Machine learning-based development of a cytotoxicity prediction model for NK cell therapy in cancers

Abstract

Purpose: Natural killer (NK) cells mediate anti-tumor immunity through integrated signaling of inhibitory and activating receptors. The efficacy of NK cell adoptive transfer therapy varies among patients due to heterogeneous receptor-ligand expression. This study aimed to develop a predictive model based on receptor-ligand interactions to determine NK cells' therapeutic effects.

Methods: Through analyses of receptor-ligand expression profiles of NK and tumor cells and assessment of NK cell cytotoxicity, we developed a machine learning-based random forest model using 11 key receptor-ligand pairs selected through database mining and experimental screening. Flow cytometry was used to obtain receptor-ligand profiles, and combined predictors were calculated for each pair. The model was validated using independent datasets and evaluated for generalizability across different tumor types.

Results: The model showed significant predictive performance, achieving an accuracy of 84.2% and an area under the curve (AUC) of 0.908 in ovarian cancer cohorts. This predictive capability was validated in both in vitro experiments and clinical samples, revealing complex non-linear interactions between receptor-ligand expression and NK cell killing efficacy. Cancer-specific ligand expression patterns were identified. While showing optimal performance in studied cancer types, it exhibited moderate applicability to other cancers and demonstrated potential compatibility with transcriptomic data for prediction.

Conclusions: This model provides tools and foundations for the precise treatment of tumors using NK immune cells and may be applied in clinical practice.

Supplementary Information

The online version contains supplementary material available at 10.1007/s13402-025-01113-1.

Background

Cellular immunotherapy, a safe and effective approach, is becoming increasingly prominent in the field of tumor immunotherapy. Among the immune cells being studied, natural killer (NK) cells are particularly valuable. NK cells are cytotoxic lymphocytes that play key roles in the innate immune system [1]. Compared with T cells, NK cells provide several advantages, including pan-specific recognition without prior sensitization, high safety, good tissue compatibility, abundant sources, and a singular function [2–4].

The functional activity of NK cells is precisely regulated through a sophisticated array of surface receptors that maintain a delicate balance between activating and inhibitory signals [5, 6]. These receptors can be broadly categorized into several major families: killer cell immunoglobulin (Ig)-like receptors (KIRs), natural cytotoxicity receptors (NCRs), and C-

type lectin-like receptors [7]. Inhibitory receptors, including KIR2DL and NKG2A/CD94, recognize self MHC class I molecules on healthy cells and generate signals that prevent NK cell activation [8]. Conversely, activating receptors such as NKG2D and natural cytotoxicity receptors (NCRs; NKp30, NKp44, and NKp46) recognize stress-induced ligands or viral proteins to trigger NK cell cytotoxicity [9]. Additional co-stimulatory receptors like 2B4, and DNAM-1 provide complementary activation signals [9]. This complex receptor network enables NK cells to effectively distinguish between normal and abnormal cells, forming the foundation for their immune surveillance capabilities. NK cell function can be modulated through in vitro stimulation, expansion, and genetic editing, ensuring both efficacy and safety against tumors [7].

Currently, a key focus of research on NK cell-based tumor treatments is adoptive NK cell transfer therapy [3]. This involves transferring NK cells derived from peripheral blood, stem cells, or induced pluripotent stem cells into bodies of patients for treatment [7, 10]. Our previous studies have established robust protocols for expanding NK cells from peripheral blood, typically obtained from healthy donors, using in vitro expansion systems to yield highly purified NK cells for treating patients with cancer [11]. Although NK cell therapy is a promising treatment option for solid tumors, it has not yet achieved the desired efficacy for tumor treatment. A previous study systematically analyzed 31 clinical trials involving NK cell infusion in 600 patients [12]. The objective response rate was 28.2%, and the disease control rate was 63.2%. Significant heterogeneity was observed among different samples, with marked variability in therapeutic responses among different patients. These differences in efficacy may be associated with variations in cell surface marker expression. Heterogeneity in tumor and NK cells between individuals leads to differences in the receptor and ligand expression profiles of cells from various sources [13]. This differential expression can influence the integrated killing signals of NK cells toward tumor cells, ultimately resulting in inconsistent clinical outcomes of NK cell immunotherapy [14].

Several ligands related to NK cell activation and inhibition have been identified as potential biomarkers for predicting therapeutic efficacy; however, these molecules do not fully explain the mechanisms underlying the sensitivity or resistance of tumor cells to NK cell-mediated killing [15, 16]. This is primarily because NK cell receptors do not act independently of each other. Recent years have witnessed a paradigm shift in predicting immunotherapy outcomes, moving from single biomarker-based approaches to multi-dimensional prediction models. Within NK cell therapy, machine learning has showed promise for efficacy prediction. Key advances include: the 19-gene NK cell signature for triple-negative breast cancer prognosis derived from single-cell RNA sequencing (scRNA-seq) by Liu et al. [17]; the integrated prognostic model of Zhao et al. that combines NK cell biomarkers with clinical parameters in multiple myeloma [18]; and the pan-cancer applicable NK cell-related gene signature developed by Li et al. using The Cancer Genome Atlas (TCGA) glioma data [19]. For mechanistic insights, CellPhoneDB has been widely adopted to decode NK cell-tumor cell interactions at single-cell resolution [20]. However, current models largely overlook the coordinated expression patterns of NK cell receptors and their cognate tumor ligands as predictive features. Therefore, further investigation into how receptor–ligand (R-L) interactions between NK cells and tumor cells collectively affect overall NK cell efficacy is of significant scientific and clinical importance.

In this study, we used conditionally reprogrammed (CR) tumor cells and human NK cells that were expanded and cultured in vitro. By analyzing the molecular expression profiles of NK and ovarian cancer (OC) cells and conducting in vitro NK cytotoxicity assays, we established a research model to predict tumor cell sensitivity to NK cell-mediated killing. We aimed to investigate the impact of R-L expression data pairing on the efficacy of adoptive NK cell transfer therapy and explore the clinical application value of this model in personalized NK cell therapy for cancer.

Materials and methods

Data sets collection

We obtained the mRNA expression profiles of various tumor types (ovarian serous cystadenocarcinoma (OV), skin cutaneous melanoma (SKCM), stomach adenocarcinoma (STAD), breast invasive carcinoma (BRCA), colon adenocarcinoma (COAD), liver hepatocellular carcinoma (LIHC), esophageal carcinoma (ESCA), and lung adenocarcinoma (LUAD)) from Oncolnc (http://www.oncolnc.org/). The expression profiles of the tumor ligands required for the study were obtained from the Human Protein Atlas (HPA) database (https://www.proteinatlas.org/), and with immunohistochemistry images from patients with different tumor samples. Survival correlation analyses for immune cell levels in tumors were conducted using the Gene Expression Profiling Interactive Analysis 2021 (GEPIA2021) database (http://gepia2021.cancer-pku.cn/). Datasets for ovarian cancer (GSE54388), gastric cancer (GSE79973), breast cancer (GSE36295 and GSE32526), bladder cancer (GSE120736), prostate cancer (GSE68138), liver cancer (GSE144269), lung cancer (GSE27556), and NK cells (GSE116660, GSE75091, and GSE15743) were downloaded from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/). Preliminary analyses were performed using GEO2R (https://www.ncbi.nlm.nih.gov/geo/geo2r/), and volcano plots were generated using GraphPad Prism 9.0 (GraphPad Software, San Diego, CA, USA). Analysis of tumor ligands associated with overall survival (OS) in patients with cancer] and transcriptomic sequencing data for several cancer types were obtained from GEPIA (http://gepia.cancer-pku.cn/). The gene expression profiles of several tumor cell lines were retrieved from the Dependency Map (DepMap) database (https://depmap.org/portal/). Finally, correlation analyses between the gene expression of tumor ligands and NK cell infiltration were conducted on the Tumor Immune System Interactions Database (TISIDB) website (http://cis.hku.hk/TISIDB/) and relevant data from the Tumor Immune Estimation Resource 2.0 (TIMER2.0) website (http://timer.cistrome.org/) were obtained to create correlation bubble charts.

Tumor cells culture

CR tumor cells were provided through collaboration with Hefei PreceDo Medical Laboratory Co., Ltd. (Anhui, China) and originally established using the feeder-dependent culture protocols as described previously [21–23]. For experimental use, the cells were maintained based the standardized culture protocol of the provider. The CR OC (D84, A09, L8, L37, L48, L57, L58, L62, L74, L82, L83, L91, L92, L93, L96, L100, and L106), breast (BC-1), esophageal (EC-1), gastric (GC-1), and lung (LC-1) cancer cells were maintained with their corresponding reprogramming media with Human Ovarian Carcinoma Conditional Reprogramming Medium (PRECEDO, PRS-OCM-CR), Human Mammary Epithelial Cell Medium (PRS-BCM-2D), Human Esophagus Carcinoma Cell Medium (PRS-ECM-2D), Human Gastric Carcinoma Cell Medium (PRS-GCM-2D), and Human Lung Carcinoma Cell Medium (PRS-LCM-CR), respectively.

MDA-MB-468 (RRID: CVCL_0419), MDA-MB-231 (RRID: CVCL_0062), BT549 (RRID: CVCL_1091), HEYA8 (RRID: CVCL_8878), and MCF-7 (RRID: CVCL_0031) cell lines were provided by Zhenye Yang (University of Science and Technology of China, Anhui, China). COV504 (RRID: CVCL_2424), OVCAR8 (RRID: CVCL_1629), and HEPG2 (RRID: CVCL_0027) cell lines were provided by Ying Zhou (University of Science and Technology of China). A375 (RRID: CVCL_0132) and SKOV3 (RRID: CVCL_0532) cell lines were provided by Zhigang Tian (University of Science and Technology of China). These cell lines were maintained in DMEM (VivaCell Biosciences, C3113–0500) supplemented with 10% fetal bovine serum (FBS; Biological Industries, 04–001-1ACS). K562 (RRID: CVCL_0004), U266 (RRID: CVCL_U266), and HO8910 (RRID: CVCL_6868) cell lines were purchased from the cell bank of the Chinese Academy of Sciences (Shanghai, China) and maintained in RPMI-1640 (VivaCell, C3010–0500) supplemented with 10% FBS. SW620 (RRID: CVCL_0547), MKN74 (RRID: CVCL_2791), MKN7 (RRID: CVCL_1417), SK-BR-3 (RRID: CVCL_0033), HCC1954 (RRID: CVCL_1259), AGS (RRID: CVCL-0139), HGC27 (RRID: CVCL_1279), and MKN45 (RRID: CVCL_0434) cell lines were purchased from Procell Life Technology (Wuhan, China). SW620 cells were maintained in Leibovitz’s L-15 Medium (Pricella, PM151010B) supplemented with 10% FBS. AGS cells were cultured in DMEM/F-12 Medium (VivaCell Biosciences, C3130–0500) supplemented with 10% FBS, HGC27 cells were cultured in RPMI-1640 supplemented with 20% FBS, and HCC1954, SK-BR-3, MKN74, MKN7, and MKN45 cells were cultured in RPMI-1640 supplemented with 10% FBS. All cells were cryopreserved to maintain viability and stored in liquid nitrogen according to the manufacturer’s instructions.

CR tumor cell subculture

Cell-forming colonies were observed under a microscope, and cells were passaged after reaching 80–90% confluence. Cells were removed from the incubator, the old culture medium was discarded, and cells were washed twice with phosphate-buffered saline (PBS; Biosharp, BL302A). Next, 0.05% trypsin (Beyotime Biotechnology, C0201-500 mL) was added, and the cells were incubated at 37 °C for 2–3 min. The digestion progress was stopped using a serum-containing culture medium. Cells were then collected, transferred to a centrifuge tube, and centrifuged at 300 × g for 5 min, after which the supernatant was discarded, and cells were resuspended in primary tumor culture medium, with 20 μL of this solution used for counting. The cell suspension was inoculated into a culture flask at a density of 0.52–0.56 × 10⁵/cm². Then, γ-irradiated 3T3 cells (CR cells:3T3 cells = 3:1; irradiation dose: 35 Gy) were added and mixed with culture medium, gently shaken, and returned to the incubator at 37 °C in 5% CO₂ for further culture.

Peripheral blood mononuclear cell (PBMC)-NK cell expansion culture

PBMC-NK cell expansion was performed as described previously [11]. Briefly, fresh buffy coats (Leukocyte Source) were used as the PBMC source. This study was conducted in accordance with the Declaration of Helsinki and approved by the Ethics Committee of Taicang First People’s Hospital (Approval No. 2022-SR-020; approval date: November 10, 2022). All participants provided written informed consent prior to enrollment. Clinical trial number: not applicable. PBMCs were isolated from buffy coats using Ficoll-Paque (Biosharp, BL672A). Peripheral blood-derived NK cells were expanded using a feeder-free NK cell expansion system (feeder-free expanded natural killer cells, FF-NK), feeder cell (K562)-based expansion system (feeder cell expanded natural killer cells, feeder-NK), and a protein-conjugated magnetic bead expansion system (protein-conjugated magnetic bead-expanded NK cells, MB-NK). Although we did not compare the efficacy of these three methods, we only used products with CD56⁺ cell purity exceeding 95%. The purity of NK cells (CD3^– CD56⁺) expanded by the different methods was assessed using flow cytometry (Supplementary Figure 1).

For feeder-free expansion, we followed our laboratory’s established protocol [11]. In brief, PBMCs were cultured in complete KBM-581 medium (Corning, 88–581-CM) supplemented with 5% autologous plasma and 1000 IU/mL recombinant human interleukin-2 (rhIL-2; Sihuan Biological 8,797,873). Cells were seeded at 1.5 × 10⁶ cells/mL in T25 flasks (Corning 430,639) pre-coated with anti-CD16 and anti-CD3 monoclonal antibody to facilitate NK cell enrichment. The cells were cultured for 17–27 days at 37 °C with 5% CO₂, with fresh medium added every 2–3 days to maintain the cell density at 1.8–3.2 × 10⁶ cells/mL until the desired cell number was reached.

For feeder cell-based expansion, we adapted the protocol described by Fujisaki H et al. [24] with the following modifications: Genetically modified K562 cells (engineered to express CD86, 4-1BBL, and rhIL-21) were irradiated (100 Gy) and resuspended in KBM581 medium. PBMCs were seeded at 1.25 × 10⁵ cells/mL with irradiated feeder cells (PBMCs: feeder cells = 2:1) in T25 flasks using KBM581 medium supplemented with 5% AB serum (GeminiBio, 100–512), and 50 IU/mL rhIL-2. Depending on the state of feeder cell lysis, the medium was changed, and rhIL-2 was added again on day 3 at 50IU/ml in cultures. On days 6–7, cells were transferred to T75 flasks at 2.5 × 10⁵ cells/mL with fresh feeder cells (NK cells: feeder cells = 1:1). Thereafter, NK cells were subcultured at the same density (2.5 × 10⁵ cells/mL) for continued expansion. No additional feeder cells were required beyond this point, and cultures were maintained until Day 14 for subsequent analysis.

