Experimental design

Accurate deconvolution of functional readouts requires reliable estimation of cellular composition. We therefore developed CLIMB, a signature matrix–free deconvolution method that infers cell subtype proportions and expression profiles from bulk RNA-seq data using a single-cell RNA-seq reference dataset (Figs 1A, S1, S2). CLIMB models each bulk transcriptome as a non-negative linear combination of individual single-cell profiles, yielding cell-level coefficients that map bulk samples to single cells and leverage the full intercellular variability of the reference dataset. Cell-type proportions are obtained by aggregating these coefficients by annotation, with reference-induced cellular imbalance biases mitigated through an iterative Empirical Bayes sub-sampling procedure. This bulk-to-single-cell mapping further enables the estimation of cell-type–specific expression by combining reference-based expression with learned residual components that capture expression absent from the single-cell reference. As a clinically relevant application, we used CLIMB to estimate leukemic stem cell fractions in bulk samples from patients with acute myeloid leukemia.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Experimental Design.

(A) Overview of CLIMB. CLIMB deconvolves bulk RNA-seq data by mapping each bulk transcriptome to individual single cells from a reference scRNA-seq dataset, using iterative Empirical Bayes sub-sampling to reduce reference bias, and enables estimation of cell subtype proportions and cell-type–specific expression. (B) Overview of CLIFF. Building on CLIMB output, CLIFF is a probabilistic model that infers cell-type–specific drug sensitivities from bulk survival measurements by modeling observed responses as weighted mixtures of latent cell-subtype survival probabilities, with parameters estimated via an Expectation–Maximization algorithm. (C) Framework overview. Schematic of the two-step CLIMB–CLIFF framework, illustrating the links between input data, learned variables, and outputs across both methods.

https://doi.org/10.1371/journal.pcbi.1014862.g001

Building on the output of CLIMB, we developed CLIFF, a probabilistic model that infers cell-type–specific drug sensitivities from bulk drug response measurements by modeling observed survival rates as weighted mixtures of latent cell-subtype survival probabilities, using known cell-type proportions (Figs 1B, S1). Cell-subtype drug sensitivity is linked to gene expression through a sigmoid model, and parameters are estimated via an Expectation–Maximization algorithm. By integrating CLIMB-derived single-cell coefficients, CLIFF can be naturally extended to single-cell resolution (CLIFF-SC), enabling inference of drug sensitivity at the level of individual cells without requiring direct single-cell perturbation data. The framework can optionally incorporate somatic mutations as covariates and supports a high-resolution mode in which drug sensitivity is estimated within each bulk sample (S1B Fig).

This framework was evaluated using both experimental and computational strategies. First, we generated mixtures of human leukemic cell lines, providing a controlled system with known cell proportions and experimentally measurable cell line–specific drug sensitivities. Second, we applied CLIMB and CLIFF to bulk RNA-seq and ex-vivo drug sensitivity data from the beatAML cohort, enabling translation of bulk drug response measurements into cell subtype–resolved profiles.

Evaluating bulk deconvolution methods for predicting cellular composition in cancer cell line mixtures

CLIMB infers cell-subtype proportions and can thus predict the composition of a cell mix (Fig 2A). To test its efficiency and compare it with published methods, we generated in-vitro pools containing various ratios of HL60, SUDHL4, K562, and THP1 human leukemic cell lines. We additionally performed single-cell RNA sequencing of an equimolar mixture of the four cell lines and used this dataset as a reference for all deconvolution methods, applying the same cell and gene selection criteria (employing genes with higher estimated biological variance than technical variance; see Methods). We could thus use these pre-determined ratios as ground truth to compare the predictive value of CLIMB (Fig 2B, 2C) and a panel of seven bulk deconvolution methods: BayesPrism [5], CiberSortX [6], TAPE [7], Scaden [8], MuSiC [9], BisqueRNA [36], and NNLS [9], using random proportions as a negative control. The overall ranking across four performance metrics–-root mean square error (RMSE; lower values indicate better performance), Pearson correlation coefficient (PCC), Spearman correlation coefficient (SCC), and coefficient of determination (R-square)–-was computed by averaging the ranks obtained for each metric (mean ± standard deviation). This analysis identified CLIMB and Scaden as the top-performing deconvolution methods (Fig 2D), followed by BayesPrism. Notably, CLIMB reached a higher accuracy than other deconvolution methods based on the root mean square error (RMSE, Fig 2E). Scaden reached similar PCC, SCC, and R-squared, than CLIMB (Fig 2C–2F). A wilcoxon test taking four metrics into account showed significant advantage of CLIMB over other deconvolution methods, except Scaden. In summary, CLIMB demonstrated top-ranking accuracy in deconvoluting bulk RNA-seq samples from in-vitro cell mixtures, matching the performance of the deep learning-based method Scaden and surpassing other deconvolution methods.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Cell Line Abundance Deconvolution using an In-Vitro Experiment with Known Proportions.

(A) Experimental design for in-vitro experiment involving mixtures of human leukemic cell lines. (B) Heatmap displaying predicted cell line proportions by CLIMB (left) and actual proportions (ground truth, right). (C) Scatter plot comparing true versus predicted cell line abundances using CLIMB (left) and Scaden (right), with colors representing cell lines (6 bulk samples and 4 cell lines). (D) Overall results ranked based on median of four metrics (PCC, SCC, RMSE, and R-square) for bulk sample-level comparison. Wilcoxon rank test used for statistical comparison between CLIMB and other methods. (E) Left: Comparison of deconvolution accuracy using Pearson correlation coefficient (PCC) between predicted and true cell line proportions. Statistical significance assessed using Pearson and Filon method with cocor R package. Left boxplot depicts PCC per sample (4 true versus 4 predicted cell line abundances for 6 bulk samples). Right: same approach using Spearman Correlation Coefficient. (F) Left: Similar to (E) but for R² metric. Right: similar to (E) but for root mean square error (RMSE) metric.

https://doi.org/10.1371/journal.pcbi.1014862.g002

Evaluating non-negative least squares performance relative to CLIMB in bulk and pseudo-bulk settings

Although non-negative least squares (NNLS) is expected to perform well for deconvolution, it did not outperform random proportions in the in-vitro cell mixture experiments (Fig 2). To investigate this unexpected result, we generated six pseudo-bulk samples from the same scRNA-seq reference dataset, using exactly the same cell proportions as the six real bulk RNA-seq samples obtained from experimental cell mixes. Half of the cells were used to construct the pseudo-bulk samples, while the remaining cells were retained as the reference. Deconvolution of these pseudo-bulk samples resulted in high accuracy (>95%) for CLIMB and for two independent NNLS implementations (from MuSiC library or with a custom NNLS implementation based on average cell-type expression), with CLIMB performing marginally better (S9A Fig). In contrast, when deconvolving real bulk RNA-seq samples using the same scRNA-seq reference dataset, CLIMB retained high accuracy whereas both NNLS implementations performed poorly (S9B Fig). This difference was observed specifically when reference and target data were generated using different technologies. Consistently, expression correlations between average cell-type profiles and pseudo-bulk samples were substantially higher than those observed with real bulk RNA-seq samples (S9C Fig), and pseudo-bulk and real bulk samples themselves showed weak expression correlation (S9D Fig). Together, these results show that NNLS performance is sensitive to cross-platform discrepancies between scRNA-seq–derived reference profiles and real bulk RNA-seq data, whereas CLIMB maintains stable accuracy when applied to both pseudo-bulk and real bulk RNA-seq samples.

