One of the central reasons cancer is so difficult to treat is that tumors are not made of a single uniform population of cells. Instead, they contain multiple distinct subpopulations - often called clones - that differ from each other in their genetic makeup, behavior, and sensitivity to drugs. This diversity is called intratumoral heterogeneity.
In acute myeloid leukemia (AML), one of the most aggressive blood cancers, this heterogeneity is well documented. Different clones within the same patient can have different survival strategies, meaning that a drug that kills one clone may leave another unaffected - allowing that surviving clone to repopulate the cancer and drive relapse. This is a major reason why AML so frequently recurs after initial treatment.
The same problem exists in solid tumors. In high-grade serous ovarian carcinoma (HGSC), a particularly deadly form of ovarian cancer, tumors contain a complex mix of cancer and non-cancerous cells, and different patients' tumors have different molecular drivers even when they share the same diagnosis. No single drug works reliably for all patients.
Ideally, physicians would treat each patient with a combination of drugs specifically matched to the unique constellation of cancer clones present in that individual's tumor. But identifying such personalized combinations is extraordinarily difficult: the number of possible drug-and-dose combinations vastly exceeds what could ever be tested directly in each patient's scarce cancer cells.
This paper presents scTherapy, a machine learning tool designed to solve the personalized treatment problem using single-cell RNA sequencing (scRNA-seq) data. Single-cell RNA sequencing measures the gene activity of thousands of individual cells simultaneously, allowing researchers to map the full diversity of cell types within a tumor sample.
The key innovation of scTherapy is that it can take only the scRNA-seq data from a patient's tumor sample and automatically predict which drugs or drug combinations would most effectively kill the cancer cells while sparing the normal cells in that same sample. This selectivity - targeting malignant cells without harming healthy ones - is the holy grail of cancer treatment.
scTherapy uses a type of machine learning called a gradient boosting model (LightGBM), pre-trained on a massive database of over 394,000 gene expression profiles measured in cancer cell lines after treatment with nearly 20,000 different drug treatments. This training allows the model to learn the relationship between a cell's gene activity and how it responds to drugs.
When applied to a patient sample, scTherapy identifies genetically distinct cancer subclones, predicts which drugs would reverse each subclone's abnormal gene expression back toward normal, and then combines the top predictions for each subclone into a personalized multi-drug treatment regimen. The entire pipeline requires only the scRNA-seq count matrix as input.
The scTherapy pipeline operates in several automated steps. First, it processes the raw scRNA-seq data to identify distinct cell populations in the tumor sample. It uses three complementary tools - ScType, CopyKAT, and SCEVAN - to classify cells as malignant or normal, using a majority voting strategy to maximize accuracy. This step distinguishes the cancer cells from surrounding normal cells in each patient sample.
Next, the tool identifies genetically distinct cancer subclones by inferring large-scale chromosomal gains and deletions from the RNA data using a method called InferCNV. Cancer cells often acquire abnormal chromosome copies, and detecting these patterns reveals which cells share a common genetic ancestor. The subclone structure is then visualized as an evolutionary tree showing which clone gave rise to which.
For each identified subclone, scTherapy calculates which genes are differentially expressed compared to normal cells from the same patient. These gene expression differences are then fed into the pre-trained LightGBM model, which matches them against known drug-induced gene expression changes to find compounds that would push the cancer cells' activity back toward a normal state while effectively inhibiting their growth.
The model provides dose-specific predictions - specifying not just which drugs to use but at what concentration - and filters out low-confidence predictions using conformal prediction statistics. It also enforces a 1 micromolar dose ceiling to favor clinically feasible concentrations and minimize the risk of off-target toxicity to normal cells.
The team tested scTherapy on 12 bone marrow samples from AML patients at various disease stages: initial diagnosis, relapse, and treatment-refractory disease. Each patient's sample contained a heterogeneous mixture of leukemic cells (blasts) and normal blood cells. scTherapy generated personalized treatment predictions for each patient, with most predicted treatments being unique to a single patient.
Retrospective validation confirmed that drugs predicted to be effective by scTherapy showed significantly better actual cell-killing in lab assays compared to drugs predicted to be ineffective (p less than 0.0001). This comparison was done using existing drug sensitivity data from the same patient samples - data that had not been used in training the model.
For four patients with sufficient cells available for combination testing, the team designed two-drug combinations targeting the two major cancer subclones in each patient. All predicted combinations showed either synergistic or at least additive effects in bulk cell viability assays - meaning the drugs worked better together than either alone. More importantly, flow cytometry assays showed that in most patients, the combinations selectively killed leukemic blast cells while leaving normal immune cells largely intact.
Of 24 tested drug combinations across the four AML patients, 21 (88%) led to increased killing of leukemic cells compared to single agents. Only 3 combinations (13%) showed high toxicity to normal cells, and these were appropriately flagged for exclusion. This high selectivity ratio demonstrates that scTherapy's predictions are both effective and cancer-specific.