For magnetic bead-based expansion, we adapted established protocols [25] with optimized modifications. In brief: The PBMC suspension was adjusted to a cell density of 1.2 × 10⁶ cells/mL in KBM581 medium supplemented with 5% AB serum. 5.04 mg of magnetic beads (Affimag PSC core-shell polystyrene magnetic microspheres, Tianjin Baseline ChromTech, 3451 & 3452) were conjugated with stimulatory proteins, sterilized (75% ethanol for 5 min followed by PBS washes), and aseptically transferred to culture flasks. The prepared cell suspension was then added to flasks containing the conjugated beads. rhIL-2 was supplemented to a final concentration of 1000 U/mL. Subsequent expansion followed our standard feeder-free NK cell culture protocol.

Induced pluripotent stem cell (IPSC)-NK cell derivation

For IPSC-NK cell generation, we implemented optimized modifications to the expansion protocol originally described by Hermanson et al. [26] and Knorr et al. [27]. Briefly, iPSCs were maintained under feeder-free conditions and differentiated through embryoid body formation, followed by culture in NK differentiation medium. The emerging CD34⁺ hematopoietic progenitors were subsequently expanded with rhIL-15 (Peprotech, AF-200–15-50UG) and rhIL-2 supplementation to generate functional CD56⁺ NK cells.

All expanded NK (FF-NK/feeder-NK/MB-NK/IPSC-NK) cell products in this study were designated with batch-specific identifiers following the format FF/feeder/MB/IPSC-NK-XXXXX (such as FF-NK-0330O), where the alphanumeric code represents the unique production batch number.

Real-time cell analysis (RTCA)-based killing assay

The killing assay was conducted using RTCA (ACEA Biosciences, xCELLigence RTCA S16), according to the manufacturer’s instructions. Briefly, adherent cells were washed and resuspended in a medium containing 10% FBS. The E-Plate 16 was removed, and 1.5–3.0 × 10⁴ cells per well thoroughly mixed in 100 µL of medium added. The E-Plate 16 was placed on the RTCA station in a constant-temperature incubator. The total monitoring time was set to 48 h, with a data collection interval of 15 min. NK cells were added approximately 12 h after plating. The expanded NK cells were washed and resuspended in RPMI-1640 medium containing 10% FBS. NK cell suspensions were diluted and added to tumor cell cultures at an effector-to-target ratio of 5:1. The plate was returned to the RTCA instrument for continued monitoring. The NK cells were co-cultured with tumor cells in a constant-temperature incubator at 37 °C for 4 h. Each cytotoxicity assay was performed in triplicate to evaluate data consistency.

CCK8-based killing assay

Approximately 100 μL of tumor cell suspension was added to a 96-well plate at a concentration of approximately 1 × 10⁵ cells/mL. The plate was incubated at 37 °C and 5% CO₂ for approximately 12 h. Thereafter, 50 μL of expanded NK cells was added to the plate as effector cells at an effector-to-target ratios of 5:1. After placing the plate in the incubator for 4 h, 15 μL of CCK-8 solution (Biosharp, BS350B) was added to each well (avoiding bubble generations in the wells). The plate was placed back in the incubator for another 2 h. The absorbance of the cell culture plates was measured at 450 nm using a microplate reader (SpectraMax iD5 Multi-Mode Microplate Reader, Molecular Devices 2,020,736,010). Each cytotoxicity assay was performed in triplicate to evaluate data consistency.

Flow cytometry analysis of cell surface markers

Flow cytometry was performed to analyze cell surface markers. Fc receptors were initially blocked by incubating cells with 10% mouse serum (Future, F001008) at 4 °C for 15 min, prior to incubation with appropriate antibodies for 30 min at 4 °C. For intracellular staining, such as PCNA and TGF-β, cells were incubated with 2.5 ng/mL Monensin Solution (Sigma-Aldrich 22,373–78–0) for 4 h at 37 °C in a 5% CO₂ incubator, followed by antibody staining for extracellular markers. The fluorescent antibodies used in this study are listed in Tables 1 and 2. DAPI (10 µg/mL; BioLegend 422,801) was used to differentiate dead cells 5 min before the test. Next, 13-color fluorescence-activated cell sorting analysis of cell markers was performed using a flow cytometer (Beckman Coulter, CytoFLEX). The mean fluorescence intensity (MFI) protein expression levels were normalized to MFI ratios (the ratio of MFI in the experimental group to that in the isotype control group) to ensure consistent comparisons across different samples. Protein expression levels for all cells used in the predicted model were verified with at least five replicates to ensure stable and reliable expression profiles.

Table 1.

Antibodies used for flow cytometric detection of NK cells

Fluorescence channel	Protein	Cat.	Company
Brilliant Violet 785	CD56	362550	Biolegend
PerCP/Cyanine5.5	CD3	300430	Biolegend
Brilliant Violet 605	CD16	302040	Biolegend
PE	NKG2A	FAB1059P	R&D
APC	NKp30	325210	Biolegend
PE/Cyanine7	NKp44	325116	Biolegend
BV510	2B4	329534	Biolegend
APC	CD158b	312612	Biolegend
PE	PD-1	329906	Biolegend
PE/Cyanine7	DNAM-1	338316	Biolegend
APC	CD158d	347006	Biolegend
PE	TRAIL	308206	Biolegend
Alexa Fluor 488	LFA-1	363404	Biolegend
APC	TIGIT	372706	Biolegend
FITC	CD2	300206	Biolegend
APC	NKG2D	320808	Biolegend
PE	CD158a	374904	Biolegend
FITC	CD158e1	312706	Biolegend
APC	CD200R	329308	Biolegend
PE/Cyanine7	SIGLEC9	351520	Biolegend
APC	CD56	362504	Biolegend
PE	CD56	362508	Biolegend
PE/Cyanine7	CD56	318318	Biolegend
FITC	CD56	362546	Biolegend

Table 2.

Antibodies used for flow cytometric detection of tumor cells

Fluorescence channel	Protein	Cat.	Company
APC	MICA	FAB1300A–100	R&D
APC	MICB	FAB1599A–100	R&D
PE	B7–H6	FAB7144P–100	R&D
APC	ULBP1	FAB1380A	R&D
APC	ULBP3	FAB1517A	R&D
PE	ULBP4	FAB6285P	R&D
APC	DR4	307208	Biolegend
PE	DR5	307406	Biolegend
PerCP/Cyanine5.5	HLA-E	342610	Biolegend
PE	HLA-C	566372	BD
PE/Cyanine7	HLA-G	335912	Biolegend
PE/Cyanine7	CD155	337614	Biolegend
PE	CD112	337410	Biolegend
Alexa Fluor 700	CD54	353126	Biolegend
PerCP-Cy5.5	FGF-β1	562423	BD
FITC	PD-L1	393606	Biolegend
PE/Cyanine5	CD58	330909	Biolegend
PerCP/Cyanine5.5	CD48	336716	Biolegend
PE/Cyanine7	CD200	399806	Biolegend
PerCP/Cyanine5.5	B7–H3	351010	Biolegend
PE	PCNA	307908	Biolegend
-	CA125	MAB56091–100	R&D
APC	Mouse F(ab)2 IgG(H+L)	F0101B	R&D
PE	MICB	FAB1599P–100	R&D
APC-Cyanine7	CD45	368516	Biolegend
FITC	CD34	343504	Biolegend
Brilliant Violet 605	CD117	313218	Biolegend
APC	CD58	330918	Biolegend
PE	CD58	330928	Biolegend
PE/Cyanine7	CD58	330916	Biolegend

Construction and validation of cell scores

After screening the targets, integrated cell signaling scores were assigned to NK and OC cells. All subsequent analyses were performed using IBM SPSS Statistics (IBM, Armonk, NY, USA). Principal component analysis was used for dimensionality reduction, where the principal components were characterized by two key metrics: the factor coordinates (FAC), representing the loading scores of variables on principal components, and the eigenvalues (λ), quantifying the proportion of total variance explained by each principal component. A composite score was calculated by combining the principal component scores (FAC1, FAC2, FAC3, etc.) with their corresponding eigenvalues (λ1, λ2, λ3, etc.), that is, cell score = FAC × λ. Multiple NK cell samples were tested for their cytotoxicity against CR-D84 OC cells to evaluate the correlation between NK cell scores and killing efficacy. To compare the killing effects across different tumor cell lines, the cytotoxicity of each NK cell sample was normalized to its killing effect against the reference cell line HO8910. The correlation between tumor cell scores and killing efficacy was then evaluated. Linear regression analysis was used to compute the Spearman’s rank correlation coefficient (r), p-value (two-tailed), and root mean square error (RMSE) to evaluate the significant correlation between the cell score and NK cell-killing effect. Subsequently, the efficacy of NK cell therapy was categorized into two levels (better; worse) based on the median. The optimal cell score threshold was determined by maximizing Youden’s J statistic (sensitivity + specificity − 1), followed by group comparisons. An unpaired t-test was used to determine the significance of differences in overall treatment efficacy between low and high-score groups (p < 0.05).

Next, we developed a linear regression model using the two types of cell scores as independent variables and the killing effect as the dependent variable:

where W is the killing effect and β₀, β₁, and β₂ are the equation coefficients. The coefficients for the predictors in the cell score model were as follows: β₀ = 0.438, β₁ = 0.020, and β₂ = 0.029. The standardized residuals of the regression equation were analyzed using a Probability-Probability Plot (P-P plot), and the performance of the model was evaluated through goodness-of-fit and analysis of variance (ANOVA).

Construction and verification of the R-L model

For precise screening of NK and tumor cells, we normalized protein expression across different flow cytometry fluorescence channels. The fluorescence channel coefficient k was defined as the ratio of protein expression levels between different fluorescence channels, with one channel serving as the reference (Table 3). We then calculated the normalized MFI ratio for each molecule, where TIGIT-CD155 and DNAM1–CD112 were analyzed through separate integrated computations, and identified the combined predictor (CP) for each R-L interaction as follows:

Table 3.

The values of k (fluorescence channel coefficient)

Ovarian cancer cell	PE/APC	PC5/APC	PC7/APC	Pan-cancer tumor cell	PE/APC	P.P/APC	PC7/APC
OVCAR8	0.75	0.91	1.01	AGS	0.66	0.89	1.02
HO8910	0.69	0.90	0.91	HGC27	0.71	0.93	0.99
SKOV3	0.64	0.88	0.94	MKN45	0.74	0.92	1.54
COV504	0.65	0.88	0.93	MCF-7	0.67	0.91	0.96
HEYA8	0.72	0.95	0.98	MDA-MB-231	0.69	0.89	0.91
D84	0.68	0.90	0.87	MDA-MB-468	0.69	0.88	0.94
L93	0.67	0.84	0.83	BT549	0.69	0.90	0.95
A09	0.64	0.84	0.85	K562	0.65	0.90	0.96
LQQ	0.68	0.88	0.86	HEPG2	0.71	0.89	0.92
L82	0.68	0.83	0.82	U266	0.75	0.90	0.93
L106	0.63	0.79	0.79	A375	0.76	0.92	0.94
L48	0.66	0.85	0.80	SW620	0.73	0.92	1.02
L74	0.68	0.83	0.88	SK-BR-3	0.63	0.83	0.80
L83	0.68	0.90	0.97	MKN7	0.71	0.81	0.81
L100	0.66	0.84	0.82	MKN74	0.69	0.81	0.82
				HCC1954	0.68	0.79	0.82

Using a multi-model comparison approach, we identified the optimal predictive model. The OC dataset was partitioned into training (70%) and test (30%) sets using the ‘train_test_split’ function (‘test_size’ = 0.3). Four machine-learning algorithms were implemented: logistic regression, random forest, extreme gradient boosting (XGBoost), and support vector machine (SVM). Model performance was assessed through iteration using accuracy, positive predictive value (PPV)/precision, negative predictive value (NPV), and area under the curve (AUC). ROC analysis was performed using the ‘roc_curve’. The optimal classification threshold was determined through cross-validation on the training set. Using this threshold, the killing effect was dichotomized into binary outcomes: 1 (high susceptibility) and 0 (low susceptibility). Data preprocessing included handling missing values and Z-score standardization, followed by parameter optimization.

Univariate analysis was conducted to assess individual feature predictive power. Features with mean AUC > 0.6 across 20 iterations were selected. Feature importance was determined using model-specific metrics: Gini impurity (Random Forest), gain metrics (XGBoost), absolute coefficient values (logistic regression), and permutation importance (SVM).

The random forest model was implemented with Python 3.7, binary outcome probabilities were predicted with ‘RandomForestClassifier’ in scikit-learn, and data were manipulated using the ‘pandas’ and ‘numpy’ libraries. Model performance was evaluated using comprehensive metrics including the accuracy, precision, NPV, recall, F1 score, ROC curves, and AUC values. Hyperparameter optimization was conducted through GridSearchCV with 10-fold cross-validation, tuning the number of trees (100–300), tree depth (5–7), and minimum samples split (5–10). The hyperparameter combination yielding the best average performance across all folds was selected for the final model. Training and validation were repeated 20 times, with mean values reported. Feature importance was extracted using the built-in ‘feature_importances_attribute’, and visualizations were generated using the ‘matplotlib’ and ‘seaborn’ libraries. Model performance was evaluated in training, testing and external validation cohorts. A workflow of R-L model is shown in Supplementary Figure 2.

Cytotoxicity assay of acute myeloid leukemia (AML) cells

The use of peripheral blood samples from patients with AML was conducted in accordance with the Declaration of Helsinki and approved by the Ethics Committee of the Second Hospital of Anhui Medical University (Approval No. YX2023–046, valid from March 20, 2023 to March 19, 2024). All participants provided written informed consent prior to enrollment. Clinical trial number: not applicable. Blood samples were processed as follows: erythrocytes were lysed (lysis buffer: peripheral blood = 3:1) for 15 min and centrifuged at 500 × g for 5 min. The supernatant was discarded, and the cells were washed with 1× PBS, blocked for 15 min, and labeled with AC7–CD45, FITC-CD34, and BV610–CD117 antibodies. Subsequently, the cells were washed and transferred to new tubes, and flow cytometry was performed to detect the proportion of AML cells (CD45^+dim CD34⁺ CD117⁺). Simultaneously, the total number of peripheral blood cells was determined using a hemocytometer, and the density of AML cells was calculated based on their proportion. The AML cells were plated at 1 × 10⁵ cells per well. NK cells were resuspended in RPMI-1640 medium containing 10% FBS, and plated at an effector-to-target ratios of 5:1. The total volume was adjusted to 1 mL, mixed thoroughly, and placed in an incubator for co-culturing. Then, the cells were counted and labeled with antibodies to determine the proportion of AML cells remaining after treatment with NK cells. Each cytotoxicity assay was performed in triplicate to evaluate data consistency.