Evaluating potential bias towards cellular proportions contained in reference dataset

The deconvolution process of CLIMB, which employs the entire single-cell matrix as input, may introduce bias mirroring the cellular proportions in the reference matrix. To mitigate such bias, CLIMB’s model executes three sequential deconvolutions. After deconvoluting cellular proportions during its first iteration, CLIMB subsets cells in the reference matrix based on these proportions, which subsequently serve as a reference for the next deconvolution. This process is iterated twice, progressively adjusting the reference matrix to align its cellular proportions with those of the target bulk sample (S2 Fig, Methods). To demonstrate CLIMB’s model accuracy and robustness, we studied how the correlation between cellular proportions in the reference matrix and target bulk samples impacts deconvolution accuracy. We simulated eleven scenarios ranging from strong to zero-correlation contexts, using reference and pseudo-bulk data derived from AML cells in the Van Galen et al. dataset (S3A Fig). We examined accuracy and resilience to induced bias, which we inferred from the variability in accuracy across our range of bias contexts. By comparing CLIMB-corrected (the default cell subtype proportions provided by CLIMB) with the first iteration of CLIMB algorithm (’CLIMB-init’) to a panel of seven deconvolution methods, we observed that CLIMB-corrected was the second most accurate method, and the third less biased method (S3B-S3D Fig). Interestingly, CLIMB-corrected demonstrated enhanced accuracy and resilience to bias over CLIMB-init in most cross-dataset pseudo-bulk analysis, particularly in cases with significant cellular proportion imbalances between reference and target bulk (S3E-S3F Fig). Overall, we demonstrate that CLIMB iterative deconvolution exhibits high accuracy while maintaining low bias.

Benchmarking Bulk Deconvolution Methods using scRNA-seq Derived Pseudo-bulk Samples in a Cross-Dataset Context

To extend our in-vitro findings, we conducted cross-dataset deconvolution by integrating five pairs of scRNA-seq datasets in five cancer types, totaling ten cross-dataset analyses [28,37–44]. These cross-dataset analyses first involved transferring cell-subtype labels between datasets. Then, one dataset served as the reference single-cell dataset while the other was used to generate pseudo-bulk samples (Fig 3A). Importantly, the reference dataset and the pseudo-bulk datasets were constructed using only non-batch-corrected raw counts to preserve a realistic batch effect across datasets. The comprehensive results of this large-scale analysis demonstrated that CLIMB significantly outperformed other deconvolution methods (Fig 3B), based on an aggregated ranking across four evaluation metrics (PCC, SCC, RMSE, and R2). In the context of Acute Myeloid Leukemia (AML), CLIMB outperformed other methods (Fig 3C) in (i) ‘Van Galen> Naldini’ deconvolution, with CLIMB achieving the best PCC, SCC, and R2 (Figs 3D, 3F, S4E), (ii) ‘Naldini> Van Galen’ deconvolution, where CLIMB achieved the highest PCC, RMSE, and R2 (Figs 3E, 3F, S4G), and (iii) ‘200 simulated pseudo-bulks’, a dataset of 200 simulated bulk samples generated from the Van Galen et al. dataset, where CLIMB demonstrated the top scores for all four metrics (S4F Fig). Interestingly, CLIMB demonstrated superior performance in predicting the proportion of leukemic stem cells (‘HSC-like’ in Van Galen dataset) in both analysis with available HSC-like labels (Fig 3G). Moreover, CLIMB ranked first for CRC cancer analysis (S4H-S4J Fig). CiberSortX and CLIMB emerged as the best methods in Melanoma analysis (‘Tirosh> Jerby’ and ‘Jerby> Tirosh’, S4K-S4M Fig). In the Glioblastoma context, CLIMB was the top-performing method both in the ‘10x> SmartSeq2’ and the ‘SmartSeq2> 10x’ analysis (S4N-S4P Fig). In breast cancer context, MuSiC performed best, based on ‘Gray> Wu’ and ‘Wu> Gray’ analyses (S4Q-S4S Fig). We also analyzed a nasal swab dataset of healthy and COVID-19 samples [45], using 29 patients as an scRNA-seq reference and the remainder to generate pseudo-bulk samples. In this disease context, CLIMB showed the best deconvolution performance (S9E Fig). Overall, these findings highlight CLIMB’s top-ranking accuracy in cross-dataset pseudo-bulk analyses across diverse disease and cancer contexts.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Cell Subtype Composition Deconvolution in Acute Myeloid Leukemia.

(A) Experimental design outline for cross-dataset deconvolution analysis. (B) Average rank across four metrics (PCC, SCC, R2, and RMSE) for 10 cross-dataset pseudo-bulk deconvolutions, including in-vitro mixes and 200 simulated AML samples, totaling 12 analyses. (C) Subset of (B), focusing on the four AML context analyses. (D) Pseudo-bulk analysis using Van Galen dataset as a reference for Naldini’s pseudo-bulks. Boxplots depict PCC/RMSE values for each pseudo-bulk sample (13 samples, 19 cell subtypes’ predictions). (E) Similar to (D), but in the Naldini > Van Galen context, with 16 data points for the 16 samples in the Van Galen dataset. (F) Scatter plot comparing true proportions with deconvoluted proportions using CLIMB (left) and Scaden (right) for both Van Galen > Naldini and Naldini > Van Galen analyses. (G) Predictive accuracy of HSC-like cell subtype proportions for Van Galen > Naldini analysis, comparing 13 predicted proportions with true proportions. (H) Inferred cell-subtype proportions using CLIMB for bone marrow in BeatAML, Leucegene, and TCGA-AML cohorts (N = 628), along with scores of seven signatures (LSC17 and six Van Galen signatures) and FAB labels. (I) HSC-like fraction vs known signatures.