To test whether scTherapy could work in solid tumors - where direct drug testing in patient cells is much harder - the researchers applied it to samples from three patients with high-grade serous ovarian carcinoma (HGSC). These patients had metastatic disease with poor responsiveness to standard chemotherapy, representing one of the most challenging treatment scenarios in oncology.
Because the proportion of cancer cells in the ovarian samples was relatively small (making reliable subclone identification difficult), scTherapy was adapted to predict single-agent therapies targeting all cancer cells collectively, rather than subclone-specific combinations. Cancer cells were identified using established marker genes including PAX8 and MUC16.
Predictions were validated in patient-derived tumor organoids - three-dimensional cancer cell cultures grown in the laboratory from each patient's own tumor cells. These organoids closely mimic the behavior of the original tumor. Of 54 evaluated treatments across three patients, 31 (57%) caused more than 50% inhibition of the cancer organoid cells, and across all samples, predicted effective treatments showed significantly higher tumor cell inhibition than normal stromal cells (p less than 0.01).
Certain drug classes showed non-selective responses - meaning they also substantially inhibited normal cells - including proteasome inhibitors (bortezomib, ixazomib) and an HSP90 inhibitor (ganetespib). These findings underscore the importance of experimental validation before clinical application, as even AI-predicted treatments may carry toxicity risks that require careful assessment.
To assess scTherapy's broad applicability, the team applied it to publicly available single-cell RNA datasets from five cancer types: AML, HGSC (ovarian), lung adenocarcinoma, pancreatic ductal adenocarcinoma, and triple-negative breast cancer - totaling dozens of patient samples across highly treatment-resistant malignancies.
The analysis revealed a striking balance in predicted treatments: 25% of predictions were shared among patients with the same cancer type (disease-specific), 22% were common across all five cancer types (pan-cancer), and 19% were unique to a single patient (truly personalized). This distribution suggests that while there is some consistency in treatment patterns within cancer types, a substantial fraction of patients have truly unique therapeutic needs.
The cancer type-specific predictions were biologically meaningful: AML patients, whose cancer cells often have hyperactive cell signaling, received frequent predictions of MEK inhibitors and PLK inhibitors - both classes extensively studied in AML. Ovarian cancer predictions were more heterogeneous, reflecting the diverse biology of that disease.
Notably, 22 of 131 total predicted treatments (17%) were already in phase 3 or 4 clinical trials - indicating that scTherapy is identifying clinically relevant treatments, not just theoretical predictions. This alignment with ongoing trials provides independent validation that the model's outputs are biologically grounded.
scTherapy was compared head-to-head against two existing single-cell drug prediction methods, BeyondCell and scDrug, using the AML patient dataset. In both predicting effective and ineffective drugs, scTherapy consistently outperformed the other methods, with the Area Under the ROC Curve (a measure of prediction accuracy) being significantly better (p less than 0.001).
A key advantage of scTherapy is that it addresses three limitations of previous methods: it predicts drug combinations rather than just single agents; it provides dose-specific predictions critical for clinical translation; and it has been experimentally validated in primary patient cells rather than just benchmarked computationally.
The current version of scTherapy has important limitations. It cannot currently incorporate point mutations in individual genes - meaning that cancer driven primarily by a specific point mutation (such as melanoma with BRAF-V600E) would not fully benefit from the model's predictions without additional genomic data integration, which is planned for future versions.
The authors envision future extensions of scTherapy to include multi-omics data (combining genomic, transcriptomic, and proteomic information), larger drug libraries, and application to rare cancer types. The tool is designed to make ex vivo drug testing in scarce patient cells more cost-effective by narrowing down the testing space to the most promising candidates - a critical practical advantage in clinical settings.
scTherapy represents a significant step toward realizing the promise of precision oncology - the goal of matching each cancer patient with the treatment most likely to work for their specific tumor. By reading the gene activity of individual cancer cells and translating that information into treatment predictions, the tool provides a clinically actionable output from single-cell sequencing data.
The approach is particularly valuable for patients with relapsed or refractory cancer who have exhausted standard treatments. In these patients, understanding which new combinations might target the specific clones driving their resistant disease is urgently needed and often impossible to determine through conventional means.
The tool could also find application in upfront treatment planning - predicting effective combinations before any treatment begins - potentially preventing the clonal selection that occurs when a single-agent treatment kills sensitive cells while allowing resistant ones to expand. Treating multiple clones simultaneously from the start may reduce the rate of relapse.
With single-cell RNA sequencing becoming faster and cheaper each year, the feasibility of incorporating scRNA-seq profiling into routine clinical cancer care is increasing. scTherapy provides a blueprint for how this technology can be translated into direct patient benefit by converting complex molecular data into clinically actionable, personalized treatment recommendations.