Tumor cell killing experiment using aneuploid drugs combined with NK cells

Well-maintained tumor cells at passage density were digested and counted, then seeded at an appropriate density in culture dishes. Aneuploid drugs BAY1217389 (50 nM; Selleck, S8215) and Hesperadin (100 nM; Selleck, S1529) were added to cells for treatment for 48 h. Thereafter, cells were digested and collected, washed twice with PBS, and the resulting cell suspension was used for both flow cytometry analysis and in vitro killing experiments. For the RTCA killing assay, cells were plated in E-plates at 1.5–2.0 × 10⁴ cells per well and placed in an incubator for monitoring. When cells reached plateau phase (at approximately 13 h), effector cells were added. After calculating and adjusting the NK cell density, the plate was removed and 50 μL of NK cell suspension was added to each well in the experimental groups and continuously monitored. After 4 h, the killing assay was stopped. The experiment was repeated three times independently (n = 3). Each cytotoxicity assay was performed in triplicate to evaluate data consistency.

Statistical analysis

Statistical analysis of all relevant data was performed using GraphPad Prism 9.0 (GraphPad Software), IBM SPSS Statistics (IBM), and Python software. Pearson’s or Spearman’s coefficients were used to analyze the correlations between variables. When comparing inter-group differences in data, the t-test (paired or unpaired) was used for data with a normal distribution and homogeneity of variance; otherwise, non-parametric tests such as the Mann–Whitney U test (for unpaired samples) or Wilcoxon signed-rank test (for paired samples) were used. ROC curves were used to calculate the accuracy of the predictive model. A p-value < 0.05 was considered statistically significant.

Results

Cell cytotoxicity sensitivity and ligand expression profiles differ by tumor type

To assess differential NK cell-mediated cytotoxicity across distinct tumor cell lines, we performed in vitro cytolytic assays using FF-NK cells (batch number: FF-NK-0330O) against various cancer cell lines (Supplementary Figure 3A). Tumor cells from different cancer types exhibited varying sensitivity to NK cells. Although conventional cell lines lack the complex heterogeneity of primary tumors and cannot accurately represent tumor responses to treatment, CR cells maintain a similar phenotype, genetic, and heterogeneity of primary tumors while being capable of unlimited proliferation in vitro [21–23]. Based on the results of NK cell cytotoxicity assays (batch number: FF-NK-1130AB) conducted using CR OC cells, cells from different patients responded differently to NK cell-mediated therapy (Fig. 1A).

Fig. 1.

Screening receptor-ligand pairs for predictive model construction. (A) Cytotoxicity of FF-NK-1130AB cells against conditionally reprogrammed OC cells and cell lines was assessed using RTCA. (B) Immunohistochemical staining of B7–H6 and PCNA expression in OC tissues from different patients (data from the Human Protein Atlas database). (C) Differential expression analysis of OC (vs. normal) samples (n = 32, GSE54388) for 25 selected ligands. Significantly upregulated (log₂ FC > 1, p ＜ 0.05; red) and downregulated (log₂ FC < −1, p ＜ 0.05; blue) genes are highlighted with direct annotations in the panel. (D) Bubble chart showing the impact of ligands on NK cell infiltration in the Tumor Immune Estimation Resource 2.0 database (x-axis = cancer type; y-axis = NK cell-related ligands). Color and size of symbols indicate the impact of ligand expression on NK cell infiltration; red (blue) represents a positive (negative) correlation; symbol size is -log10 (adj. p-value), where larger symbols indicate stronger correlations

To validate differences in the protein expression profiles of tumor cells of different origins, we investigated the expression of several important ligands in various cancers using the Oncolnc website. Significant differences in ligand expression were observed among the different cancer types (Supplementary Figure 3B, Table 4). We selected several tumors (OV, SKCM, STAD, BRCA, COAD, LIHC) with potentially high sensitivity to NK cell therapy for further study. First, we analyzed ligand expression in OC using the HPA database. Immunohistochemistry revealed that the expression of proteins, such as B7–H6 and PCNA, varied among tumors from different patients, confirming expression profile heterogeneity within a specific cancer type across different individuals (Fig. 1B).

Table 4.

Expression of ligands involved in NK cell-mediated cytotoxicity across various cancer types (oncolnc database)

	MICA	MICB	ULBP3	DR4	DR5	PCNA	HLA-C	B7–H6	CD155	ICAM1	CD112	CD58
OV	276.61	171.04	36.64	67.05	736.28	2531.96	15680.79	32.64	838.60	2121.68	5088.89	190.19
SKCM	407.61	176.94	7.80	25.95	1531.29	2597.11	23048.47	37.86	992.04	4669.67	1326.39	349.50
STAD	271.40	196.10	52.16	203.35	1550.06	2962.94	19420.74	311.84	1609.49	1944.41	4425.33	356.54
BRCA	285.81	154.57	7.94	102.61	744.76	2319.17	16625.30	34.10	509.52	1117.68	4622.13	189.21
COAD	313.27	127.60	54.35	237.99	1869.09	3698.95	26481.78	271.51	1704.31	1024.43	2750.92	308.51
LIHC	293.96	95.92	2.70	68.55	1149.21	1628.35	23601.61	7.36	2053.19	1885.32	5019.20	175.33
ESCA	226.41	203.7	94.15	202.82	1537.71	2862.71	15616.68	223.49	1550.85	1402.16	2625.19	367.41
LUAD	334.12	196.84	35.11	167.66	1779.86	1648.76	26810.15	24.48	1192.33	8246.9	3011.82	267.14

Selection of R-L pairs for model construction

To investigate the relationship between cytotoxicity and expression profile differences, we screened predictive variables based on three criteria: (1) tumor immune response relevance, validated through literature and databases (GEPIA, TISIDB, and TIMER2.0) for NK cell activity and patient survival; (2) differential expression patterns, confirmed by GEO database and flow cytometry; and (3) tumor cell expression, verified through experimental methods. First, we selected NK cell sensitivity-related ligands (B7–H6, MICA, MICB, ULBP1, ULBP3, BAT-3, HSPG2, NID1, MLL-5, CFP, DR4, DR5, PCNA, HLA-E, HLA-C, HLA-G, CD200, CD155, CD112, CD54, CA125, CD58, CD48, PD-L1, and TGF-β1) and their corresponding receptors based on previous reports (Table 5).

Table 5.

A summary of NK cell function-related receptors and ligands identified from prior studies

Function	NK Receptor	Tumor ligand	References
NK cell activating receptors	NKp30	B7–H6	[28, 29]
		BAT-3	[30]
		HSPG2	[31]
	NKG2D	MICA	[32, 33]
		MICB
		ULBP1
		ULBP3
	TRAIL	DR4	[34, 35]
		DR5
	NKp44	PCNA	[36, 37]
		NID1	[38]
		MLL-5	[39]
	NKp46	CFP	[40]
NK cell inhibitory receptors	NKG2A	HLA-E	[41]
	KIR2DL1–3	HLA-C	[42]
		HLA-G	[43]
	CD200R	CD200	[44]
	TIGIT	CD155	[45]
	SIGLEC9	CA125	[46]
	PD-1	PD-L1	[47]
Adhesion receptors	DNAM1	CD112	[48]
	LFA-1	CD54	[49]
Co-Stimulatory receptors	CD2	CD58	[50]
	2B4	CD48	[51]
Cytokine receptors	TGF-βR	TGF-β1	[52]

We then performed differential expression analysis on NK cell-related ligands using multiple cancer datasets from GEO. In OC (GSE54388), volcano plot analysis revealed significant upregulation of PCNA, HSPG2, and CA125, whereas B7–H6 was downregulated compared to normal tissues (Fig. 1C). Similar expression patterns were observed in breast (GSE36295), gastric (GSE79973), and liver cancer (GSE144269) datasets (Supplementary Figure 3C, Table 6). In the GEPIA database, univariate Cox regression analysis was used to screen tumor ligands associated with prolonged OS of patients. In OC, MICB and HLA-C expression was significantly positively correlated with OS, whereas ULBP4 expression was negatively correlated (Supplementary Figure 3D). Key ligand correlations varied across cancer types, with distinct positive and negative regulatory patterns observed in breast, gastric, colorectal, and liver cancer (Supplementary Figure 4A). Finally, we examined the influence of ligand expression across different cancer types on NK cell infiltration using TISIDB and TIMER2.0 databases. Spearman correlation analysis revealed that CD200 had a greater impact on NK cell infiltration (Fig. 1D). In OC, MICB, HLA-G, CD58, CD155, DR4, DR5, HLA-C, and CA125 were significantly positively correlated with NK cell infiltration, whereas PCNA, CD200, and ULBP3 were significantly negatively correlated (Supplementary Figure 4B). The ligands in other tumor cell types are listed in Supplementary Figure 5.

Table 6.

Analysis of differential expression of ligands (GEO database)

NK cell receptors	Tumor cell ligands	OV		BRAC		STAD		LIHC
		P.Value a	logFC b	P.Value	logFC	P.Value	logFC	P.Value	logFC
NKp30	B7–H6	0.000	−1.630	0.032	0.798	0.168	0.603	0.000	0.938
	BAT-3	0.454	0.085	0.024	0.597	0.103	0.346	0.002	0.214
	HSPG2	0.005	1.206	0.673	0.270	0.001	0.956	0.000	0.702
NKG2D	MICA	0.300	−0.501	0.227	0.256	0.596	0.148	0.168	−0.197
	MICB	0.187	0.873	0.007	1.049	0.145	0.492	0.000	0.960
	ULBP1	0.459	−0.165	0.027	0.492	0.581	−0.293	-	-
	ULBP3	0.927	0.011	0.561	−0.228	0.389	0.501	-	-
TRAIL	DR4	0.035	0.449	0.132	0.367	0.409	0.191	0.090	0.190
	DR5	0.256	−0.213	0.280	0.260	0.007	1.604	0.000	−0.447
NKp44	PCNA	0.000	2.160	0.005	0.830	0.036	0.562	0.000	1.301
	NID1	0.082	−0.512	0.140	−0.759	0.336	−0.228	0.719	0.046
	MLL-5	0.065	0.810	0.025	0.363	0.601	−0.080	0.000	0.272
NKp46	CFP	0.015	−0.850	0.015	−0.511	0.509	−0.416	0.000	−2.489
NKG2A	HLA-E	0.277	0.472	0.141	0.463	0.272	−0.212	0.002	−0.347
KIR 2DL1–3	HLA-C	0.109	0.632	0.009	0.650	0.715	−0.092	0.202	−0.168
	HLA-G	0.932	0.037	0.012	0.603	0.268	0.299	0.075	0.332
CD200R	CD200	0.307	−0.932	0.906	−6.441	0.018	0.676	0.000	1.670
TIGIT	CD155	0.070	−0.323	0.542	−6.262	0.187	−0.729	0.116	0.172
DNAM1	CD112	0.084	−0.312	0.002	−1.596	0.022	0.358	0.128	0.150
LFA-1	CD54	0.620	0.218	0.113	0.717	0.013	0.774	0.025	0.408
SIGLEC9	CA125	0.013	2.363	0.484	0.378	0.168	1.028	0.934	−0.027
CD2	CD58	0.014	−0.400	0.557	−0.136	0.188	0.328	0.000	0.703
2B4	CD48	0.418	−0.104	0.451	0.628	0.048	−0.561	0.499	−0.144
PD-1	PD-L1	0.086	−0.277	0.017	1.425	0.283	0.416	0.087	−0.291
TGF-βR	TGF-β1	0.056	0.416	0.624	−0.198	0.391	−0.216	0.007	0.457

^a P.Value: Adjusted p-value (Benjamini-Hochberg) < 0.05, indicating statistically significant differential expression

^b LogFC: Log2 fold-change (log2-transformed ratio of expression between groups); thresholds of |logFC| > 1.0 (2-fold change) highlight biologically relevant genes

To further observe the actual expression levels of these ligands for screening, we performed flow cytometry on the ligand expression profiles of CR OC cells from nine different patients. MICA, ULBP1, DR4, CA125, CD155, CD58, and CD54 showed notable expression, whereas PD-L1, CD48, HLA-E, and TGF-β1exhibited minimal expression (Fig. 2A, Table 7). In parallel, receptor profiling of ten in vitro expanded peripheral blood NK cells revealed significant expression variations in NKp30, NKp44, KIR2DL1, KIR2DL4, and DNAM1, with minimal differences in NKG2A, PD-1, and LFA-1 (Fig. 2B, Table 8). Coefficient of variation analysis confirmed heterogeneous expression patterns across both cell populations (Fig.s 2C and D). Therefore, we excluded some receptors and ligands with minimal expression and ultimately selected 11 R-L pairs (NKG2D-MIC/ULBP, NKp30-B7H6, NKp44-PCNA, TRAIL-DR4/5, KIR2DL1–3-HLA-C, KIR2DL4-HLA-G, CD200R-CD200, TIGIT-CD155, SIGLEC-9-CA125, DNAM1–CD112, and CD2–CD58) as predictive variables to evaluate the interaction between NK and tumor cells (Fig. 2E).

Fig. 2.

Confirmed and established predictive variables in the prediction model. (A) Heatmap of protein expression profiles for nine reprogrammed OC cell lines. Data represent natural log-transformed ratios of mean fluorescence intensity (MFI) from flow cytometry analysis, with the expression level of each protein calculated as the mean of five independent experimental replicates. Color scale indicates relative expression (blue: low; red: high). (B) Heatmap of receptor expression profiles for NK cells expanded from ten different sources of peripheral blood mononuclear cells, with the data Ln-transformed. (C) Intercellular heterogeneity of ligand expression in OC cell lines. The line plot displays the dispersion coefficient (CV = SD/mean), the expression of each ligand across nine reprogrammed OC cell lines. (D) Calculate the dispersion coefficient (CV) for each receptor expression in NK cells expanded from ten different sources. (E) Schematic of predictive variables in the cytotoxicity model. Key receptor-ligand pairs (validated predictive variables) are shown, with NK cell receptors (left) and tumor-expressed ligands (right)

Table 7.