https://doi.org/10.1371/journal.pcbi.1014862.g003

Bulk deconvolution of cell subtype abundance across three aml patient cohorts

We performed deconvolution of cell subtype proportions using bulk RNA-seq samples from three AML cohorts: beatAML [12], Leucegene [46], and TCGA-AML [47], using as a reference only cells in pre-treatment conditions from Van Galen dataset [28]. Employing CLIMB, we delineated a landscape of cell subtype proportions in AML (Fig 3H). To assess deconvolution accuracy in patient samples, we juxtaposed CLIMB-derived proportions with FAB classification [48], a scheme categorizing patients according to blast (cancer cell) differentiation patterns. Specifically, blasts from M0/M1 patients, characterized by minimal differentiation, were expected to correlate with higher proportions of HSC-like and/or Progenitor-like cancer cells, a correspondence verified by CLIMB-derived proportions (Figs 3H, S5A). Similarly, M3 samples (PML-RARA classified patients) were anticipated to exhibit enrichment for GMP-like cells, in line with our deconvoluted GMP-like proportions (S5A Fig). Moreover, we observed monocyte-like cancer cells to be enriched in M4/M5 patients, along with erythrocyte enrichment in M7 patients, both of which were expected (S5A Fig). To evaluate CLIMB’s accuracy in predicting leukemic stem cells, we correlated deconvoluted HSC-like proportions with the LSC17 signature [35]. As expected, the highest correlations were observed with Prog-like and HSC-like cell subtypes (S5B Fig). Interestingly, proportions of HSC-like cells predicted by CLIMB showed a noticeable alignment with both the HSC-like signature established by Van Galen et al. [28] (S5C Fig) and the LSC17 signature, when compared to other deconvolution methods (Fig 3I). To evaluate the consistency of results across different deconvolution algorithms, we assessed cross-method correlations across all three AML cohorts. We observed that CLIMB, Scaden, and TAPE formed a distinct cluster (S5D Fig). In summation, CLIMB provides a robust depiction of the cellular landscape of AML patients across three cohorts, laying foundations for deconvoluting cell subtype drug sensitivity.

Benchmarking bulk expression deconvolution into cell subtype expression and simulated over-expression

CLIMB enables the deconvolution of bulk-level RNA expression into cell subtype-specific expression. This can be achieved through either an ‘overall’ approach, which accounts for expression averaged across all bulks used as input, or a ‘high-resolution’ mode that disentangles cell subtype expression within each bulk sample. In our in-vitro experiment with leukemic cell lines, CLIMB’s ‘overall’ deconvoluted expression achieved high accuracy when compared to bulk RNA-seq expression obtained from pure cell lines (PCC of 0.94, S6A-S6C Fig). BayesPrism displayed similar patterns and accuracy to CLIMB, achieving a PCC of 0.92 (S6C Fig). Subsequently, we deconvoluted AML cell-subtype expression based on the ‘Van Galen> Naldini’ cross-dataset framework. Notably, the ‘overall’ mode of CLIMB demonstrated compelling deconvoluted expression patterns for cancer cell markers (PCC of 0.75, S6D-S6F Fig), surpassing the performance of BayesPrism (0.59) and CiberSortX in ‘group mode’. We then subjected CLIMB’s deconvoluted expression to a ‘high-resolution’ test, simulating the over-expression of 20 genes in the HSC-like subtype. This induction was applied to 7 out of 13 pseudo-bulk samples, with varying fold changes. The simulated expression was then compared to true expression values, showing CLIMB’s deconvoluted expression high accuracy at fold change 10 and higher (S6G Fig). To assess the precision in detecting differentially expressed genes (DEGs), we conducted a DE analysis between wild-type (WT) and over-expression (OE) conditions. We employed the AUC-ROC metric to evaluate DEG detection accuracy (S6J-S6K Fig). BayesPrism exhibited the highest accuracy in recovering DEGs followed by CLIMB (S6J Fig). As a positive control, we checked that true simulated expression efficiently recovered DEGs (AUC > 0.95 with fold change > 2, S6J, S6K Fig). We then computed correlations (PCC) between true and deconvoluted expression (S6L Fig), showing BayesPrism as the most accurate (median PCC of 0.76), followed by CLIMB (0.66). In summary, CLIMB demonstrates accuracy in expression deconvolution, particularly in ‘overall’ mode.

Benchmarking drug sensitivity deconvolution with in-vitro cell line mixtures

In this next step, we integrate drug sensitivity data (measured as a cell viability rate) with CLIMB’s bulk deconvolution results. To achieve this, we developed CLIFF, a method that deconvolutes bulk-level cell viability into cell-subtype-specific cell viability (Figs 1, 4A). We generated drug sensitivity data from in-vitro mixtures comprising six combinations of four human leukemia cell lines, as previously depicted in Fig 2 (HL60, SUDHL4, K562, and THP1). These mixtures were exposed to seven drugs commonly employed in AML consolidation therapies (5-Aza-2’-deoxycytidine, cytarabine, imatinib, nilotinib, doxorubicin, rapamycin, mitomycin C). To assess the accuracy of drug sensitivity deconvolution, we determined the ground truth cell viability of the four individual cell lines when exposed individually to the same drug panel, doses, and time points. We initially evaluated the theoretical basis of CLIFF’s model by reconstructing the experimentally-assessed cell mix viability from the experimentally-determined viability of pure cell lines. This was achieved using CLIFF’s model equations (Equations 15, and 16) to correlate both metrics (S7A Fig), which showed a high Pearson’s correlation of 0.88. In the next step, we evaluated CLIFF’s deconvoluted cell viabilities against ground truth, showing a PCC of 0.80 across all 21 conditions (Fig 4B, 4C). When utilizing high-resolution cell-subtype expression as input (‘CLIFF-highres’), the accuracy of CLIFF was diminished (PCC of 0.59, Figs 4C, S7B). Subsequently, we conducted a comparison of CLIFF’s accuracy with the approach employed by Zeng et al. [49], which relies on a correlation between cell subtype proportions and bulk-level drug sensitivity; with Bottomly et al.’s approach [50], wherein the PC1 of cell subtype signatures is correlated with bulk drug sensitivity; and a linear regression using CLIMB proportions as input and bulk-level cell viability as ouput (LinReg). When assessing drugs and timepoints individually, CLIFF exhibited the highest correlation (median PCC of 0.82, Fig 4D), followed by Bottomly (0.80), ‘CLIFF-highres’ (0.68), Zeng (0.68), and LinReg (0.51). Interestingly, CLIFF is the only method that predicts cell viability rather than providing a drug sensitivity score based on correlation metrics, which results in a significantly lower RMSE compared to all other methods, both in ‘overall’ and ‘high-resolution’ modes (Fig 4E). Thus, within a framework comprising six cell mixes encompassing four cell lines, CLIFF exhibited precise deconvolution of drug sensitivity and proved effective to integrate cell subtype proportions estimated via CLIMB.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Deconvolving Cell Line Drug Sensitivity: In-Vitro and Simulated AML Context.

