panel plugs directly into the parameterization that Sheng et al. established. To handle the protein-turnover problem they flagged — Mcl-1's short half-life dominates response — we measure at a time-course of post-treatment points rather than a single snapshot, estimating turnover at the population level while acknowledging that per-cell degradation rates remain out of reach of a static panel.
Second, the paired modality. For every condition (cell line, drug, dose, time), we collect the single-cell protein panel and bulk RNA sequencing of the same sample, with sample handling in sodium citrate tubes, which the IRF4 study showed preserves both mRNA and surface antigens over 48-hour delays where other anticoagulants fail (Optimal conditions for quantification of IRF4 mRNA and protein as a correlative biomarker in clinical cancer studies). Cell lines are drawn from the CCLE panel Sheng et al. used (starting with MCF-7 and MDA-MB-231, where they demonstrated a combination index of 0.25 for ABT-737 plus A-1210477 at a 1:2 ratio), plus the Kasumi-1 leukemia line used for the 753b dual BCL-xL/BCL-2 degrader (Development of a BCL-xL and BCL-2 dual degrader with improved anti-leukemic activity,). Perturbations span the BH3 mimetic ratios the ODE model optimizes and the degrader, so the model must generalize across mechanism classes.
Third, the model and transfer test. The predictor maps the single-cell ratio distribution to population kill, using linear-attention pooling over the raw point cloud in the style of ReluFormer, which processes up to 500,000-cell flow cytometry clouds without downsampling and whose attention heads spontaneously specialize by lineage (Explainable Techniques for Analyzing Flow Cytometry Cell Transformers). Training is in vitro only; evaluation is in vivo. Xenografts of the same lines — under both normoxic and hypoxia-mimicking conditions, since hypoxia is documented to reset the stoichiometric threshold (Modulating the Stoichiometric Balance of Bcl-2 Family Proteins) — receive the same drugs, and excised tumors are measured by the identical flow panel and RNA-seq. The decisive quantity is how well each representation (single-cell ratios, bulk ratios, transcriptome) trained in vitro predicts the measured in vivo caspase-3-positive fraction in held-out lines.
A secondary deliverable maps systemic safety: the same panel run across blood, spleen, and liver compartments after 753b dosing, following the compartmentalized flow cytometry design of the melanoma chemokine-receptor study, which showed organ-specific immune signatures (R6R5 cells driving liver toxicity) that isolated-Cell assays cannot see (A Conserved Landscape of Chemokine Receptor Co-expression Defines the Functional States of CD8+ T Cells in Melanoma). The 753b platelet-sparing mechanism was validated only in isolated cell-line conditions; whether dual BCL-xL/BCL-2 degradation remodels immune compartments differently across organs is unmapped.
Baselines
The naive reference point is direct carryover of in vitro potency: rank in vivo response by in vitro IC50 or viability score, with no model. This is the implicit default of large screens, and the corpus documents its failure mode — in vitro-validated pro-apoptotic agents neutralized by host systemic countermeasures in vivo, and monolayer assays that never see hypoxic, perfused tissue (Systematic Validation; Systematic Validation). Beating it means any representation achieving significant rank correlation between predicted and measured in vivo kill where IC50 ordering does not.
The strongest mechanistic competitor is the ODE weights model of Sheng et al., which fits bulk Bcl-2/Mcl-1 levels and turnover kinetics to predict drug-combination synergy and dependency (Systems analysis of the "weights" of Bcl-2 and Mcl-1 in mitochondrial apoptosis pathway establishes a predictor for best drug combination ratio). This is the closest existing work and must be run on exactly our data to be a fair comparison: same lines, same drug ratios, same time-course, with its bulk inputs supplied by our calibrated panel averaged over cells (so it gets the best possible bulk measurements, not inferior Westerns), and the same tuning budget on its affinity parameters. Its reported in vitro performance sets the bar: it correctly predicted dependency for 27 out of 30 CCLE cell lines, and its closed-form combination-ratio equations matched ODE simulations with $R^2$ of 0.85 and 0.71. Crucially for us, there is no reported result found on Lacuna for in vivo prediction from that model — the in vivo regime is exactly where it has never been tested. We beat it if the single-cell distribution model predicts in vivo caspase-3 fractions materially better, especially on tumors where bulk ratios agree between in vitro and in vivo but the distributions differ.
