High intratumoral heterogeneity and clonal evolution in cancer drive therapy resistance. In AML, single-cell genomic analyses have mapped clonal evolution and cellular reprogramming related to chemoresistance and relapse. Similarly, solid tumors exhibit significant inter- and intra-patient heterogeneity, impacting diagnosis and treatment. The tumor microenvironment (TME) influences treatment outcomes. Single-cell RNA sequencing (scRNA-seq) provides insights into genomic, transcriptomic, and epigenomic variations, enabling personalized treatment strategies. However, identifying effective combinatorial therapies that selectively co-inhibit multiple pathways in heterogeneous populations remains a challenge. Existing computational approaches using scRNA-seq data associate cells with disease attributes but don't identify multi-targeting drugs or combinations at a single-cell, patient level. This study addresses this limitation by presenting scTherapy, a machine learning model that identifies cancer-selective and low-toxic multi-targeting options based solely on scRNA-seq data, thus overcoming limitations in ex vivo drug testing feasibility.

The literature extensively highlights the challenges of intratumoral heterogeneity and clonal evolution in driving therapy resistance across various cancers, including AML and solid tumors such as HGSC. Numerous studies have utilized single-cell technologies, primarily scRNA-seq, to characterize tumor heterogeneity and identify potential therapeutic targets. However, translating this single-cell data into clinically actionable personalized treatment strategies remains limited. Existing computational methods often focus on single-drug effects or lack the ability to predict effective drug dosages and selective co-inhibition of multiple cancer subclones while sparing normal cells. The authors review these existing computational approaches and identify several limitations that scTherapy attempts to address.

scTherapy uses a two-step approach. First, it pre-trains a gradient boosting model (LightGBM) using a large reference database (LINCS 2020 and PharmacoDB) containing genome-wide transcriptomic profiles and drug-induced cell viability responses measured across multiple doses in numerous cancer cell lines. This model learns drug response differences across cell populations based on fold changes in differentially expressed genes (DEGs) post-treatment. Second, the pre-trained model is applied to patient-specific scRNA-seq data. The input consists of fold changes in DEGs between normal and cancer cell populations in individual patient samples. The model generates a ranked list of effective multi-targeting options, considering both efficacy and selectivity. Low-confidence predictions and non-tolerated doses are filtered out. For AML, the workflow involves scRNA-seq data processing, cell clustering, identification of normal and malignant cells using an ensemble approach (ScType, CopyKAT, and SCEVAN), subclone identification using InferCNV, and DEG analysis between normal cells and subclones. The DEGs are then input to the LightGBM model to predict multi-targeting options. For HGSC, due to smaller cancer cell proportions, the focus is on predicting multi-targeting monotherapies, using all cancer cells as a collective malignant entity. Experimental validation in both AML and HGSC involves ex vivo drug testing in patient-derived cells (bone marrow cells for AML and tumor organoids for HGSC), using bulk and flow cytometry assays to assess efficacy and selectivity.

In AML, scTherapy predicted combinatorial treatments targeting two major subclones in each patient. Ex vivo validations showed that 96% of the predicted multi-targeting treatments exhibited selective efficacy or synergy, and 83% demonstrated low toxicity to normal cells. Most treatments were patient-specific; a few shared treatments showed variable responses across patients. In HGSC, scTherapy predicted multi-targeting monotherapies due to the low proportion of cancer cells. Experimental validation in patient-derived tumor organoids showed that 57.4% of the treatments resulted in >50% inhibition of tumor cells, with only 20.4% showing similar inhibition of non-cancerous cells. Pan-cancer analysis across five cancer types revealed that 25% of predicted treatments were shared among patients of the same tumor type, while 19% were patient-specific, and 22% were common across all five cancer types. Comparative analysis with BeyondCell and scDrug in AML patients showed scTherapy consistently outperformed these methods in predicting both effective and ineffective single drugs.

scTherapy addresses the need for multi-targeting therapies in heterogeneous cancers by integrating scRNA-seq data with large-scale drug response profiles. The findings highlight the potential for personalized medicine by identifying both shared and patient-specific treatments. The high success rate of experimental validations supports the model's accuracy and clinical relevance. The ability to predict effective doses and prioritize combinations for ex vivo testing improves the efficiency of personalized drug development. The model's applicability to both hematological and solid tumors demonstrates its versatility.

scTherapy provides a novel computational framework for identifying personalized multi-targeting treatments in cancer. The approach successfully predicts effective and selective drug combinations based solely on scRNA-seq data. Experimental validations confirm its efficacy and safety. Future directions include incorporating multi-omics data, exploring additional cancer types, and refining prediction strategies for different cellular contexts.

The current version of scTherapy primarily relies on scRNA-seq data and might benefit from the integration of other omics data (e.g., point mutations) for more comprehensive predictions. The limited number of patients in some cohorts restricts the generalizability of certain findings. The reliance on pre-trained models from cell lines means that in vitro findings need to be validated in more complex ex vivo models. Differences in growth dynamics between cell lines and organoids could slightly affect prediction accuracy.