(A) Experimental setup: Known cell line mixtures are deconvoluted to estimate cell abundances, which are integrated with bulk drug sensitivity for predicting individual cell line drug responses. (B) Predicted vs. actual cell line drug sensitivity for multiple timepoints, cell lines, and drugs. (C) Heatmap displays predictive accuracy as PCCs for each drug and timepoint. (D) Method comparison of deconvolution accuracy using PCCs. (E) Method comparison of deconvolution accuracy using RMSE values. (F) CLIFF’s predictive accuracy as PCC, grouped by drugs. (G) Drug sensitivity simulation on 200 pseudo-bulk samples derived from cell subtype sensitivities (H). (H) Simulated cell subtype drug sensitivity pattern for 10 drugs, used for generating bulk drug sensitivity. (I) Predictive accuracy as PCC for CLIFF and other methods, deconvoluting cell subtype sensitivity (G) using bulk drug sensitivity (H). (J) Same as (I), but using RMSE metric. (K) Simulated high-resolution drug sensitivity across pseudo-bulk samples and cell subtypes. (L) Deconvoluted high-resolution drug sensitivity based on bulk samples from (K).

https://doi.org/10.1371/journal.pcbi.1014862.g004

Simulated Ex-Vivo drug sensitivity assay for benchmarking deconvolution of drug sensitivity

We then simulated bone marrow samples from AML patients to generate a benchmarking dataset similar to the beatAML cohort; as the latter lacks ground truth data for testing deconvolution. Using the Van Galen single-cell dataset, we generated 200 pseudo-bulk RNA-seq samples along with their corresponding bulk cell viability data for ten simulated drugs (Fig 4G). To do this, we simulated cell viability for 19 cell subtypes across ten drugs (Fig 4H), which serves as our ground truth for evaluating CLIFF’s deconvolution accuracy. We proceeded to assess the accuracy of drug sensitivity deconvolution using the same five approaches mentioned earlier (Fig 4D–4E). In this context, CLIFF-overall demonstrated the highest accuracy, with a median PCC of 0.82 (Fig 4I), followed by LinReg (0.75), Zeng (0.68), CLIFF-highres (0.54), and Bottomly (0.35). As observed in the in-vitro experiment, CLIFF’s advantage, both in overall and high-resolution modes, lies in generating realistic cell viability values, which translates into significantly lower RMSE values compared to every other methods (Fig 4J). Investigating the effect of the number of input genes on CLIFF’s performance, we found that CLIFF-overall maintained consistent accuracy regardless of the gene subset (S7D Fig), while CLIFF-highres excelled particularly with higher gene counts (>500). Additionally, introducing L2 regularization improved CLIFF-highres predictions, although it had no significant impact on CLIFF-overall predictions (S7D Fig). The impact of input size on CLIFF’s accuracy was examined by varying the number of bulk samples (S7E Fig). Notably, approximately 25 cell mixes proved sufficient for CLIFF to achieve a PCC > 0.6 when deconvoluting 19 cell subtypes, and about 75 input mixtures were required to attain a PCC > 0.7. By simulating patient-specific mutations (S7F Fig) that were associated with heightened drug sensitivity in HSC-like/Prog-like cell subtypes and lowered sensitivity in Mono-like/ProMono-like cell subtypes, we examined how well CLIFF-highres could discern patient-level variability (Fig 4K, 4L). Our observations showed that CLIFF-highres detected shifts in drug sensitivity between the two sample groups in the correct cell subtypes (HSC-like, Prog-like, Mono-like, and ProMono-like). However, it also introduced false positive predictions in cDC-like and GMP-like subtypes. Benchmarking CLIFF on in-vitro and simulated data shows the ‘overall’ mode best predicts average cell-subtype drug sensitivity, so we restrict subsequent analyses to this mode.

Robustness, Identifiability, and Scalability of CLIMB and CLIFF methods

We assembled a compact stress test of CLIMB and CLIFF’s ground-truth accuracy, identifiability, and computational scaling, summarized in seven panels (S10A-S10G Fig). Across 11 ground-truth datasets (in-vitro mixtures and 10 cross-dataset pairs), CLIMB’s empirical-Bayes correction improved accuracy in 9/11 cases (mean RMSE 0.095 to 0.076; S10A Fig), and this benefit extended beyond the hematopoietic tissue in which CLIMB was developed: on a non-hematopoietic dataset (COVID-19 nasal-swab), single-cell resolution alone (props.init vs. NNLS) improved with reference coverage (RMSE 0.109 to 0.096 from 5 to 500 cells/type), while NNLS remained flat (0.123-0.131; S10B Fig). Turning to CLIFF’s downstream sensitivity model, identifiability held through cross-sample cell-type proportion correlation (RMSE 0.15-0.17) before degrading sharply at (RMSE 0.26), consistent with the multicollinearity expected as the proportion matrix loses rank (S10C Fig). This near-collinear regime also made optimization less robust: 25 restarts from randomized starting coefficients converged to distinctly worse optima in a minority of cases at -0.9 but a majority at , though across a sweep of possible starting points, CLIFF’s default choice – starting all coefficients at zero – remained among the best-performing initializations tested (S10D-S10E Fig). On the computational side, runtime scaled approximately linearly with sample count, gene count, reference cell-type count, and reference size; at the largest scale tested (100 samples and 15,000 reference cells together), the combined climb()/cliff() pipeline completed in just over two minutes (128s), which remains practical for the dataset sizes considered here (S10F Fig). Finally, injecting calibrated proportion error matched to each method’s real accuracy showed CLIFF’s sensitivity RMSE rising monotonically from 0.158 (true proportions) to 0.185/0.194/0.231 (CLIMB-corrected/CLIMB-init/NNLS), with a non-zero floor even at perfect proportions reflecting measurement noise (S10G Fig) – underscoring that CLIFF’s accuracy is fundamentally bounded by the quality of the proportions it receives.

Deconvoluting bulk-level drug sensitivity to cell subtype sensitivity using beatAML cohort