Expression profiles of ligands in CR ovarian cancer cells during the pre-discovery stage

Tumor cell ligands	L48	L62	L96	L92	L8	L57	L58	L37	L83
B7–H6	1.13	1.06	1.16	1.18	1.11	1.25	0.71	0.86	1.60
MICA	1.03	1.12	1.11	1.44	1.34	1.53	0.93	1.37	1.40
MICB	1.18	0.69	1.04	1.05	1.08	0.82	0.79	0.75	0.87
ULBP1	1.12	1.07	1.20	1.22	1.23	1.39	0.70	0.88	1.18
ULBP3	1.80	1.28	1.59	1.60	1.41	1.40	0.72	0.95	1.14
ULBP4	1.07	1.08	1.21	1.13	1.33	1.20	0.65	1.86	1.49
DR4	1.54	1.06	1.81	1.33	2.02	2.57	0.82	1.22	1.84
DR5	7.75	6.34	7.20	7.39	8.11	10.96	11.75	4.85	1.15
PCNA	2.70	1.42	2.91	2.54	2.97	2.74	-	1.79	3.04
HLA-E	1.11	1.00	1.14	1.08	1.21	1.32	0.71	1.11	1.31
HLA-C	1.34	1.24	1.45	1.32	3.04	1.46	0.79	0.90	1.04
HLA-G	1.49	1.49	2.19	1.88	1.10	2.44	1.05	1.19	1.51
CD200	142.43	112.68	186.24	85.32	143.83	210.67	118.78	52.42	10.53
CD155	28.30	35.88	46.83	79.21	41.34	45.75	117.25	23.90	225.96
CD112	20.99	10.53	14.83	16.42	21.71	23.23	11.81	17.72	4.59
CD54	138.35	188.55	570.38	459.83	245.54	274.17	2364.74	413.33	35.88
CA125	11.13	36.38	27.42	20.38	40.71	44.84	24.46	31.09	173.48
CD58	12.70	8.12	16.72	15.74	18.26	19.95	25.21	11.56	24.95
CD48	1.34	0.90	1.41	1.36	1.32	1.42	0.84	0.85	1.07
PD-L1	1.13	0.77	1.05	1.02	0.92	0.76	0.86	0.32	1.05
TGF-β1	1.05	0.98	1.09	1.06	1.06	1.00	-	0.94	1.20

Table 8.

Expression profiles of receptors in PBMCs-NK cells during the pre-discovery stage

NK cell receptors	FF-NK-0317O	FF-NK-0329A	FF-NK-0419A	FF-NK-0531A	FF-NK-0607O	FF-NK-0628A	FF-NK -0705A1	FF-NK-0705A2	FF-NK-0811O	FF-NK-0818B
NKG2D	73.43	156.89	636.57	164.05	175.51	421.57	274.26	195.08	209.28	101.46
NKp30	144.37	54.24	346.02	39.18	61.49	282.01	129.08	135.79	47.35	60.60
NKp44	35.37	16.56	75.66	26.23	34.50	27.07	154.54	24.62	8.03	17.45
TRAIL	1.78	5.15	5.39	3.88	1.51	7.22	8.65	4.90	11.14	8.23
NKG2A	0.89	1.13	1.39	1.18	1.12	1.29	1.21	1.25	1.08	1.54
KIR2DL1	40.47	93.35	60.99	10.82	16.78	73.30	35.36	115.44	188.55	108.84
KIR 2DL2/3	446.59	590.56	911.90	504.61	314.79	834.27	634.86	526.44	985.17	236.76
KIR3DL1	7.12	17.21	17.45	1.19	6.35	12.04	5.82	16.51	1.00	9.48
KIR2DL4	2.17	4.57	5.12	1.50	2.17	10.34	2.63	3.17	2.01	6.59
PD-1	1.09	1.04	1.13	1.06	1.18	1.07	1.79	1.49	1.57	1.09
CD200R	2.71	3.47	24.91	5.40	10.68	2.61	6.49	4.06	4.29	9.59
TIGIT	24.91	29.43	57.15	34.97	34.24	33.46	42.09	18.32	11.53	11.33
Siglec-9	1.96	1.65	2.12	1.08	3.91	1.57	1.32	1.90	0.59	1.13
DNAM1	60.09	47.27	259.84	64.85	60.78	54.78	218.22	59.03	43.06	29.82
LFA-1	1.23	1.58	1.53	1.12	1.21	1.12	1.11	1.20	1.11	1.40
CD2	79.73	59.20	24.40	75.61	98.10	68.61	83.56	75.40	121.22	68.15
2B4	42.51	31.69	10.52	5.55	16.74	15.84	11.65	12.58	18.54	6.64

Integrated R-L interactions determine NK cell cytotoxicity through complex nonlinear relationships

To investigate the relationship between expression profile differences and NK cell therapy efficacy, we evaluated the cytotoxicity of 76 distinct NK cell samples against 14 OC cell lines. Different cell lines responded differently to individual NK cell samples, and individual cell lines responded differently to different NK cell samples (Supplementary Figure 6A, Fig. 1A). Next, based on the markers screened in Sect. 3.2, we used flow cytometry to determine the expression profiles of both NK and OC cells (Fig. 3A, Supplementary Fig. 6B, Table 9, Supplementary Table 1). NK cells were categorized based on killing efficacy from five cytotoxicity experiments. Comparison of receptor expression between high and low cytotoxicity NK cells revealed no significant differences, except for CD200R (Fig. 3B). We ranked OC cells by their sensitivity to NK cells and analyzed corresponding ligand expression trends. Statistical analysis revealed no significant correlation between NK cell-mediated cytotoxicity and ligand expression levels on OC cells (Supplementary Figure 7A). Multidimensional analysis further confirmed the absence of a direct linear relationship between cytotoxicity and the expression of R-L pairs (Fig. 3C).

Fig. 3.

Cytotoxic effect of NK cells is influenced by multifactorial synergistic interactions. (A) Heatmap of ligand expression profiles for OC cells, with natural log transformation of detected MFI ratio data. (B) Comparison of receptor expression profiles between better (≥70% tumor lysis) and worse (<30% tumor lysis) killing effect NK cell subsets. (C) Bubble chart showing correlation trends between differences in NK cell killing and expression levels of R-L pairs. x-axis: NK cell receptor expression (ln MFI radio); y-axis: Tumor cell ligand expression (ln MFI radio); Color and size: Magnitude of NK cell killing effect. (D) A significant linear relationship exists between the established NK cell score and cytotoxicity (y-axis = killing effect of NK cells on CR-D84 OC cells; p = 0.0081 (< 0.05)). (E) Linear relationship between tumor cell score and NK cell cytotoxicity (y-axis = the normalized killing effect of NK cells; p = 0.0022 (< 0.05)). (F) Three-dimensional scatter plot of the linear regression model (x-axis = tumor cell score; z-axis = NK cell score; y-axis = killing effect). (G) Predicted versus observed scatter plot of the model (slope = 0.12; line Y = X represents the ideal prediction line); x-axis: actual cytotoxicity assay results; y-axis: model-predicted cytotoxicity values

Table 9.

Expression profiles of ligands in ovarian cancer cells during the modeling phase

	OVCAR3	HO8910	SKOV3	HEYA8	L91	L96	L92	L93	D84	L83	L48	L62	A09	LQQ
MIC-A	7.59	7.47	1.13	2.13	0.93	1.11	1.44	1.42	1.01	1.77	1.04	1.12	1.21	1.35
MIC-B	0.92	1.09	1.05	1.19	0.95	1.04	1.05	1.06	0.99	1.04	1.05	0.69	1.10	1.16
B7–H6	1.28	1.15	1.09	1.01	0.96	1.16	1.18	1.01	1.04	1.13	1.05	1.06	1.03	1.02
ULBP1	5.28	2.00	6.86	1.86	1.06	1.20	1.22	1.02	1.04	1.30	1.06	1.07	1.11	1.06
ULBP3	4.89	1.05	1.85	2.09	1.36	1.59	1.60	1.09	1.15	1.42	1.80	1.28	1.06	1.09
ULBP4	1.16	1.20	1.16	1.04	1.12	1.21	1.13	1.03	1.05	1.18	1.01	1.08	1.07	1.08
DR4	8.86	4.95	3.37	6.67	1.51	1.81	1.33	2.22	1.86	1.29	2.12	1.06	1.44	1.84
DR5	11.88	3.85	3.91	17.80	4.93	7.20	7.39	5.09	4.16	1.83	7.75	6.34	2.52	5.79
PCNA	6.43	3.10	4.91	6.52	2.41	2.91	2.54	2.97	2.27	2.77	2.70	1.42	2.05	2.49
HLA-C	0.98	2.46	1.29	1.12	1.41	1.45	1.32	2.66	1.93	1.27	1.38	1.24	2.32	6.63
HLA-G	1.51	1.35	1.58	1.86	1.91	2.19	1.88	1.42	1.44	1.54	1.53	1.49	1.40	1.50
CD200	247.59	1.21	1.48	2.84	55.34	186.24	85.32	107.71	146.39	20.29	164.37	112.68	124.74	160.04
CD155	411.51	519.15	601.60	420.07	16.02	46.83	79.21	41.61	64.15	418.58	36.03	35.88	18.91	56.90
CA125	812.63	18.52	24.16	1.69	30.73	27.42	20.38	9.96	38.93	705.28	14.41	36.38	23.76	141.07
CD58	188.53	108.99	59.88	88.93	33.02	16.72	15.74	13.18	17.74	30.61	14.33	8.12	12.89	20.86
CD112	10.18	11.13	5.52	25.75	29.50	14.83	16.42	13.52	15.01	6.72	20.99	10.53	12.04	18.18

To model NK cell cytotoxicity, we analyzed composite cell scores derived from PCA (NK cell score and tumor cell score; Tables 10 and 11). A significant linear relationship was observed between the cellular composite scores and NK cell cytotoxicity (p = 0.0081, RMSE = 0.203, F = 7.718, R² = 0.1552; Fig. 3D). For tumor cells, linear regression also revealed a strong association between tumor cell scores and killing efficacy (p = 0.0022, RMSE = 0.17, F = 15.12, R² = 0.5576; Fig. 3E). This indicates that the cytotoxic effect was significantly associated with the receptor and ligand expression profiles. Stratification of cells into high- and low-score groups based on optimal ROC thresholds further supported this association, with significantly reduced treatment effects observed in low-score groups (NK cell score: p = 0.0012; tumor cell score: p = 0.0089; Supplementary Figure 7B, C). These findings indicate that the cytotoxic effect of NK cells results from the synergistic action of multiple factors, with no single R-L pair exerting a decisive influence.

Table 10.

The NK cell scores

Sample	Score	Sample	Score	Sample	Score
FF-NK-1014O	1.17	MB-NK-B	−0.44	FF-NK-0202A	−0.31
IPSC-NK-0923	−0.55	FF-NK-0628	2.28	feeder-NK-1011A	−1.16
FF-NK-1103D	0.89	FF-NK-0303 H	0.00	feeder-NK-1011O	−1.09
FF-NK-1117–2	−0.31	FF-NK-0114	−0.36	FF-NK-0303D	−0.38
FF-NK-0929O-2	−0.16	FF-NK-1117O	−0.22	feeder-NK-0929O	−0.53
MB-NK-1116	0.75	FF-NK-0929B-2	−0.25	FF-NK-0222O	0.38
feeder-NK-1103	0.93	FF-NK-0315A	−0.67	feeder-NK-1011A-2	−0.14
FF-NK-0607O	−0.28	FF-NK-0921B	1.25	FF-NK-0217B	−0.50
FF-NK-1014	0.88	feeder-NK-0511	−0.16	feeder-NK-0929B	−0.61
FF-NK-0315D	1.02	FF-NK-0825A	−0.20	FF-NK-0823	0.25
feeder-NK-1116	−0.19	MB-NK-0315	0.69	feeder-NK-0217B	−0.50
FF-NK-0531A	−0.56	FF-NK-0818B	−0.56	IPSC-NK-1102	−0.69
MB-NK-A	−0.47	FF-NK-0427A-2	0.31	feeder-NK-0202	0.20
FF-NK-1014AB	0.14	feeder-NK-1014	−0.33	FF-NK-1129B	−0.86
MB-NK-1027	1.30	FF-NK-0427A-3	0.04	-	-

Table 11.

The ovarian cancer cell scores

Tumor cell	score	Tumor cell	score
OVCAR3	2.25	L83	−0.22
HO8910	0.27	L48	−0.27
SKOV3	0.34	L62	−0.71
HEYA8	1.57	A09	−0.95
L91	−0.17	LQQ	−0.94
L96	0.15	COV504	0.27
L93	−0.71	OVCAR8	0.71
L92	−0.01	L82	−0.37
D84	−0.68	L106	−0.52
L8	0.83	-	-

When expanded NK cells are used to treat tumors, both the donor and recipient vary; thus, understanding the dynamic pairing between NK and tumor cells is crucial. Here, we assigned cell scores to all NK and tumor cells to construct a predictive model of cytotoxicity using linear regression. Although the standardized residuals exhibited normal distribution (Supplementary Figure 7D), the poor fit indicates that the model may lack meaningful explanatory power (R² = 0.008, F = 0.644, p = 0.527). The 3D scatter plot revealed distinct nonlinear patterns (Fig. 3F), and the predicted versus observed scatter plot showed significant deviation from the ideal prediction line (slope = 0.12; Fig. 3G). Thus, cell score-based prediction models may not be suitable for modeling NK cell cytotoxicity.

In contrast to previous results, the slopes for both tumor and NK cell scores were not statistically significant (p > 0.05), indicating no linear relationship with cytotoxicity. The variance inflation factor analysis ruled out multi-collinearity between cell types (VIF = 1.015, tolerance = 0.985), these results reveal a complex nonlinear interaction between tumor and NK cells that collectively determines cytotoxicity (the data used to build the model are provided in Supplementary Table 2).

R-L pairing enables precise screening in the predictive model

To improve the model and achieve precise screening of NK and tumor cells, we identified the combined predictor for each R-L interaction as an independent variable and established the R-L predictive model through machine learning. First, we used the OC dataset comprising 52 NK and 14 OC cell samples to evaluate the performance of logistic regression, random forest, XGBoost, and SVM algorithms based on R-L pairs. The dataset was split 7:3 into training and test sets, and the training–validation process was repeated 20 times to estimate confidence intervals for accuracy, precision, NPV, recall, and AUC based on the test data.

Following visualization of feature distributions and threshold optimization, we converted the target variable to binary classification (Supplementary Figure 8A). Random forest models provided the best overall performance, whereas XGBoost showed overfitting on the training set (Fig.s 4A). Univariate analysis identified ten preprocessed variables with predictive value (AUC > 0.6, Fig. 4B), with random forest and XGBoost algorithms showing balanced performance across variables. Pairwise combination analysis revealed the predictive value of R-L combinations (Supplementary Figure 8B). The variable importance rankings for all models are shown in Supplementary Figure 8C. After evaluating performance metrics, we selected the random forest algorithm as the optimal machine learning model.