The strongest transcriptomic competitor is the transfer-learning family: DeepSADR pre-trains on 225,781 DepMap cell-line samples and fine-tunes on 555 TCGA patient samples, reaching an AUC of 0.957 for Sorafenib response in patients (DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction); CSG²A pre-trains gene-gene attention on ~200,000 LINCS L1000 profiles and achieves an RMSE of 2.119 on drug-blind GDSC prediction, beating the next best model's 2.268 (Transfer learning of condition-specific perturbation in gene interactions improves drug response prediction). We re-implement a CSG²A-style pipeline on our paired RNA-seq (same cell lines, same perturbations, same compute and tuning budget), so the only difference is the input representation. Beating it means higher accuracy on the in vivo apoptotic fraction, the outcome both models are asked to predict. Note that these models predict gross viability or retrospective clinical labels and cannot distinguish apoptosis from other death modes — our ground truth is mode-specific.
The key ablation is the same model with the single-cell distribution collapsed to its population mean, so the input is a bulk ratio vector rather than a point cloud. This isolates the proposal's central ingredient: if the mean-collapsed model matches the full-distribution model in vivo, the single-cell claim is wrong and the honest conclusion is that calibrated bulk ratios suffice. A second ablation replaces protein ratios with mRNA abundances of the same five genes, isolating the measurement-layer claim.
Why not just…?
Why not just do bulk RNA-seq transfer learning, since DeepSADR reaches 0.957 AUC? That number is real and strong, but it is retrospective prediction of clinical response labels from TCGA bulk expression, and the corpus gives three reasons to doubt it transfers to mechanism-specific, prospective prediction. First, transcript-only models collapse on out-of-distribution validation: the TCGA random-forest models reached 86% on training cohorts but fell to 52.9% — near chance — on independent validation for 5-FU (Leveraging TCGA gene expression data to build predictive models for cancer drug response). Second, mRNA and protein discord: IRF4 protein stays stable across collection conditions while its mRNA fluctuates wildly (Optimal conditions for quantification of IRF4 mRNA and protein as a correlative biomarker in clinical cancer studies), and static transcript levels miss Mcl-1's turnover dynamics that Sheng et al. proved control sensitivity. Third, the LINCS L1000 backbone covers only 978 landmark genes, compressing pathway signal (Transfer learning of condition-specific perturbation in gene interactions improves drug response prediction). Our decisive experiment settles this directly: the re-implemented CSG²A-style baseline runs on the identical paired RNA-seq, so if it matches our protein-ratio model on in vivo kill, the RNA route wins and we will report that.
Why not just extend the ODE model with bulk Western blots on xenografts? This is the most serious alternative, because it reuses a validated mechanistic model and adds only in vivo bulk measurement. It could plausibly work, and our ablation is designed to concede gracefully if it does. The case against it is heterogeneity: bulk ratios from two tumors can match while their single-cell distributions differ — one unimodal, one bimodal with a resistant tail — and the kill fraction depends on the tail. The ODE model is also strictly in vitro to date, with no reported in vivo result found on Lacuna, and hypoxia resets the very thresholds its parameters encode (Modulating the Stoichiometric Balance of Bcl-2 Family Proteins). The head-to-head, with the ODE model fed the best possible bulk inputs from our calibrated panel, is the experiment that answers this question rather than an argument.
Why not just measure Annexin V and caspase-3 in vivo directly and skip the predictive model? Because that measures what happened, not what will happen. The translational question is prospective: given an in vitro profile, which drug ratio will kill a given tumor in vivo? Direct in vivo measurement cannot rank untested combinations without running them all in animals, and it gives no portable representation. Our model uses in vivo measurements as ground truth, not as a substitute for prediction. The related cheaper option — running viability screens and trusting IC50 — fails for a further reason: gross viability cannot tell MOMP-driven apoptosis from necrosis or pyroptosis, which are molecularly distinct outcomes with different resistance mechanisms.