Utilizing cell subtype proportions and expression data estimated by CLIMB from the beatAML cohort, we integrated this information with bulk ex-vivo drug sensitivity data using CLIFF, which revealed a heterogeneous landscape of drug sensitivity patterns among distinct cell subtypes (Fig 5A). We assessed the findings based on 14 studies that documented cell-subtype level drug sensitivity and featured 22 pairs of drug-to-cell subtype interactions (S1 Table). CLIFF’s predictions aligned with existing literature for venetoclax [16,17], erlotinib [20], A-674563 [21], AZD1480 [22,23], and crenolanib [24], indicating elevated sensitivity in HSC-like and Prog-like cancer cells. Interestingly, midostaurin [25] and selinexor [26] were predicted to target HSC-like cell subtypes, but not progenitor cancer cells (Prog-like). We also documented drug sources that targeted differentiated blasts. Doramapimod, predicted by CLIFF to target monocyte-like cells, aligns with previous evidence of heightened activity in M4/M5 patients showing expression of CD14 monocyte marker [15]. Similar agreement was observed for trametinib [16] and selumetinib [13], known to target differentiated blasts, and projected by CLIFF to exhibit strong affinity for monocyte-like cancer cells. Bromodomain and Extraterminal Domain Inhibitors (BETi) were also shown by Romine et al. [14] to target differentiated blasts, which coroborates CLIFF’s predictions on BET-inhibitor JQ1, predicted to preferentially target monocyte-like cancer cells. Ruxolitinib [16], however, did not display a clear pattern, as it was predicted to target pro-monocyte-like but not monocyte-like cancer cells. CLIFF’s predictions also indicate a notable resistance of monocyte-like cells to venetoclax, validated by four distinct studies [15–19]. Among the 22 documented instances of cell subtype-specific drug sensitivity, CLIFF demonstrated 16/22 cases with clear concordance (72.7%), 3/22 with partial alignment, and 3/22 in direct contradiction (13.6%). Furthermore, we evaluated the concordance between CLIFF’s predictions and results reported by Bottomly et al. using the same dataset [50], revealing an overall correlation of 0.44 for all drugs (S8A Fig), with stronger correlations observed for established drug-to-cell subtype associations (S8B Fig).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Deconvolving Cell Subtype-Specific Drug Sensitivity in AML ex-vivo Data.

(A) Deconvoluted drug sensitivity across cell subtypes in the beatAML cohort dataset, with dot size indicating AUC of sensitivity and color representing standardized AUC, indicating relative drug sensitivity among cell subtypes for a specific drug. (B) Differentiated (Mono/ProMono/cDC) vs. undifferentiated (HSC/Prog/GMP) myeloid cells’ sensitivity (grouped) on the x-axis and normal vs. cancer myeloid cells’ sensitivity on the y-axis. (C) HSC-like sensitivity score computed from AUC and standardized AUC. (D) Drug sensitivity of HSC-like cells shown on a scatter plot with AUC on the x-axis, standardized AUC on the y-axis, and color indicating the combined score.

https://doi.org/10.1371/journal.pcbi.1014862.g005

Subsequently, we focused on evaluating drug sensitivity in HSC-like cells, ranking drugs according to their potential effectiveness against this particular cell subtype (Fig 5B–5D). Among the top ten drugs, only A-674563 and midostaurin had available validation from the literature (S1 Table). The eight undocumented drugs predicted to be the most effective against HSC-like cells were, in order, SNS-032, flavopiridol, foretinib, dovitinib, bortezomib, quizartinib, Go6976, and RAF265. In summary, the combined use of CLIMB and CLIFF successfully unveiled a comprehensive landscape of drug sensitivity at the cell subtype level in AML, along with the ability to predict drug sensitivity in the clinically relevant HSC-like cell subtype.

Single-cell resolution prediction identifies resistant cancer cell subpopulations

We further developed CLIFF-SC, which computes drug sensitivity prediction at single-cell resolution (Fig 6A, Methods), leveraging coefficients derived by CLIMB for each individual cell. We focused on three drugs predicted to target the HSC-like compartment: venetoclax, A-674563, and foretinib, and one drug with an opposite pattern, nilotinib, which targets monocyte-like cancer cells (Fig 5A). We visualized single-cell survival probability on 2D projections (UMAP) using all cancer cells from Van Galen dataset [28], which showed considerable heterogeneity in distributions across the four drugs and the six cancer cell subtypes (Fig 6B–6E). We observed high concordance between average cell subtype sensitivity predicted by CLIFF-SC and the one predicted by CLIFF-overall (Fig 6F). Interestingly, we observed a long upper tail of seemingly resistant HSC-like cells for venetoclax compound (Fig 6B, right panel, circled in red). We thus decided to study in more detail this population of resistant HSC-like. We performed differential expression analysis between resistant and sensitive HSC-like populations, which were clustered using single-cell venetoclax sensitivity predictions (25% most sensitive versus 25% most resistant, Fig 6G). We could show a downregulation of NPM1 expression and upregulation of ANXA1 and TMSB4XP8 in the 25% with highest predicted resistance to venetoclax. These data suggest that expression of these genes could be predictive of venetoclax sensitivity at single-cell level.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Single-Cell Resolution Drug Sensitivity Prediction.

(A) Framework overview: CLIMB deconvolution generates coefficients for each single cell (left), integrated with ex-vivo drug screening data (DSS) for single-cell survival probabilities (right). (B) Single-cell survival probability for venetoclax, based on UMAP projections of cancer cells from Van Galen et al. [28], highlighting upper tail of resistant HSC-like cells in red circle, and survival probability distributions among six cancer cell subtypes (right). (C) Similar view for A-674563 single-cell drug sensitivity. (D) Similar view for foretinib single-cell drug sensitivity. (E) Similar view for nilotinib single-cell drug sensitivity. (F) CLIFF-overall vs. CLIFF-SC concordance scatter plot with correlation coefficients. (G) Left, volcano plot showing gene expression comparison between venetoclax-resistant (25%) and sensitive cells (25%) with p-value versus log fold change. Right, 3 genes misregulated in venetoclax-resistant population of HSC-like cells.

https://doi.org/10.1371/journal.pcbi.1014862.g006

Gene and cell selection

In our approach to gene subset selection prior to bulk deconvolution, we utilized the Scran::ModelGenVar function on the expression matrix derived from single-cell RNA sequencing (scRNA-seq) data. By applying this methodology, we identified genes with estimated biological variance surpassing technical variance, as indicated by positive biological variance outcomes from ModelGenVar. This strategy eliminated the reliance on arbitrary thresholds for the determination of variable or deferentially expressed genes and was consistently employed across all analyses in this study.

Regarding cell selection, in the context of our in-vitro experiment, we observed cells forming five clusters, with four clusters being distinctly attributed to the four mixed cell lines. The fifth cluster lacked a clear marker expression pattern, exhibited smaller size (only few cells) compared to the other clusters, and displayed elevated mitochondrial read content. Consequently, we excluded this ambiguous cluster from our analysis, while retaining the remaining cells in four clusters that were attributed their known cell line labels according to their marker expression pattern.

Regarding the re-analysis of publicly available scRNA-seq datasets, we harnessed the cell selection information provided in the processed data tables, abstaining from further cell filtering based on specific criteria. In instances where datasets encompassed over 30,000 cells, we subsampled them to reach a count of 30,000 cells, adopting the set.seed(1) option for reproducibility. This enabled us to execute the Seurat integration tool without memory constraints. The datasets subjected to subsampling included Naldini et al., Gray et al., Wu et al., and Khaliq et al.