Fig. 4.

Establishment of the R-L predictive model in OC. (A) Comparative evaluation metrics (prediction accuracy, positive predictive value (PPV), negative predictive value (NPV), and AUC) of four models trained on the OC dataset. x-axis: the types of models; y-axis: the model metrics. (B) AUC performance of univariate analysis for ten features across different machine learning models. x-axis: predictive variables of the model; y-axis: the AUC performance of the model; bar colors: different model types. (C) Comparison of four prediction metrics of the random forest model at different classification thresholds in the test set, with an optimal classification threshold of 0.55. (D) Variable importance of each variable in the OC random forest model. (E) ROC curve of the OC R-L prediction model. Red line: performance on the independent external test set; Blue line: performance on training set. (F) Scatter plot showing the correlation between predicted probabilities and actual cytotoxicity values. x-axis: measured cytotoxicity (NK cell killing efficacy from experimental assays); y-axis: predicted probability of cytotoxicity (R-L model output, where values ≥ threshold classify as 1 (high efficacy)); point color: dataset classification (train set vs. test set)

The final random forest model was trained using all ten variables, with an optimal classification threshold of 0.55 (Fig. 4C). CD200, TIGIT, and DNAM1were the most influential R-L pairs (Fig. 4D). The trained model was also tested on an independent cohort (Supplementary Figure 8D and E, Tables 12 and 13, Supplementary Table 3), achieving an accuracy of 0.842 and AUC of 0.908; the same model showed an accuracy of 0.888 and AUC of 0.921 for the training set (Fig. 4E, Table 14). Finally, matrix scatter plots revealed a significant positive correlation between the R-L model’s predicted probabilities of cytotoxicity and experimentally measured NK cell cytotoxicity levels (r = 0.7433, p < 0.0001, Fig. 4F), showing applicability of the model for quantitative analysis beyond basic stratification. These results collectively reveal the good overall performance of the model using R-L pairs as predictive variables.

Table 12.

Expression profiles of ligands in ovarian cancer cells from the test set

	OVCAR8	COV504	L106	L82	L74	L100
MIC-A	0.99	0.91	1.11	0.90	5.94	0.79
MIC-B	1.92	1.19	0.94	1.05	1.14	1.02
B7–H6	1.08	0.94	0.96	0.95	1.11	0.99
ULBP1	13.08	9.61	1.15	2.24	5.14	0.87
ULBP3	1.44	1.13	0.97	1.09	1.38	1.08
ULBP4	0.99	0.88	0.93	0.94	1.08	1.10
DR4	2.15	2.79	4.08	2.07	3.38	1.65
DR5	9.72	10.28	11.34	22.52	10.94	23.23
PCNA	3.82	13.88	11.57	8.20	4.89	5.25
HLA-C	1.02	1.23	5.08	3.41	3.27	1.21
HLA-G	1.65	1.72	1.23	1.28	1.51	1.16
CD200	1.25	1.88	558.68	670.89	2.23	211.65
CD155	405.32	411.64	225.14	262.55	107.96	104.51
CA125	27.79	8.45	3.79	10.13	15.66	3.63
CD58	148.58	96.76	51.53	37.98	35.79	48.74
CD112	76.93	34.51	115.45	79.64	15.82	27.45

Table 13.

Expression profiles of receptors in NK cells from the test set

	NKp30	NKG2D	TRAIL	NKp44	KIR2DL 1/2/3	KIR2DL4	CD200R	TIGIT	Siglec-9	CD2	DNAM1
feeder-NK-0307	123.81	274.98	13.17	21.89	445.29	2.63	6.88	28.61	2.20	680.42	79.57
feeder-NK-0718	172.53	266.72	11.30	9.66	717.75	9.81	13.49	96.42	1.64	3049.61	212.50
feeder-NK-1206	119.69	147.95	14.50	5.98	321.62	4.96	7.38	20.37	4.05	902.89	110.85
feeder-NK-1220	188.85	245.93	28.74	23.65	258.86	5.13	9.20	28.74	4.80	1276.43	197.96
FF-NK-0212	265.34	180.38	27.89	414.24	1037.56	33.37	18.12	32.98	3.10	1445.33	221.20
FF-NK-0227–2	78.17	177.47	10.14	243.90	1768.09	11.54	3.36	31.68	3.61	5157.06	196.80
FF-NK-0227	96.11	214.72	6.50	21.25	592.87	8.32	9.10	15.53	3.46	1393.17	316.01
FF-NK-0307	100.57	156.21	22.96	221.22	1969.22	4.57	8.07	23.13	1.95	6050.92	150.20
FF-NK-0423	337.61	345.53	22.26	371.17	2882.28	3.05	27.94	121.81	1.62	2361.22	176.70
FF-NK-0423DC	313.35	226.35	29.77	214.12	2405.87	8.59	21.16	66.54	2.46	3136.55	183.10
FF-NK-0423–2	237.72	206.17	10.76	353.97	1707.03	7.72	13.48	50.46	10.65	2882.66	271.41
FF-NK-0509	242.30	297.29	10.55	98.45	1179.84	2.19	17.74	40.96	2.90	3060.71	134.35
FF-NK-0514	130.81	235.23	27.23	71.81	1458.17	7.81	11.44	48.14	1.96	2737.28	250.32
FF-NK-0514–2	196.21	454.99	16.13	102.14	1784.46	10.17	9.36	150.48	2.61	4854.99	242.08
FF-NK-0606	78.34	416.93	16.26	4.16	889.60	8.73	22.88	61.94	0.92	4712.50	287.22
FF-NK-0618	181.56	509.05	37.13	100.59	42.69	13.32	13.28	63.86	5.46	3263.85	197.25
FF-NK-0618–2	160.72	246.80	20.01	190.90	1227.08	7.67	13.30	47.68	3.08	3402.77	232.99
FF-NK-0704	349.71	275.38	21.65	118.47	1114.25	9.46	30.01	53.56	3.37	5451.60	225.61
FF-NK-0704D	94.00	176.63	2.52	72.77	90.74	4.33	3.83	51.69	12.82	1561.70	325.61
FF-NK-0718	286.20	565.34	73.00	166.53	947.28	11.34	22.48	96.64	1.49	3283.25	88.69
FF-NK-0718–2	309.19	371.96	64.36	5382.66	3685.42	20.70	32.11	109.49	6.60	43616.92	1198.82
FF-NK-0813	529.50	512.87	45.36	230.56	1862.98	25.86	26.17	66.49	3.52	4324.69	221.39
FF-NK-0813D	386.64	900.88	13.48	897.02	7678.78	28.79	36.89	81.65	6.33	5681.55	794.58
FF-NK-0830	576.50	828.78	126.16	93.34	2938.49	60.71	16.81	133.73	4.51	21588.05	769.40
FF-NK-0830–2	124.86	118.68	2.43	110.48	1189.61	8.63	5.04	32.68	2.02	5015.96	222.19
FF-NK-0910	374.29	340.64	15.59	248.93	1047.86	24.38	13.17	73.50	3.27	1832.88	157.93
FF-NK-0910–2	195.59	205.03	8.86	162.65	735.22	11.43	7.11	35.79	5.41	2383.36	101.47
FF-NK-1128	167.85	432.03	19.42	68.18	1758.15	27.96	31.70	79.95	3.79	8362.22	371.22
FF-NK-1128–2	1068.36	669.90	16.81	537.82	3266.70	112.39	81.68	102.59	2.43	7835.05	290.45
FF-NK-1206	357.47	144.96	6.17	96.29	1007.84	21.51	18.86	19.77	27.70	1132.60	135.79
FF-NK-1220	295.51	350.83	31.38	205.42	669.19	27.64	19.65	45.06	5.97	1810.59	159.51
FF-NK-1220–2	348.35	315.79	29.12	461.77	1694.19	44.84	29.09	51.65	2.05	2311.32	99.09
IPSC-NK-0405	75.75	46.80	65.95	21.71	79.27	4.92	24.45	4.50	1.83	573.51	227.97
IPSC-NK-0412	69.46	35.67	22.76	26.48	7.28	2.11	20.39	1.16	1.00	113.23	147.49
IPSC-NK-0608	88.88	100.33	16.91	68.93	18.90	2.55	25.00	1.36	1.63	120.71	217.59
IPSC-NK-0621	62.91	50.24	23.68	18.17	8.62	2.79	21.46	1.54	0.92	58.40	152.87

Table 14.

Prediction performance of the OC R-L model in the test set

Prediction accuracy (training data)	Prediction accuracy (test data)	PPV a (training data)	PPV (test data)	Recall b (training data)	Recall (test data)	AUC c (training data)	AUC (test data)
0.888	0.842	0.862	0.833	0.920	0.833	0.921	0.908

^a Positive predictive value: true positive samples/(true positive samples + false positive samples)

^b Recall (True Positive Rate): true positive samples/(true positive samples + false negative samples)

^c AUC: area under receiver operating characteristic curve

Applicability of the predictive model to pan-cancer studies

Based on the established R-L model for OC, we expanded our scope to include pan-cancer analysis, examining the applicability of the model to other cancers. Similarly, we assessed the ligand expression profiles of different tumor cells, along with their corresponding NK cell sensitivity characteristics (Table 15, Supplementary Figures 9A and B). Preliminary correlation analysis verified that the expression profiles of various tumor cells, including OC cells, were not significantly correlated with killing sensitivity (Supplementary Figures 9C and D). The same result was observed for NK cell surface receptors. No individual R-L pair exerted a notable impact on cytotoxicity.

Table 15.

Expression profiles of ligands in tumor cells across multiple cancer types from the training set

	AGS	HGC27	MKN45	MDA-MB-231	MDA-MB-468	BT549	MCF-7	K562
MIC-A	1.02	1.01	2.19	3.21	0.85	4.50	1.92	1.27
MIC-B	1.32	5.61	1.56	1.47	1.32	1.35	1.67	5.16
B7–H6	1.05	3.70	0.99	1.12	0.79	1.05	0.98	2.81
ULBP1	1.11	1.09	2.65	1.11	18.91	1.24	3.16	2.25
ULBP3	1.12	2.03	1.34	1.87	1.26	1.00	1.77	1.06
ULBP4	1.14	1.58	1.57	1.03	1.07	1.15	1.08	1.35
DR4	14.98	6.36	3.69	7.56	4.22	1.25	3.75	22.37
DR5	8.13	13.16	19.18	10.49	1.89	4.26	9.96	8.35
PCNA	6.35	5.08	9.05	18.59	5.40	3.81	6.46	2.55
HLA-C	1.18	1.53	2.42	1.20	1.05	1.02	1.25	1.18
HLA-G	2.37	4.39	2.12	1.98	2.04	1.40	2.81	2.92
CD200	1.48	1.49	1.63	1.35	1.33	1.41	1.46	1.45
CD155	747.08	281.34	71.80	578.68	454.30	401.82	232.41	84.84
CA125	3.32	3.09	18.51	1.85	5.02	2.42	43.20	18.72
CD58	54.82	244.08	70.79	266.47	162.40	40.51	323.73	185.98
CD112	79.96	16.01	9.12	46.79	69.37	6.55	98.73	63.57

Next, we used pan-cancer training data to build a pan-cancer model, following a similar process to that used for the OC model. We initially split the data 7:3 into training and test sets to validate model performance. After averaging 20 random samplings, the accuracy rates for the training and test sets were 0.759 and 0.750, respectively, and AUC values were 0.884 and 0.781, respectively (Fig. 5A, Table 16). Based on variable importance ranking, the CD200–CD200R interaction ranked first in the model (Fig. 5B). Further validation using an independent external test set (Fig. 5C, Supplementary Fig. 1E, Table 17) and ROC curve analysis revealed inferior model performance, with a test accuracy of 63.52% and AUC of 0.638 (Table 16). The distribution of predictive variables is shown in Supplementary Figure 9F. Matrix scatter plots showed significant discrepancies between the R-L model’s predicted probabilities of cytotoxicity and experimentally measured NK cell cytotoxicity levels (Supplementary Figure 9 G). Overall, the pan-cancer model exhibited markedly inferior performance to the OC model (training and test sets are provided in Supplementary Tables 4 and 5).

Fig. 5.

Performance of the R-L predictive model in pan-cancer studies. (A) ROC curve of the pan-cancer model performance on the test set (30% of total data, mean AUC from 20 random 7:3 train-test splits). The shading represents 95% confidence intervals across iterations. (B) Importance ranking of predictive variables in the pan-cancer R-L model. (C) ROC curve of the pan-cancer predictive model. Red line: performance on the independent external test set; Blue line: performance on training set. (D, E) ROC curves for the breast cancer (BC) (D) and gastric cancer (GC) (E) predictive models on test sets (30% of total data, mean AUC from 20 random 7:3 train-test splits). Blue shading represents 95% confidence intervals across iterations. (F, G) Ranking of ten predictive variables in the random forest models for BC (F) and GC (G). (H, I) ROC curves of BC (H) and GC (I) models. Red line: performance on the independent external test set; Blue line: performance on training set

Table 16.

Prediction performance of the Pan-cancer R-L model in the test set

	Prediction accuracy (training data)	Prediction accuracy (test data)	PPV (training data)	PPV (test data)	Recall (training data)	Recall (test data)	AUC (training data)	AUC (test data)
Test set 1 a	0.759	0.750	0.770	0.730	0.707	0.934	0.884	0.781
Test set 2 b	0.759	0.635	0.770	0.598	0.707	0.930	0.884	0.638

^a Test set 1: a test set partitioned from the original data at a 7:3 ratio

^b Test set 2: an independent external test set

Table 17.