Why not just improve the GSEA-style analysis of the RNA-seq we already collect? Because gene-set enrichment is unstable with respect to the reference collection: MSigDB C6 was decisively outperformed, with Wilcoxon p-values as low as $1.24 \times 10^{-33}$, and data-driven collections beat curated ones (Importance of collection in gene set enrichment analysis of drug response in cancer cell lines). Enrichment is a descriptive tool; it does not produce a per-cell, threshold-crossing representation that can be trained in vitro and evaluated in vivo. We will still run it as a sanity check on the paired RNA-seq, but as annotation, not prediction.
Proposed Experiments and Decisive Checks
Experiment 1 validates the panel: calibrated molecules-per-cell measurements against Western blots across the CCLE-derived lines, then a dose-response of ABT-737/A-1210477 at the ODE-predicted optimal ratios, confirming we reproduce the reported combination index of 0.25 in MDA-MB-231 (Systems analysis of the "weights" of Bcl-2 and Mcl-1 in mitochondrial apoptosis pathway establishes a predictor for best drug combination ratio) — if we cannot, the measurement layer is not sound and nothing downstream matters.
Experiment 2 builds the paired in vitro dataset: the five-protein single-cell panel, the time-course for Mcl-1 turnover, and matched RNA-seq, across roughly a dozen lines spanning Bcl-2-dependent, Mcl-1-dependent, and double-dependent genotypes (Sheng et al. showed 27 of 30 CCLE lines are covered by their dependency classes), under BH3 mimetics and the 753b degrader, whose DC50 of 6 nM for BCL-xL and 48 nM for BCL-2 in Kasumi-1 (Development of a BCL-xL and BCL-2 dual degrader with improved anti-leukemic activity,) sets expected effect sizes.
Experiment 3 is the transfer test: xenografts of a held-out subset of lines under normoxic and hypoxia-mimicking conditions, same drugs, same panel and RNA-seq on excised tumors. Primary metric: Spearman correlation between predicted and measured in vivo caspase-3-positive fraction across lines and conditions, for each representation; secondary metric: AUROC for separating responder from non-responder tumors. The decisive checks are three. First, the single-cell distribution model must beat the mean-collapsed ablation specifically on heterogeneous and hypoxic tumors — that is where the distribution carries information the mean cannot. Second, it must beat the RNA-only transfer model on the same data; if RNA wins, the protein-stability argument fails and the field should know that. Third, on the ODE baseline, prediction must hold on tumors whose bulk ratios match their in vitro values but whose distributions have shifted, since those are the cases that distinguish a distribution claim from a bulk claim. The compartment toxicity arm runs in parallel: multi-organ flow after 753b dosing, testing whether platelet-sparing and immune-organ remodeling hold under systemic conditions, with any liver or spleen immune remodeling treated as a finding either way.
References
- The hypoxia-mimetic agent CoCl2 induces chemotherapy resistance in LOVO colorectal cancer cells
- Systems analysis of the "weights" of Bcl-2 and Mcl-1 in mitochondrial apoptosis pathway establishes a predictor for best drug combination ratio
- A Conserved Landscape of Chemokine Receptor Co-expression Defines the Functional States of CD8+ T Cells in Melanoma
- Optimal conditions for quantification of IRF4 mRNA and protein as a correlative biomarker in clinical cancer studies
- Development of a BCL-xL and BCL-2 dual degrader with improved anti-leukemic activity,
- Explainable Techniques for Analyzing Flow Cytometry Cell Transformers
- Effects of miR-135a-5p and miR-141 on proliferation, invasion and apoptosis of colorectal cancer SW620 cells
- Aldehyde dehydrogenase activity is a cancer stem cell marker of tongue squamous cell carcinoma
- DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction
- Transfer learning of condition-specific perturbation in gene interactions improves drug response prediction
- Leveraging TCGA gene expression data to build predictive models for cancer drug response
- Importance of collection in gene set enrichment analysis of drug response in cancer cell lines