Prototypical bulk deconvolution model

The goal of a bulk deconvolution model is to fit coefficients corresponding to each cell-subtype k through a linear model. Fitted coefficients directly relates to cell-subtype proportions. Thus, bulk deconvolution solves the following problem:

(1)

with representing a vector of gene expression of length G, for a given mixture n. The matrix S is a G by K matrix containing the average expression for each cell-subtype k and each gene g. To obtain proportions that sum up to 1, the model can either impose a constraint , or the coefficients can be normalized after fitting. In this model, the genes are our observations, and corresponds to genes shared between and S or to a subset of marker genes. Note that every mixture n are resolved with an independent linear model, and that the average expression of cell-subtypes is usually derived from a scRNA-seq that does not contain mixture n. Thus, we assume that . Of note, this simple model do not apply for advanced deconvolution algorithms like CiberSortX, BayesPrism, Scaden, BLADE, or TAPE.

CLIMB: Cell-level Linear model through bulk-to-single-cell deconvolution

We developed the CLIMB method, which proposes an innovative approach to solve a bulk deconvolution problem, by applying a mapping from each bulk sample to each single-cell from a reference dataset. We thus fit the coefficients for each single-cell i in an intial run of CLIMB, and thus we find the best linear combination of single-cell to predict bulk expression . Thus, we use the relation , shown below in more details:

(2)

where the vector represent the bulk expression matrix for a mixture n, C is the raw counts matrix from scRNA-seq, and is the vector of coefficients fitted for each single-cell i for a given mixture n. Thus, CLIMB loss function is defined by:

(3)

Thus, we impose a non-negativity constraint on the coefficients, and no regularization. We optimize coefficients associated with each single-cell i with glmnet R library [54], with the options lower.limits = 0, lambda = 0, and scale = TRUE, intersect = FALSE. We indeed observed no improvements of the deconvolution with neither l₁ nor l₂ regularization.

We then group and normalize coefficients obtained with (3) to obtain the cell-subtype proportions for each cell-subtype k and mixture n:

(4)

With being an indicator variable of length C, equal to 1 whenever cell i is of type k.

Empirical Bayes sub-sampling mitigates bias induced by reference matrix

The equations 2, 3, 4 yield biased estimates of cellular proportions towards the cellular proportions contained in the reference single-cell matrix C. This bias acts as a prior, impacting CLIMB’s predictions. To enhance cellular composition depiction, we’ll correct C by informed cell sub-sampling, creating a new matrix via an Empirical Bayes approach estimating the prior from the data.

We use the initial CLIMB-derived cellular proportions from equations 2 3, 4 to sub-sample half of the cells in reference matrix C based on our predicted biased proportions. Each sub-sample ensures a minimum of 50 cells (or 2% of total cell count) per cell subtype. In total, five sub-sampling are done, and this process is repeated three times. Ultimately, the proportions from the second iteration - and thus third deconvolution - become the final CLIMB output (stored in ‘props.corrected’). CLIMB’s iterative algorithm is summarized in S2 Fig.

Additionally, a gene selection aids CLIMB’s deconvolution. First, dataset-specific top-variable genes are selected based on cell subtype average expression per gene. Genes with significant fold change compared to the average expression and compared to the second top-expressed cell subtype, are chosen (threshold set to 95% and 75% fold changes quantiles). cell subtypes with top expression are recorded, allowing a maximum of 100 genes per cell subtype. This initial gene selection is followed by bulk sample-specific gene selection: Genes are ranked according to their maximum cell subtype expression multiplied by their expression in the target bulk sample, and the top 500 genes are selected. These gene selection are thus bulk sample-specific.

Deconvolution of cell-subtype expression

CLIMB allows the prediction of cancer cell expression in a two step manner. Let first define our target high-resolution expression , which is the cell-subtype-specific and bulk sample-specific expression for each cell-subtype k. We assume that for a given group of samples, the grouped expression relates to the high-resolution expression with the following formula:

(5)

First, we take advantage of the bulk to single-cell mapping to directly derive the cell-subtype expression from the bulk-to-single-cell deconvolution or mapping. We can simply sum up the expression of cells mapped by CLIMB to get a high-resolution cell-subtype-specific and bulk-specific expression :

(6)

with being the coefficient fitted by GLMNET on each single-cell, linking each cell i to each bulk n. denotes the single-cell expression of cell i and gene g, and is an indicator vector equal to 1 if cell i is of type k.

Although this is potentially sufficient in cases where scRNA-seq reference dataset maps all cell states present in bulk samples, it can be insufficient to detect genes deferentially-expressed (DE) at cell-subtype level. As DE genes will likely induce error on the cell-subtype expression deconvolution, we use the error itself to fit our model. We can also make the reasonable assumption that DE genes only originates in cancer cells. Thus, for a given gene g and a given bulk n, we have:

(7)

Thus, we use as input for a model to predict cancer cell-subtype expression. We decompose our error between the error originated from DE genes , and an additional error term . Similar as other methods, we use the information from all bulk samples N to deconvolute the high-resolution cell-subtype specific expression . We consider that only originates from cancer cells C^(C) of type K^(C), thus

(8)

Where is the vector of coefficients we wish to predict. We used Glmnet with options lower.limits = 0, lambda = 0, and scale = TRUE, fitting the following loss function:

Our final prediction of cell-subtype expression for a cancer cell-subtype k, a given bulk n, and a given gene g is:

(9)

Following the assumptions made above, we set to zero for normal cells, which imply that for normal cell subtypes.

CLIFF - Cell fate inference for drug sensitivity deconvolution

Intuitively, CLIFF treats each observed bulk drug-response measurement as a blend of hidden, cell-subtype-specific responses, weighted by how much of each cell subtype is present in the sample (Fig 1B). Because we never observe which individual cells survived a given drug exposure, we treat survival as a latent variable and use an Expectation-Maximization algorithm to resolve it: the E-step asks, given the current estimate of each cell subtype’s drug sensitivity, how much of the observed bulk survival should be attributed to each subtype; the M-step then uses that attribution to re-estimate each subtype’s sensitivity, and the two alternate until the estimates stabilize. The sigmoid link (Eq. 16) simply ensures that whatever combination of expression and mutation covariates the model considers, the resulting survival probability stays within [0,1].

Here we propose to infer the drug sensitivity of individual cell subtypes using only drug sensitivity data observed on mixed cell populations, considering K different cell populations in N mixed samples, with proportions than can be defined as , with and . Drug sensitivity observed in the bulk samples we wish to demix are represented by a vector y^(N), with being the empirical surviving rate of cells, and for sample n. Therefore the target matrix we wish to predict is , and the predicted values of drug sensitivity is assumed to depend on the underlying drug sensitivity of cell subtypes for patient n, and cell-subtype k, represented as a surviving rate of cells.