Expression profiles of ligands in tumor cells across multiple cancer types from the test set

	A375	HEPG2	SW620	U266	NIH-716	MKN74	MKN7	HCC1954	SK-BR-3	BC-1	EC-1	LC-1	GC-1
MIC-A	1.05	4.20	1.19	4.15	1.61	0.93	1.82	2.27	1.28	1.06	0.91	1.39	0.98
MIC-B	1.14	1.09	1.48	2.61	1.00	1.02	1.16	1.09	1.17	1.23	1.15	1.01	1.12
B7-H6	1.04	0.94	1.17	2.18	1.28	0.88	2.50	1.03	0.93	1.01	1.10	0.98	1.00
ULBP1	1.02	1.72	8.25	1.64	1.35	7.49	1.72	1.47	3.59	4.99	4.41	3.29	1.46
ULBP3	1.43	1.02	1.77	0.93	1.12	0.97	0.97	1.44	1.13	1.67	1.58	1.04	1.17
ULBP4	1.04	1.02	1.22	0.97	1.31	1.02	1.60	1.24	1.64	1.06	1.06	0.98	1.11
DR4	2.10	6.36	10.64	4.50	1.64	9.74	4.07	4.99	2.30	2.63	7.23	1.37	1.70
DR5	30.02	9.72	18.33	3.49	5.14	2.25	3.34	1.28	3.08	4.53	8.20	5.46	3.64
PCNA	9.74	4.94	2.99	2.04	2.36	4.82	9.14	3.73	3.63	5.42	6.36	6.31	4.58
HLA-C	1.55	2.13	1.22	2.56	1.13	0.94	1.28	3.57	1.23	1.35	1.26	1.23	1.26
HLA-G	1.44	1.43	2.22	1.22	3.83	1.41	2.08	1.63	1.56	1.96	2.09	1.38	1.38
CD200	2.39	1.22	1.34	1.08	1.27	1.73	1.81	1.27	1.18	95.90	2.89	18.52	106.29
CD155	132.98	675.76	319.91	96.52	534.47	72.20	1112.36	365.24	355.77	228.26	347.15	117.46	227.68
CA125	8.76	40.03	115.70	10.09	40.13	14.51	49.94	213.41	61.35	14.90	57.57	1.44	13.42
CD58	204.78	135.64	225.77	343.30	68.15	247.24	125.18	247.24	147.45	231.20	84.25	35.06	69.73
CD112	6.01	34.10	30.42	2.09	4.43	14.18	32.86	27.69	19.89	10.56	7.71	9.08	10.00

Impact of tumor type on predictive models

We hypothesized that the poor predictive performance of the pan-cancer mode was related to tumor type. In a previous study, the NK cell receptors NKG2D and NKP46 and their ligands played different roles in cytotoxicity across different types of tumor cell lines [53]. As tumor ligands exerted variable effects on immune infiltration across different cancers (Fig. 1D), we attempted to categorize all data based on tumor type, establishing distinct cohorts with breast and gastric cancer as examples. Subsequently, 20 random samplings were performed, with 70% of the dataset used for training and 30% for testing. Random forest R-L models were established for breast and gastric cancer (BC and GC models, respectively) using averaged predictions.

After tumor data classification, the predictive performance of the models significantly outperformed that of the pan-cancer model (Table 18). In the test set, BC and GC models achieved accuracy of 88.46% and 86.87%, respectively, with AUC values of 0.850 and 0.873, respectively (Figs. 5D and E). Variable importance ranking revealed differential ligand pair effects across tumor types, with CA125–Siglec.9 and B7–H6–NKp30 most influential in the BC and GC model, respectively (Figs. 5F, G). Scatter plots generated from the matrices showed high consistency and significant correlations between the R-L model’s predicted probabilities of cytotoxicity and experimentally measured NK cell cytotoxicity levels: r = 0.7162, p < 0.0001 for the BC model and r = 0.8645, p < 0.0001 for the GC model (Supplementary Figures 10A and B). Similarly, external validation using independent tumor–NK cell pairs showed good predictive performance, with BC and GC models achieving accuracy of 83.52% and 84.60%, respectively, with AUC values of 0.831 and 0.865, respectively (Figs. 5H and I, Table 18). Thus, cancer-specific classification models outperformed the pan-cancer model, ligand expression patterns differed based on cancer type.

Table 18.

Prediction performance of the BC and GC R-L model in the test set

	Prediction accuracy (training data)	Prediction accuracy (test data)	PPV (training data)	PPV (test data)	Recall (training data)	Recall (test data)	AUC (training data)	AUC (test data)
Test set 1 (BC) a	0.813	0.885	0.766	0.917	0.750	0.846	0.902	0.850
Test set 2 (BC) b	0.813	0.835	0.766	0.800	0.750	0.815	0.902	0.831
Test set 1 (GC) c	0.867	0.869	0.957	0.889	0.842	0.900	0.969	0.873
Test set 2 (GC) d	0.867	0.846	0.957	0.818	0.842	0.800	0.969	0.865

^a Test set 1 (BC): a test set partitioned from the original data at a 7:3 ratio in BC model

^b Test set 2 (BC): an independent external test set in BC model

^c Test set 1 (GC): a test set partitioned from the original data at a 7:3 ratio in GC model

^d Test set 2 (GC): an independent external test set in GC model

To investigate the performance of various tumor datasets in predictive models across different cancer types, we performed model validation using four established R-L models. When testing with melanoma, hepatocellular carcinoma, colorectal cancer, and hematological malignancy cell lines, the pan-cancer model showed optimal predictive performance (Supplementary Figure 10C). The pan-cancer model showed superior performance in predicting novel cancer types, whereas cancer-specific models achieved higher accuracy in predicting their corresponding cancer types.

Validation of ligand expression changes following NK cell-mediated cytotoxicity

Bernareggi et al. showed that tumor cells alter their expression of NK-sensitivity-related surface molecules following NK cell-mediated cytotoxicity in vitro [15]. In our study, we tested two NK cell types (NK-A and NK-O) against the OC cell lines HO8910 and SKOV3, analyzing their ligand expression profiles before and after cytotoxicity assays (Table 19). The predicted probability of cytotoxicity by the OC R-L model matched the observed cytotoxicity data (Fig. 6A). Post-cytotoxicity analysis revealed decreased predicted probability of cytotoxicity, indicating reduced NK sensitivity and acquired resistance in tumor cells after exposure to cytotoxic effects (Fig. 6B, Supplementary Fig. 2A), which is consistent with previous findings that tumor immune escape reduces secondary NK cell killing efficiency [15].

Table 19.

Ligand expression changes in HO8910 and SKOV3 after cytotoxicity

	HO8910-controla	HO8910-NK-Ab	HO8910-NK-Oc	SKOV3-control	SKOV3-NK-A	SKOV3-NK-O
MIC-A	5.75	6.75	6.94	1.18	0.75	0.76
MIC-B	0.87	0.87	0.91	0.87	0.67	0.63
B7–H6	1.02	0.99	1.04	1.05	0.71	0.75
ULBP1	1.94	2.03	1.04	1.84	1.04	1.09
ULBP3	1.54	1.59	1.63	2.24	1.29	1.36
ULBP4	1.21	1.22	1.17	1.37	0.96	1.05
DR4	6.76	7.40	7.77	3.96	3.43	3.63
DR5	6.73	6.55	5.11	4.00	3.94	4.04
PCNA	3.17	3.06	3.39	9.25	9.40	9.75
HLA-C	3.42	3.60	3.57	1.01	0.83	0.81
HLA-G	2.19	2.16	2.21	3.23	1.10	1.52
CD200	2.38	2.73	1.69	3.47	2.12	2.46
CD155	398.60	379.94	323.10	667.75	339.48	329.11
CA125	27.97	37.86	58.79	24.51	8.17	8.04
CD58	120.36	116.63	113.79	105.02	57.63	64.54
CD112	12.80	12.85	12.61	7.96	7.03	6.83

^a HO8910-control: HO8910 cells (NK cell-uncultured)

^b HO8910-NK-A: HO8910 cells co-cultured with NK-A

^c HO8910-NK-O: HO8910 cells co-cultured with NK-O

Fig. 6.

Application of the model to tumor cell dynamics and NK cell therapy for patients with AML. (A) Line graph comparing model predicted probabilities of cytotoxicity and actual killing effects. (B) Treatment-induced changes in tumor cell predicted probability of cytotoxicity by OC model. (C) Experimental results of NK-92 cell therapy against OC (HEYA8, SKOV3, OVCAR8) and BC (MDA-MB-231, MDA-MB-468 and MCF7) cells, where six tumor cell lines were treated with DMSO/ Reversine/ Hesperadin for 48 h. (D) Comparison between model predicted probabilities of cytotoxicity and actual killing effects. (E) NK cell-mediated cytotoxicity against five primary AML samples was assessed using different production batches of FF-NK cells (x-axis). The y-axis represents fractional cytotoxicity (range: 0-1), quantified at an effector-to-target (E:T) ratio of 5:1 in cytotoxicity assays. Each colored segment indicates the specific killing efficiency of individual FF-NK batches against distinct AML cell lines, with total bar height reflecting cumulative cytotoxicity across all AML targets. (F) Radar chart comparing the performance of four models (GC, OC, BC, and pan-cancer model) in patients with AML across five metrics: accuracy, precision, recall, F1 score, and AUC. (G) ROC curve showing the performance of the pan-cancer model in patients with AML

Validation of the R-L model using aneuploid drug-treated cancer cells

Aneuploid drugs can modulate tumor cell sensitivity to immune cells (such as NK cells) by regulating the genomic instability. In this study, we investigated the effects of two aneuploid drugs, Reversine and Hesperadin, on NK cell cytotoxicity using OC (HEYA8, SKOV3, OVCAR8) and BC (MDA-MB-231, MDA-MB-468, MCF7) cell lines. After 48 h of aneuploid drug treatment, NK-92 cell-mediated cytotoxicity was enhanced against BC cells but decreased against OC cells (Fig. 6C). Analysis of ligand expression profiles in HEYA8, MDA-MB-231, and MDA-MB-468 cells revealed significant changes following drug treatment (Supplementary Figure 11B, Table 20). The R-L prediction model outputs were consistent with the experimental results (Fig. 6D), showing decreased sensitivity in HEYA8 cells and increased sensitivity in MDA-MB-231 and MDA-MB-468 cells post-treatment. These findings show that aneuploid drugs influence tumor cell sensitivity by modulating NK cell-related ligand expression and that the R-L model effectively predicts changes in the efficacy of NK cell-based combination therapy.

Table 20.

Aneuploid drug-induced ligand expression changes in cancer cells

	HEYA8-Da	HEYA8-Rb	HEYA8-Hc	MDA-MB-231-D	MDA-MB-231-R	MDA-MB-231-H	MDA-MB-468-D	MDA-MB-468-R	MDA-MB-468-H
MIC-A	2.69	3.39	3.49	1.75	1.91	1.35	1.08	1.10	1.26
MIC-B	1.10	1.05	1.04	1.01	0.97	1.02	1.29	1.91	2.84
B7–H6	1.07	1.08	1.16	1.11	1.16	1.15	0.93	0.48	1.95
ULBP1	8.43	9.42	4.21	0.68	0.67	0.82	18.29	16.66	14.93
ULBP3	3.80	5.82	5.93	1.60	1.96	1.43	1.29	1.80	1.57
ULBP4	1.13	1.13	1.19	0.98	0.97	0.99	0.66	2.79	3.85
DR4	11.70	14.18	13.25	17.18	16.76	14.76	7.92	7.56	6.65
DR5	21.86	25.97	33.98	16.40	16.28	10.90	1.31	2.65	2.61
PCNA	12.23	9.18	9.90	15.38	14.36	12.78	4.01	5.33	4.12
HLA-C	1.11	1.07	1.20	1.06	1.06	1.16	0.51	0.81	1.79
HLA-G	2.09	2.04	2.07	1.49	1.62	1.47	2.12	2.63	1.88
CD200	4.03	6.06	4.46	1.32	1.36	1.25	1.37	1.44	1.36
CD155	719.71	696.45	571.37	856.37	709.13	451.39	1013.43	815.48	672.06
CA125	4.56	4.06	3.21	9.25	10.62	7.76	11.75	9.10	13.35
CD58	195.71	222.09	168.07	359.67	363.34	284.32	232.46	193.54	178.85
CD112	49.06	49.14	55.25	50.07	52.45	47.18	17.44	38.75	28.17

^a HEYA8-D: HEYA8 cells treated with DMSO for 48 hours

^b HEYA8-R: HEYA8 cells treated with Reversine for 48 hours

^c HEYA8-H: HEYA8 cells treated with Hesperadin for 48 hours

Predictive performance of the R-L mode for NK cell therapy efficacy in patients with AML

To further validate the actual predictive performance of the R-L model, we obtained blood samples from five patients with AML (Supplementary Figure 11C), from which we obtained AML ligand expression profiles and cytotoxicity data (Fig. 6E, Supplementary Fig. 2D, Table 21). The data were input into four R-L models (OC, GC, BC, and pan-cancer) for prediction. The pan-cancer model showed the best predictive performance, and high consistency with the in vitro killing results, with accuracy and AUC values of 81.81% and 0.900, respectively (Figs. 6F and G, Table 22).

Table 21.

Expression profiles of ligands in tumor cells from the AML patients

	AML-1	AML-2	AML-3	AML-4	AML-5
MIC-A	1.78	8.70	2.32	2.33	1.00
MIC-B	2.59	6.40	2.61	3.08	1.00
B7–H6	0.79	1.03	1.52	0.82	1.20
ULBP1	2.48	8.85	3.29	3.36	1.00
ULBP3	1.45	8.21	2.52	2.49	1.00
ULBP4	0.83	0.97	1.61	3.38	0.98
DR4	3.04	8.25	4.35	4.08	1.45
DR5	0.83	2.18	2.43	2.51	1.24
PCNA	1.68	4.96	1.43	1.71	1.40
HLA-C	0.87	2.80	11.61	2.42	3.69
HLA-G	50.35	172.68	37.46	3.74	2.41
CD200	55.92	3.30	22.21	27.73	8.68
CD155	52.14	110.83	93.55	39.75	2.17
CA125	5.54	4.28	6.96	3.28	5.79
CD58	75.08	83.48	289.27	31.81	47.62
CD112	2.10	1.31	21.40	30.03	1.71

Table 22.

Comparison of the prediction values with the actual values of NK cell killing in AML samples

Patient ID	NK cell ID	Actual valuea	Predicted valueb	Clinical efficacyc
AML-1	FF-NK042-0614A	18.69%	0	-
	FF-NK010-0614	4.54%	0	-
AML-2	FF-NK042-0628A	67.85%	0	-
	FF-NK010-0614	0.00%	0	-
AML-3	FF-NK042-0628A	10.88%	0	-
	FF-NK010-0614	10.49%	1	-
AML-4	FF-NK042-0628A	12.76%	0	worse
	FF-NK010-0614	59.05%	1	-
AML-5	FF-NK042-0628A	36.62%	0	-
	FF-NK010-0628	17.99%	0	-
	FF-NK010-0718	45.63%	0	-

^a Actual value: measured killing efficacy of NK cells in in vitro cytotoxicity assays (%)

^b Predicted value: Classification output from the R-L model

^c Clinical efficacy: Observed therapeutic response in patients treated with NK cell adoptive transfer

Specifically, patient AML-4 underwent clinical infusion therapy using expanded NK cells (FF-NK042-0628A). Clinical indicators showed suboptimal treatment efficacy (Supplementary Figure 11E, Table 23). Following NK cell infusion, the proportion of AML cells in the blood did not significantly decrease, which was consistent with R-L model predictions.