To model drug sensitivity, we assume that single-cell survival is represented as a binary output that follows a Bernoulli distribution. Thus, every single-cell i in patient n exposed to a drug can either survive () or die (). Thus a cell i has a probability to survive and to die. Let’s say we are interested to know the survival probability , given a drug dose. We could resolve the log-likelihood of the Bernoulli distribution for all the cells of a single sample n:

(10)

It turns out that in our problem, cannot be directly observed. We are indeed only able to observe the overall ratio of cells that survives among all cells for a given patient n, which we define as . It relates to the values with:

(11)

Thus, by multiplying (10) by , if we assume that every cell i share the same average probability , we can re-write (10) and obtain the following log-likelihood:

(12)

As stated above, the core of the problem is to assume different survival probability for our single-cells. To slightly simplify this, we can consider a cell-subtype-specific survival probability. Thus, we consider the survival probability observed in our mixed population of cells , as a mixture of K probabilities , one for each cell subtype. Let introduce , a variable describing the survival probability of cell-subtype k:

(13)

We also define , which corresponds to the expected number of cells surviving in patient n with drug dose d, and assume it follows a Binomial distribution, with being the ratio of surviving cells among cells in . We further define as an average of cell-subtype survival probabilities constituting the bulk, weighted with cell-subtype proportions :

(14)

with the cell-subtype proportions. is the predicted cell surviving rate in the sample n, and thus the target we wish to predict through a cross-validation procedure. We can say that if we let:

(15)

Additionally, we impose the cell-subtype-specific survival probabilities to be non-linearly related with cell-subtype gene expression (a vector of length G) through a sigmoid function:

(16)

Note that the sigmoid in (16) also optionaly depends on , which is an intersect associated with each mutation m in each patient n, attributed to the value 1 in each cancer cell subtype and to 0 in normal cell subtypes. To clarify the notation, we set that , with being a vector containing coefficients associated with each mutation m.

When running CLIFF in overall mode, there are two assumptions that we use to simplify the problem. First, we assume that every sample n share the same cell-subtype expression , for a given gene g and cell-subtype k. Thus, . Second, we assume that cells of the same cell-subtypes across the N samples share the same sensitivities . We thus assume that correspond to the expected drug sensitivity over N samples, and therefore that .

Thus, the overall log-likelihood for our problem is:

(17)

As this log-likelihood expression contains the log of a sum, it is difficult to optimize with respect to . Therefore, we use an Expectation-Maximization to iteratively define lower bound functions that we optimize with respect to , which allows solving (17).

Expectation-Maximization algorithm to solve CLIFF

By using the concavity of the log through Jensen inequality, we can define a lower bound of (17):

(18)

In order to solve (17), we will iterate over the Expectation step or E-step where we update and in (18), and the Maximization step or M-step, where we maximize (18) with respect to . We can therefore define our Maximization algorithm as follow:

E-step.

During E-step, we define our lower bound function at step t, by defining and using defined at previous step:

(19)

This gives us the lower bound function of (18), tight to the original log-likelihood (17) when and satisfies (19).

M-step.

During M-step, we optimize (18) with respect to , considering the variables q^{(0), (t)} and q^{(0), (t)} defined during E-step as constants. Also, the two last term of (18) do not depend on , thus we can solve the following problem which will produce the same optimal coefficients than if solving (18):

(20)

Interestingly, (20) can be solved with standard libraries using a weighted logistic regression model, where we define weights as being and .

We optimize coefficients through a ten-fold cross-validation (taking 90% of the samples N in the training set, and 10% in the testing set), and using early-stopping to terminate the EM algorithm. Optimized coefficients allow to compute optimized values for , our estimate of cell-subtype drug sensitivity.

CLIMB-SC - bulk deconvolution at single-cell level

The bulk deconvolution performed by CLIMB has the particularity of estimating coefficient directly on single-cells from a reference scRNA-seq dataset. Instead of transforming these coefficients into cell-subtype proportions as in Methods (3), we can compute normalized coefficients with the following formula:

(21)

With being the coefficients fitted with GLMNET library. We therefore call this single-cell bulk deconvolution, which links every bulk RNA-seq n to every single-cells i from a scRNA-seq dataset, through what we will call single-cell proportions . These single-cell proportions are thus normalized coefficients fitted by CLIMB, as shown above in (21). Thus we obtain that . As sum up to 1, they can be used by CLIFF as if they were cell-subtype proportions.

CLIFF-SC for drug sensitivity inference at single-cell level

By slightly modifying CLIFF method and by linking it with the ouput of CLIMB-SC, we propose CLIFF-SC, a method providing drug sensitivity estimates at single-cell resolution. If the input dataset is composed of paired bulk RNA-seq and drug sensitivity screening data, a so-called bulk deconvolution at single-cell level is performed using a reference scRNA-seq dataset (cf. CLIMB-SC above).

Thus, as in (15), we model the observed ratio of cells that survives in a mixed population of cells as an average of single-cell survival probabilities , associated with the bulk sample n, weighted by the coefficients :

(22)

With C being the number of cells contained in a scRNA-seq reference datasets. Thus, we model the single-cell drug sensitivity , which is also modeled with a sigmoid function, similar to (16):

(23)

Where is the ith row of the single-cell expression matrix C, thus is a vector of length G, representing the gene expression of cell i. Similarly as in the cell-subtype-centered model, we can derive a log-likelihood function:

(24)

A lower bound can be derived from (24), which gives:

(25)

Then, the Expectation-Maximization algorithm can initialized with all coefficients at 0. This implies that for all i, d. During E-step, we compute values for and variables:

(26)

Then, we optimize coefficients at each step (t), during the so-called M-step:

(27)

And we run the EM algorithm in a similar way than shown above for CLIFF.

Simulation of pseudo-bulk samples from scRNA-seq

We generated pseudo-bulk samples by summing up single-cell expression vectors per gene to obtain an artificially generated “pseudo”-bulk RNA-seq. As the precise amount of cells from each cell subtype added to the mix is known, we thus precisely know the cell-subtype proportions of the mixture and can benchmark deconvolution algorithms efficiently. We generated two datasets of pseudo-bulk in this work, both based on single-cells found in the Van Galen et al. dataset.

First, we used all single cells from the same sample (thus, the same patient), which we summed up to obtain a patient-specific pseudo-bulk RNA-seq sample. Thus, we define a variable as a binary output variable indicating if a cell i belongs to patient n. Thus, a pseudo-bulk samples is a vector of length G containing the gene expression summed over all cells from patient n:

(28)

where is a vector of length G representing the gene expression of cell i. We can therefore easily get the cell-subtype proportion , by simply counting the number of cells from each cell subtypes and dividing by the total number of cells in patient n. For this we introduce an indicator vector , which is of length C and equal to 1 whenever cell i is of type k.

(29)

Van Galen et al. dataset offers only a limited number of patients (16 AML patients at day 0 pre-treatment). To simulate an in-silico a dataset of similar size to the beatAML cohort, we generated 200 pseudo-bulk samples from the Van Galen et al. dataset allowing a random mix of cells from multiple patients. First, we generated 200 random cell-subtype proportions using a Dirichlet distribution. Thus, we assume that cell-subtype proportions , with being a vector of length K, containing one value for each cell subtype k. We fitted using cell-subtype proportions in Van Galen et al. dataset with the R library feralaes/dampack. Then, we generated cell-subtype proportions for 200 pseudo-bulks, with the R library DirichletReg and the previously fitted. Cells from Van Galen et al. dataset were selected with probabilities corresponding to cell-subtype proportions. To use a realistic number of cells to generate our pseudo-bulk samples, we used a gamma distribution fitted on Van Galen et al. dataset using fitdistr function from R library MASS. In the end, we generated a matrix of pseudo-bulk expression of size with N = 200, together with associated true cell-subtype proportions of size .

Simulation of drug sensitivity screening data

To simulate drug screening data, we assumed that the drug sensitivity observed on a mix of cells corresponds to the sum of the cell-subtype drug sensitivity constituting its bulk weighted by their respective proportions. Additionally, we added Gaussian noise with a signal-to-noise ratio of 100. Thus, the simulated survival rate for pseudo-bulk n, in cell-subtype k exposed to drug dose d was generated with the following formula, given the proportions:

(30)

with being a randomly generated number, with if f = 1 (low sensibility cell-subtypes - resisting cells) and if f = 2 (high sensibility cell-subtypes). Here we did not use the drug dose d, only the integer number related to drug dose d = {0, 1, ..., 6}, as we considered seven drug doses. We arbitrarily defined three cell subtypes with high sensitivity and the rest having low sensitivity. The constant c added was set to as it provided visually convincing results. Thus follows a sigmoid function that depends on drug dose d but in a way that is slightly different than in our model. Then, we generated the survival rate at the bulk level using the same formula as in our model but adding gaussian noise to it:

(31)

and was chosen to generate a signal-to-noise ratio arbitrarily chosen at 100.

Area under the curve for summarizing drug sensitivity of multiple data points

As CLIFF produces D predictions for each cell subtype, and CLIFF-SC produces D predictions for every single cell, it is desired to summarize these D data points into a single value. The goal is to summarize cell-subtype or single-cell drug sensitivity. To summarize drug sensitivity from D different drug doses, we normalize drug doses so that the minimum dose = 0 and the maximum dose = 1. As the cell survival rate is , we can compute an area under the curve (AUC) which will be . Conversely, we get an AUC of 1.0 if all cell survives and an AUC of 0.0 if all cells die, taking every dose into account. We computed the AUC using the function computeAUC from the R package PharmacoGx.

Culture conditions of cancer cell lines

Human cancer cell lines (K562, HL60, THP1, SUDHL4) were purchased from the American Type Culture Collection (ATCC; Rockville, MD, USA) and cultured at 37°C in a humidified 5% CO2 environment in RPMI 1640 (HCC38) supplemented with 2mM L-glutamine, 2 g/L D-glucose, and 10% fetal bovine serum (FBS; Gibco ThermoFisher).

Pharmaceutical compounds

Stock solutions for every agent (Table 1) were prepared using DMSO (Sigma-Aldrich; stored at -80°C), further diluted in H2O to the appropriate concentration, plated in Greiner 96-well flat bottom and the plates were stored at -20°C. The pharmaceutical compounds were screened at eight concentrations (Log concentration M = -9.205031395, -8.699881554, -8.130127907, -7.70415926, -7.209309101, -6.70415926, -6.125806332, -5.699837685, -5.204987526, -4.699837685) using a Log2-fold dilution series with matched DMSO concentration vehicle controls.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Table of eight coumpounds used in the in-vitro drug sensitivity experiment.

https://doi.org/10.1371/journal.pcbi.1014862.t001

Bulk proportion for bulk RNAseq and drug sensitivity

Six bulks were created by mixing the four cell lines with the proportions listed in (Table 2). Pure and mixed cell populations were subjected to RNA extraction and drug sensitivity immediately after mixing.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Table of cell line proportions added in tubes manually to generate the mixtures of cell lines.

https://doi.org/10.1371/journal.pcbi.1014862.t002

Cell viability assay

Cells were plated in the 96-well plates containing chemical compounds previously equilibrated at 37C, at a density of 1104 cells per well in 100l culture medium. Cell viability was assessed after drug treatment for 24, 48, 72 or 96 in duplicates hours by incubating the cells with PrestoBlue Cell Viability Reagent, (Thermo Fisher Scientific - Life Technologies A13261). PrestoBlue was at room temperature before adding to the cells in each well. The cells were incubated with 10l PrestoBlue (10% of cell culture volume) for two hours at 37°C. The fluorescence was measured by the bottom with a 560nm excitation filter and a 590nm emission filter in a BioTek Neo HTS microplate. The acquisitions were performed with the following parameters: Gain: 54, Light Source Xenon Flash, Lamp Energy: High, Read Speed: Normal, Delay 0 msec, Mesurements/Data Point: 10, Read Height 5.25mm.

RNA extraction and sequencing

RNA-seq Total RNA from cell lines was isolated with NucleoSpin RNA Plus kit (Machery-Nagel). cDNA was prepared with Maxima Reverse Transcriptase (Thermo Scientific). Sequencing libraries were performed with Illumina Truseq Stranded mRNA LT kit. Reads were mapped to the human (hg19) genome using Hisat2. Counts on genes were generated using featureCounts. Normalization for sequencing depth has been done using the counts on genes as library size using the TMM method as implemented in the limma package of Bioconductor.

Single Cell RNAseq (MULTIome-Seq)

Single cell RNAseq was performed using 10x Genomics Chromium Single Cell Controller following the 10x Genomics Multiome (scRNA/scATAC-seq) demonstrated protocol: Chromium Next GEM Single Cell Multiome ATAC + Gene Expression User Guide (CG000338). To note, experimental parameters specific to the experiment were 3,000 for the Targeted Nuclei Recovery and 5 minutes for the lysis time.

Single-cell RNA-seq analysis

Cellranger-arc was used to obtain counts on genes using default parameters. Counts were obtained using cellranger using hg19. Only uniquely mapped reads on genes that were expressed in at least 1% of the samples were kept. Seurat’s SCTransform was used to normalize the data and correct for mitochondrial percentage and total number of reads biases.

Code Availability

The code is freely accessible on GitHub. CLIMB and CLIFF code can be accessed via an R library, “ClimbTheCliff,” available via the devtools command devtools::install_github(“alexdray86/ClimbTheCliff”). Documentation is available on https://github.com/alexdray86/ClimbTheCliff, and the code to reproduce the manuscript’s figures is available in https://github.com/alexdray86/Coudray2026. Package documentation, including installation instructions and worked examples, is available at https://alexdray86.github.io/ClimbTheCliff/.