Table 23.

The response to NK cell infusion treatment in patient 4 with AML

Time point a	Infusion stage b	Test item	NK cells/nucleated cells (%)	NK cell count × 106/L whole blood	Tumor cells/nucleated cells (%)
Baseline	Pre-treatment assessment	Bone marrow cell morphology	Bone marrow primitive cells 15.00%
C1D0	1st Pre-infusion	Whole blood CD series report	62.21%	13.48	3.50%
C1D1	1st Post-infusion (Day 1)	Whole blood CD series report	44.99%	2.13	-
C1D3	1st Post-infusion (Day 3)	Whole blood CD series report	86.99%	5.13	9.30%
C1D5	1st Post-infusion (Day 5)	Whole blood CD series report	57.64%	4.12	18.00%
C1D7	2nd Pre-infusion	Whole blood CD series report	48.82%	10.25	45.00%
C1D8	2nd Post-infusion (Day 1)	Whole blood CD series report	29.33%	4.34	-
C1D10	2nd Post-infusion (Day 3)	Whole blood CD series report	26.54%	11.71	64.00%

^a Time point and ^b Infusion stage: patients received two rounds of NK cell infusions with interval monitoring. Baseline samples were collected before treatment initiation. The first infusion cycle was monitored at C1D0 (1st pre-infusion), followed by post-infusion assessments on Day 1 (C1D1), Day 3 (C1D3), Day 5 (C1D5) and Day 7 (C1D7). The second infusion cycle began at C1D8, with subsequent monitoring on Day 1 (C1D8) and Day 3 (C1D10) post-infusion. The notation “C” represents cycle number, “D” represents day number within each cycle

Applicability of the R-L model to cancer prediction using transcriptomic data

Transcriptome data can be obtained from various samples and allows non-invasive clinical testing. To broaden the application scope of our predictive model, we assessed the feasibility of incorporating tumor transcriptional profiles into the prediction algorithm. We conducted correlation analysis between gene expression profiles from DepMap database and mRNA eukaryotic transcriptome sequencing data across six tumor cell lines (Tables 24 and 25, Supplementary Figure 12A). A significant correlation was observed between these datasets (Fig. 7A, p < 0.0001). Furthermore, the ratio of ligand mRNA transcript abundance and corresponding protein expression levels was relatively stable across different cell types (Fig. 7B).

Table 24.

Expression profiles of ligands in tumor cell lines from the DepMap database

	AGS	BT549	HEYA8	HGC27	K562	MCF7	MDA-MB-231	MDA-MB-468	MKN45	OVCAR3	SKOV3
MIC-A	5.27	6.22	4.00	2.16	4.67	5.61	6.20	5.41	5.01	4.84	5.08
MIC-B	4.55	4.36	2.65	3.04	4.57	3.94	5.55	4.84	2.10	2.00	2.15
B7–H6	1.10	2.30	2.34	3.27	3.57	0.77	0.31	0.54	2.30	0.88	0.71
ULBP1	0.00	0.33	1.74	0.01	3.75	1.66	0.03	1.73	0.83	1.07	0.53
ULBP3	2.90	1.02	2.80	1.54	0.10	0.90	2.11	1.78	2.12	2.59	3.08
DR4	4.66	1.59	3.86	3.25	3.84	3.38	4.56	3.55	3.40	3.48	2.78
DR5	6.56	5.17	6.43	5.00	6.79	4.50	6.21	4.50	5.35	5.99	5.50
PCNA	8.99	8.72	7.80	8.39	7.56	7.45	7.98	8.93	8.33	6.80	7.97
HLA-C	4.06	7.34	6.04	4.59	2.92	5.78	8.57	8.82	6.53	8.12	9.22
HLA-G	0.14	0.00	0.08	0.18	0.00	0.08	0.38	0.66	0.07	0.00	1.35
CD200	0.04	2.01	0.10	0.00	0.08	0.00	0.14	0.10	0.08	3.82	4.00
CD155	7.21	5.63	5.75	4.25	3.94	4.37	6.05	5.99	5.22	5.14	6.15
CA125	0.04	0.12	0.10	0.03	0.00	1.47	0.03	2.52	0.07	7.48	2.09
CD58	4.50	4.74	4.70	4.32	4.97	3.93	5.44	4.78	3.44	5.24	4.37
CD112	8.00	5.43	5.32	3.82	5.72	7.07	6.61	6.47	4.75	5.66	5.81

Table 25.

Eukaryotic transcriptome sequencing of mRNA in tumor cell lines

	AGS	MDA-MB-231	L74	L82	D84	LQQ
MIC-A	15.88	7.10	19.38	9.19	8.27	14.07
MIC-B	22.46	6.99	8.43	4.44	5.59	8.15
B7–H6	0.41	0.67	2.52	3.29	4.79	7.40
ULBP1	0.00	0.00	4.98	3.30	0.00	0.02
ULBP3	2.71	0.65	8.15	1.47	1.78	2.57
DR4	14.99	9.93	5.19	1.23	4.73	7.17
DR5	53.21	40.05	66.34	56.18	49.18	53.75
PCNA	141.14	178.54	137.23	28.19	53.59	83.79
HLA-C	36.52	7.27	238.30	100.48	69.43	221.88
HLA-G	0.00	0.00	0.00	0.01	0.14	0.06
CD200	0.00	0.02	0.06	49.78	82.55	65.37
CD155	27.16	68.76	14.20	46.68	16.15	17.94
CA125	0.01	0.00	0.00	0.03	1.92	5.93
CD58	22.19	5.17	3.04	4.41	7.00	6.04
CD112	39.53	98.21	29.63	61.43	60.73	73.43

Fig. 7.

Predictive performance of the model using transcriptome data. (A) Correlation between mRNA transcriptome sequencing data (y-axis) and DepMap gene expression profiles (x-axis) in AGS (left) and MDA-MB-231(right) tumor cell lines. (B) Composite plot comparing tumor cells ligand expression levels and corresponding mRNA levels. Black dots represent the ratio of protein detection values to transcriptomic sequencing values (Pro/mRNA); red dots indicate the coefficients of variation (CV) of the ratios. Dashed line denotes the reference baseline. (C) Ranked importance of predictive variables in the model. (D) ROC curve of the model prediction results. (E) Predicted probabilities of cytotoxicity (y-axis) across cancer types (x-axis). The dashed horizontal line indicates the high/low susceptibility threshold (probability = 0.5). (F) Cytotoxicity assays of NK cells (n = 6) cytotoxicity against CR cells of BC (BC-1), OC (OC-L74), GC (GC-1), esophageal cancer (EC-1), and lung cancer (LC-1)

A similar trend was observed in NK cells (Supplementary Figure 12B), revealing the feasibility of a predictive model based on mRNA expression data. Consequently, we developed a transcriptome-based R-L prediction model integrating tumor cell transcriptome data, NK cell protein expression profiles, and associated cytotoxicity outcomes. The B7–H6–NKp30 interaction ranked highest in the model (Fig. 7C). This model achieved a prediction accuracy of 75.61% with an AUC of 0.782 in the test set (Fig. 7D, Table 26), indicating potential utility for initial transcriptome data screening while highlighting room for improvement in precision.

Table 26.

Prediction performance of the RNA R-L model in the test set

Prediction accuracy (training data)	Prediction accuracy (test data)	PPV (training data)	PPV (test data)	Recall (training data)	Recall (test data)	AUC (training data)	AUC (test data)
0.876	0.756	0.822	0.816	0.910	0.821	0.838	0.782

Subsequently, we used the model to predict NK cell therapeutic efficacy using tumor patient-derived transcriptome data (GEPIA database), incorporating averaged expression profiles from NK cell clusters with homogeneous phenotypic profiles (Supplementary Figure 12C, Tables 27 and 28). The prediction model identified ovarian, skin, lung, and hematological cancer as having higher response probabilities to NK-cell infusion therapy. These cancer types showed high predicted probabilities ( > 50%) compared to other malignancies (Fig. 7E). This is consistent with previous studies [41, 54–62]. Furthermore, we conducted experimental validation using CR ovarian, breast, gastric, esophageal cancer, and lung cancer. In vitro NK-cell-mediated killing assays showed that model sensitivity decreased as follows: lung > ovarian > breast > gastric > esophageal cancer, which was largely consistent with the R-L model predictions (Fig. 7F). Thus, the R-L model may be applicable to gene expression data from clinical patients.

Table 27.

Expression profiles of ligands in tumor cells from patients with various cancers in the GEPIA database

	OV	LAML	COAD	BRCA	SKCM	STAD	ESCA	LIHC	LUAD	READ
MIC-A	8.31	8.17	8.17	13.80	17.85	9.69	10.33	7.75	14.13	11.96
MIC-B	3.44	13.19	13.19	4.12	4.12	4.42	4.98	1.46	4.92	2.20
B7–H6	0.23	1.72	2.66	0.35	0.36	2.43	2.36	0.05	0.21	3.09
ULBP1	0.08	0.04	0.04	0.24	0.04	0.16	0.81	0.03	0.38	0.27
ULBP3	1.08	0.44	0.44	0.38	0.34	2.29	4.08	0.07	1.34	1.39
ULBP4	0.14	0.22	0.22	0.60	0.07	0.27	1.01	0.05	0.48	0.27
DR4	2.35	16.05	16.05	5.46	1.36	7.63	9.54	2.14	7.80	7.58
DR5	9.67	33.86	33.86	14.29	24.06	25.50	32.57	12.96	31.05	27.38
PCNA	92.87	118.64	156.97	117.27	122.24	121.89	136.05	48.09	73.80	171.94
HLA-C	576.91	404.67	747.33	668.23	776.05	729.01	709.91	499.66	908.16	658.50
HLA-G	0.68	0.24	0.60	0.42	0.96	0.87	0.71	0.41	1.47	0.85
CD200	16.90	15.86	3.91	7.40	13.98	5.11	3.39	1.30	5.24	4.02
CD155	9.74	9.64	27.79	9.12	16.89	25.00	31.29	20.04	18.63	28.27
CA125	85.70	0.01	0.02	0.57	0.02	0.10	0.47	0.01	5.46	0.02
CD58	9.07	32.57	18.42	12.50	21.64	18.79	22.65	6.01	16.06	18.42
CD112	123.34	24.22	81.90	162.59	41.40	115.31	94.05	104.80	91.65	77.31

Table 28.

Model prediction results using tumor patient-derived transcriptome data (GEPIA database)

Cancer type	Predicted result
	Predicted probabilities	Predicted classes	Sensitivity level
OV	0.640	1	High Susceptibility
LAML	0.779	1	High Susceptibility
COAD	0.235	0	Low Susceptibility
BRCA	0.407	0	Low Susceptibility
SKCM	0.572	1	High Susceptibility
STAD	0.340	0	Low Susceptibility
ESCA	0.205	0	Low Susceptibility
LIHC	0.186	0	Low Susceptibility
LUAD	0.544	1	High Susceptibility
READ	0.285	0	Low Susceptibility

Predictive potential of the R-L model based on GEO transcriptome data

Next, we evaluated the R-L model using tumor and NK cell transcriptome datasets from GEO to further assess its predictive potential at the transcriptome level. Specifically, we evaluated performance across different tumor cohorts using expanded NK cells (Supplementary Figure 12C) as input. In the GSE120736 bladder cancer cohort, recurrent tumors exhibited significantly higher predicted sensitivity to NK cell therapy than primary tumors (p < 0.0001; Fig. 8A), indicating potential enhanced efficacy of our NK cells against recurrent malignancies. In the GSE32526 BC cohort, although only S2N spheroid cells formed xenograft tumors, mRNA profiles alone could not predict tumorigenicity [63], our R-L model revealed significant differences in NK therapy sensitivity between spheroid and monolayer cells (p = 0.0005). S2N spheroid cells exhibited higher predicted sensitivity than S2 spheroid cells, with S2N spheroid cells showing the highest values (Fig. 8B). In the GSE68138 prostate cancer cohort, gene expression profiles from tumor tissues revealed distinct differences in predicted outcomes based on smoking status, revealing nicotine-induced progression (p = 0.0053, Fig. 8C). This finding is consistent with research showing that nicotine modulates cellular metabolism and subsequent activation of NF-κB [64]. In the GSE27556 lung cancer cohort, the predicted values for lung cancer samples were significantly higher than those for normal samples (p = 0.0069; Fig. 8D), indicating better sensitivity to NK immunotherapy.

Fig. 8.

Application and validation of the predictive model using GEO datasets. (A) Violin plot showing the differences of cytotoxicity predicted probabilities between recurrent and primary tumors in the GSE120736 bladder cancer cohort, ****p < 0.0001. (B) Bar graph comparing predicted probability of cytotoxicity between tumorigenic (S2N) and non-tumorigenic (S2) cell lines derived from primary BC in the GSE32526 cohort, *p < 0.05, ***p < 0.001, ****p < 0.0001. (C) Violin plot showing model-predicted probability of cytotoxicity in the prostate cancer cohort (GSE68138) based on smoking status. (D) Box plot comparing model predicted probabilities of cytotoxicity between lung cancer and normal samples in the GSE27556 cohort; **p < 0.01. (E) Bar plot showing the impact of hypoxia on the predicted probability of NK cells from different donors in the GSE116660 cohort. (F) Bar chart showing prediction differences for uterine decidual NK cells between lean and obese pregnant women in the GSE75091 cohort. (G) Bar graph showing increased predicted probability of cytotoxicity with prolonged interferon-α treatment in the GSE15743 cohort. (H) Differences in the predicted probabilities of cytotoxicity against Jurkat, K562, and Huh7 cells before and after IFN-α treatment

We further evaluated NK cell cohort performance using the R-L predictive mode. In the GSE116660 cohort, where NK cell-only data from NK cell cohorts was validated in the model using one-sided analysis, predicted NK cell sensitivity for different donors significantly decreased following hypoxia treatment, indicating a potential reduction in tumor-killing capacity (p < 0.0001; Fig. 8E). This aligns with previous reports that hypoxia modulates NK cell receptor expression through various pathways, impairing cytotoxicity [65]. In the GSE75091 cohort (using NK cell-only validation data), the predicted values for uterine decidual NK cells did not significantly differ among 13 lean and 11 obese pregnant women (Fig. 8F). These results are consistent with those obtained by Perdu et al., revealing that obesity does not alter certain canonical decidual NK cell markers, including cytotoxic activity [66]. Finally, in the GSE15743 cohort (using NK-tumor paired validation data from GEO), the R-L model predicted increased NK cell activity with prolonged interferon-α treatment (Fig. 8G). The predicted cytotoxicity against Jurkat, K562, and Huh7 cells was significantly higher in IFN-α-treated NK cells than in untreated controls (Fig. 8H). These results align with previous findings that IFN-α enhances human NK cell cytotoxicity in a dose-dependent manner [67]. Overall, the R-L predictive model showed strong predictive potential at the transcriptome level.

Discussion

NK cells are key innate immune components that play a significant role in malignancy development. Allogeneic NK cell adoptive transfer effectively increases the number and function of NK cells during cancer treatment [68]. Currently, NK cell therapy optimization primarily focuses on improving the source of therapeutic NK cells used for adoptive transfer and enhancing cytotoxicity and persistence in vivo [69–73]. However, in-depth research on the significant differences in efficacy among different patients receiving NK cell infusion treatment is lacking. The expression of surface ligands on tumor cells varies among patients, and receptor expression differs among NK cells expanded from different sources. These expression profile differences contribute to the varying sensitivity of tumor cells to NK cell therapy, leading to selective specificity in the responses of different patients. To examine how R-L expression determines NK cell therapy efficacy, we developed a model based on NK and OC cell expression profiles to predict therapeutic efficacy. The random forest model effectively predicted NK cell cytotoxicity against OC using the expression levels of 11 R-L pairs, with variable contributions from different R-L pairs. Validation using an independent OC test set showed satisfactory predictive performance.

The R-L pairs were preliminarily selected from pairs previously associated with NK cell-mediated cytotoxicity [32, 42, 50, 74–83]. For instance, published studies have confirmed that blocking the CD200/CD200R pathway enhances NK cell activation and cytotoxicity [82, 83]. Qiu et al. showed that knockdown of B7–H6 impaired the sensitivity of tumor cells to NK-mediated lysis [81]. In addition, Chauvin et al. reported that TIGIT blockade increased NK cell-mediated cytotoxicity against melanoma cells in vitro [80]. Furthermore, we conducted supplementary screening using database analyses, including differential expression analysis between tumor and normal cells, survival-related analysis, and immune infiltration correlation analysis, to confirm these 11 R-L pairs as predictive variables. Although the current study did not include blocking experiments as experimental validation, our predictive model effectively synthesizes established mechanistic data from the literature and database to generate robust predictions. Additionally, several limitations should be acknowledged in the selection process of predictive variables. First, owing to limited resources and experimental conditions, some receptor-ligand pairs known to influence NK cell cytotoxicity were not included in the model, such as NKp46 [40] and its tumor cell ligands, as well as KIR3DL and its ligands (HLA-A/B). Second, the screening criteria from databases were not directly associated with NK cell-mediated tumor cell killing, serving only as supplementary conditions. Nevertheless, the variable importance ranking derived from our model provides testable hypotheses for future experimental validation.

NK cell recognition and cytotoxicity are governed by a complex network of receptor-ligand interactions, rather than by a single receptor-ligand pair. Although CD200R was the only receptor showing significant differential expression (Fig. 3B), it is well-established that NK cell function is determined by the integration of signals from multiple activating and inhibitory receptors [5, 6]. Therefore, unlike previous studies, we conducted paired analysis of the 11 pairs to better understand these complex interactions [84–86]. The initial predictive model based on cell scores from tumor and NK cells exhibited poor predictive performance. Notably, while univariate analysis revealed significant correlations of NK cell cytotoxicity with tumor score and NK score separately, the integrated linear model showed no significant linear association between these scores and NK cytotoxicity. This indicated potential nonlinear interactions determining cytotoxicity. Accordingly, we constructed a novel, improved model by combining each R-L pair, enabling a more precise assessment of the impact of R-L pairs on cytotoxicity and better reflecting tumor–NK cell interactions. In addition, the predicted probability of cytotoxicity was significantly correlated with actual cytotoxicity, indicating that the predictions effectively quantify cytotoxicity beyond categorical classification. These findings highlight our model as a reliable potential tool for the quantitative prediction of NK cell-mediated cytotoxicity.

By developing a pan-cancer model based on OC, we broadened the study scope to include other cancers. Unexpectedly, ligand expression patterns varied among cancer types. This was shown using the independent BC and GC models, which showed superior predictive performance for their respective cancers. The inferior performance of the pan-cancer model compared to cancer-specific models can be attributed to several biological and immunological factors. The molecular interactions between NK cells and cancer cells, particularly the ligand-receptor combinations and their relative importance, reveal cancer type-specific patterns. As reported previously, NK cell receptors (such as NKG2D and NKp46) exhibit context-dependent functional roles across cancer types [53]. This biological heterogeneity results in different immunogenic profiles and treatment responses among cancer types. In our study, we observed distinct patterns of predictor importance across different cancer type-specific models. In the OC model, we identified the CD200–CD200R axis as one of the most important variables, highlighting the significance of this pathway in regulating NK cell anti-tumor responses in OC. In contrast, this pathway showed relatively lower importance in BC and GC models. Therefore, modeling NK cell-mediated anti-tumor immunity requires cancer-specific approaches to accurately capture the distinct molecular mechanisms of each cancer type. However, when predicting outcomes in novel cancer datasets, the pan-cancer model outperformed specific models. Therefore, although heterogeneity exists between cancer types, some molecular characteristics are shared. Predicting the NK cell therapeutic efficacy likely encompasses both cancer-specific and pan-cancer features. This indicates that the pan-cancer approach provides advantages in terms of generalizability across different cancer types, particularly for rare cancers or cases where cancer-specific training data is limited. However, when sufficient cancer-specific data is available, specialized models may provide more precise predictions for their targeted cancer types. These findings improve our understanding of tumor classification and provide novel perspectives for immunotherapy approaches in different cancers.

In this study, we primarily used flow cytometry, which clearly shows protein receptor and ligand expression. Most current predictive studies use RNA-seq, which is slower and may be less representative of cellular phenotypes, but provides deeper genetic insights and better addresses current clinical needs [84, 86, 87]. Therefore, to enhance the clinical applicability of our research, we evaluated RNA-seq data and determined that the transcriptome-based model showed promising predictive potential after refinement. Based on validation using tumor and NK cell transcriptomic datasets from GEO, the predictive model effectively differentiated between distinct cell types or treatment conditions based on gene expression profiles, whether applied to human tumor tissues, tumor cell lines, or in vitro-treated/cultured NK cells. This demonstrates that our predictive model also holds significant potential for application in transcriptomic data analysis. In addition, we approximated the possible responses of several tumor types to NK cell therapy using transcriptome data from GEPIA to validate model performance. Previous clinical studies on hematological [54], ovarian [41, 55], skin [60], and lung cancers [88] have shown that the adoptive transfer of NK cells can significantly alleviate symptoms and improve patient survival. However, more clinical data are required for other tumors, such as liver [56], gastric [62], breast [58], esophageal, and colorectal cancers [59], to further prove the effectiveness of NK cell therapy. This is consistent with our model predictions, which provide novel clinical insights regarding the most suitable tumor types to target for NK cell immunotherapy.

The development of personalized treatment plans specific to patient features can improve outcomes and avoid ineffective approaches [89, 90]. The challenge lies in accurately interpreting large amounts of cellular data and translating them into clinical practice. Our study directly connects expression profile data and treatment efficacy. By leveraging multiple biomedical detection technologies, we can profile the expression patterns of ligands at the protein or mRNA levels in the tumor cells of patients. Furthermore, expanded NK cells can be screened by analyzing receptor expression signatures and cryopreserved for future use. By integrating these ligand profiles with receptor expression signatures of expanded NK cells, this predictive approach enables identification of optimal NK cells for each patient, directly informing clinical decision-making in NK cell therapy. This method allows for precise patient stratification and treatment customization, creating a framework that maximizes therapeutic efficacy while minimizing unnecessary treatment in potential non-responders. Besides clinical trials of NK cell infusion showing potential, more and more treatments combining NK cell therapy with radiotherapy, chemotherapy, or targeted therapies to enhance efficacy are emerging [91, 92]. However, the R-L model shows promise for such combination treatments. One primary cause of medication resistance and decreased treatment efficacy is changes in tumor cell phenotype following different treatments [93, 94], necessitating continuous monitoring and adjustment. By adjusting therapeutic schemes based on real-time observations of cellular phenotypic changes, the R-L predictive model can decrease the impact of treatment-related cell resistance, showing its potential as an effective therapeutic strategy when combined with cancer treatment. Looking forward, we plan to develop an open-access web platform that will allow clinicians to input patient biomarker data and NK cell product characteristics to generate automated predictions of therapeutic outcomes.

Furthermore, the potential impact of different expansion protocols on NK cell properties warrants careful consideration. These IL-2 concentrations were empirically determined through prior optimization studies conducted by our laboratory to ensure effective expansion for each method. Our results show that the expansion protocols influence NK cell cytotoxicity, which correlates with observed differences in surface receptor expression profiles among these populations. We intentionally preserved these methodologically induced differences in our study, as they provide the different spectrum of receptor expression profiles required for our R-L predictive model, thereby enhancing its robustness and generalizability. The consistent experimental framework ensured valid cross-comparisons despite methodological variations. Furthermore, this approach aligns with real-world clinical scenarios where multiple expansion protocols may be used, making our findings more relevant to translational applications. Future investigations into how different expansion methods mechanistically regulate NK cell receptor expression and functional capacity would be valuable for optimizing NK cell-based immunotherapies.

Our study had some limitations. First, the model cannot accurately predict outcomes in all scenarios. For example, our model was developed using machine learning techniques, and although we repeatedly validated it with new experimental data, the accumulation of actual data is relatively slow, and the overall size of the test set remains small. Second, both in vivo studies and expanded clinical validation are required to strengthen our findings. Although our current analysis includes clinical AML data, the sample size remains limited, and the model requires validation through prospective clinical trials. Notably, the lack of animal model testing prevents assessment of key physiological factors like NK cell trafficking and tumor microenvironmental modulation. In addition, high-quality, large-scale clinical datasets for NK cell therapy remain scarce in existing databases. We are actively pursuing both murine xenograft experiments and collaborative clinical trials to address these limitations in future work. Third, genetic mutations can affect sensitivity to NK cells [95]; thus, an in-depth investigation into genetic mutations and other influencing factors is warranted. Finally, the complexity of the tumor immune microenvironment hinders the accurate prediction of treatment efficacy, and the state and function of immune cells change over time. In addition, in clinical settings, patient age and tumor stage can affect treatment outcomes, increasing prediction difficulty and reducing accuracy [79, 87]. Thus, the effect of extracellular factors is a topic for further investigation in future composite models.

Conclusions

In the present study, we established a predictive model of NK cell cytotoxicity against ovarian cancer and expanded the study scope to include other cancers using PBMC-derived NK cells and conditionally reprogrammed primary cells, which more closely resemble clinical treatment scenarios. This study provides a foundation for the precise treatment of tumors with NK immune cells and has clinical applications for immunotherapy against different types of cancer.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Abbreviations

AML

Acute myeloid leukemia

AUC

Area under the ROC curve

BC

Breast cancer

BRCA

Breast invasive carcinoma

COAD

Colon adenocarcinoma

CP

Combined predictor

CR

Conditionally reprogrammed

DepMap

Dependency Map

ESCA

Esophageal carcinoma

FAC

Factor coordinates

FACS

fluorescence-activated cell sorting

FBS

Fetal bovine serum

FF-NK

Feeder-free expanded natural killer cell

Feeder-NK

Feeder cell expanded natural killer cell

GC

Gastric cancer

GEO

Gene Expression Omnibus

GEPIA

Gene Expression Profiling Interactive Analysis

HPA

Human Protein Atlas

LIHC

Liver hepatocellular carcinoma

MFI

Mean fluorescence intensity

MB-NK

Protein-conjugated magnetic bead-expanded natural killer cell

NK cell

Natural killer cell

NPV

Negative predictive value

OC

Ovarian cancer

OS

Overall survival

OV

Ovarian serous cystadenocarcinoma

PBS

Phosphate-buffered saline

PBMCs

Peripheral blood mononuclear cells

PCNA

Proliferating cell nuclear antigen

PCA

Principal component analysis

P-P plot

Probability-probability plot

PPV

Positive predictive value

rhIL-2

Recombinant human interleukin-2

R-L

Receptor-ligand

RTCA

Real-time cell analysis

RMSE

Root mean square error

ROC

Receiver operating characteristic

SKCM

Skin cutaneous melanoma

STAD

Stomach adenocarcinoma

SVM

Support vector machine

TIMER2.0

Tumor Immune Estimation Resource 2.0

TISIDB

Tumor Immune System Interactions Database

XGBoost

Extreme gradient boosting

λ

Eigenvalues (variance explained per component)

Author contributions

X. W. Hua, F. Fang and M. Jie conceived the project and designed the study. M. Jie, Y. J. Jing, L. Y. Tong, and D. H. Bo performed the experiments. M. Jie analyzed the data and wrote the manuscript, and F. Fang, X. W. Hua, and L.Y. Yang revised the manuscript. All authors had full access to all of the data in the study and had final responsibility for the decision to submit for publication.

Funding

This work was supported by the Strategic Priority Research Program of the Chinese Academy of Sciences (XDB29030200) and the Key Research and Development Projects of the Ministry of Science and Technology (2023YFC2506400).

Data availability

All data generated or analyzed during this study are included in this article and its supplementary materials. The publicly available datasets utilized in this study are accessible through the following repositories: Oncolnc: http://www.oncolnc.org/, HPA: https://www.proteinatlas.org/, GEPIA2021: http://gepia2021.cancer-pku.cn/, GEO: https://www.ncbi.nlm.nih.gov/geo/, GEO2R: https://www.ncbi.nlm.nih.gov/geo/geo2r/, GEPIA: http://gepia.cancer-pku.cn/, DepMap: https://depmap.org/portal/, TISIDB: http://cis.hku.hk/TISIDB/, and TIMER2.0: http://timer.cistrome.org/. Clinical blood sample data are not publicly available to protect patient privacy but may be requested from the corresponding author under reasonable conditions.

Declarations

Ethics approval and consent to participate

PBMC-NK cells: This study was conducted in accordance with the Declaration of Helsinki. PBMCs were obtained through an institutional collaboration between our laboratory and Taicang First People’s Hospital, with approval from the hospital’s Ethics Committee (Approval No. 2022-SR-020; approval date: November 10, 2022). All participants provided written informed consent prior to enrollment. AML cells: The use of peripheral blood samples from patients with AML was conducted in accordance with the Declaration of Helsinki and approved by the Ethics Committee of the Second Hospital of Anhui Medical University (Approval No. YX2023–046, valid from March 20, 2023 to March 19, 2024). All participants provided written informed consent prior to enrollment.

Consent of publication

This study has not been published before, and all authors approved this publication.

Competing interests

The authors declare no competing interests.