RPSLearner: A novel approach based on random projection and deep stacking learning for categorizing NSCLC

RPSLearner: A novel approach based on random projection and deep stacking learning for categorizing NSCLC | bioRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-M677548'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search New Results RPSLearner: A novel approach based on random projection and deep stacking learning for categorizing NSCLC View ORCID Profile Xinchao Wu , Jieqiong Wang , View ORCID Profile Shibiao Wan doi: https://doi.org/10.1101/2025.05.01.651699 Xinchao Wu 1 Department of Genetics, Cell Biology and Anatomy, University of Nebraska Medical Center , Omaha, NE Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Xinchao Wu Jieqiong Wang 2 Department of Neurological Sciences, University of Nebraska Medical Center , Omaha, NE Find this author on Google Scholar Find this author on PubMed Search for this author on this site Shibiao Wan 1 Department of Genetics, Cell Biology and Anatomy, University of Nebraska Medical Center , Omaha, NE Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Shibiao Wan For correspondence: swan{at}unmc.edu Abstract Full Text Info/History Metrics Preview PDF Abstract Background Lung cancer is the leading cause of cancer death, and non-small cell lung cancer (NSCLC) comprises the largest subtype with most cases. Lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC) are two NSCLC subtypes that pose challenges for accurate diagnosis using conventional methods. Existing methods are histological examination and imaging which lacks definitive histologic features and requires intense time. Methods To address these concerns, we propose RPSLearner, which combines Random Projection (RP) for dimensionality reduction and stacking ensemble learning to accurately predict lung cancer subtypes. Specifically, multiple independent RP matrices were first generated to project the high-dimensional RNA-seq data into lower-dimensional space, whose features were subsequently concatenated. After that, we fed the fused features into a stack of diverse base classifiers and integrated the predictions from base models via a deep linear layer network. Results Benchmarking tests on 1,333 NSCLC patients demonstrated that RPSLearner outperformed state-of-the-art approaches for lung cancer subtype classification. Specifically, RPSLearner efficiently preserved sample-to-sample distances even after significant dimension reduction, and the meta-model in RPSLearner yielded consistently higher accuracy, F1 and AUC scores than individual base models and state-of-the-art approaches for lung cancer subtyping. Besides, the feature fusion method applied in RPSLearner shown better performance than conventional scores ensemble methods. Conclusion We developed a novel stacking learning method called RPSLearner which combines RP and stacking learning, enabling efficient and accurate identification of NSCLC subtypes. RPSLearner is a promising lung cancer subtyping model for downstream lung cancer clinical diagnosis and personalized treatment, and the framework holds the potentiality to be extended to subtyping of other types of cancer. Introduction Lung cancer remains one of the leading causes of cancer-related mortality worldwide, accounting for a substantial proportion of cancer diagnoses and deaths each year 1 , 2 . Non-small cell lung cancer (NSCLC) constitutes approximately 85% of all lung cancer cases, with subtypes such as lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC), exhibiting distinct molecular and clinical characteristics 3 . Accurate identification of NSCLC subtypes is critical for prognosis determination, therapeutic decision-making, and the development of personalized treatment strategies 4 . Traditional diagnostic methods, including histopathological examination and imaging techniques 5 , while valuable, often lack the precision required for nuanced subtype differentiation, particularly in early-stage disease 6 . Besides, histological examination increases the risk of lung cancer metastasis due to physical damage 7 . In recent years, advancements in high-throughput genomic and transcriptomic technologies have revolutionized our understanding of the molecular mechanisms of lung cancer, enabling the identification of specific biomarkers and genetic signatures associated with different NSCLC subtypes 8 . Among these techniques, transcriptomic profiling provides a dynamic and quantitative snapshot of gene expression, thereby serving as a more direct indicator of the cellular functional state 9 . However, the analysis of such high-dimensional omics data presents significant computational and methodological challenges. The complexity and volume of transcriptomic profiles require robust feature selection and dimensionality reduction techniques to extract meaning patterns while mitigating the risk of overfitting and enhancing model generalizability 10 . Machine learning (ML) approaches have emerged as powerful tools for integrating and analyzing large-scale biological data, offering the potential to improve the accuracy and efficiency of lung cancer subtype prediction. For instance, Huang et al. 11 utilized logistic regression (LR) with LASSO shrinkage for predicting NSCLC subtypes based on RNA-seq data. Their model demonstrated superior performance compared to LR using principal components and K-nearest neighbor (KNN) algorithms. Similarly, Yuan et al. 12 applied a method called Monte Carlo feature selection (MCFS) in combination with support vector machine (SVM) for subtype stratification based on microarray gene expression data, showing better results than the random forest (RF) classifier. Although these studies illustrated the effectiveness of machine learning in NSCLC subtype classification, significant challenges remain. In particular the high dimensionality of omics data can lead to overfitting and reduced generalizability, and inter-sample variability further complicates the creation of robust predictive models 13 . Recently we have witnessed the significant development of multiple deep learning approaches 14 , 15 tailored to cancer subtype diagnosis combining histological images and genomics data, which have demonstrated promising results for cancer subtyping. However, these types of methods require the availability of transcriptomics data paired with histological staining imaging data, which are very costly or laborious to obtain. In this case, we focus on developing a novel approach solely based on transcriptomics data for cost-effective and accurate NSCLC subtype classification. To address these challenges, we developed RPSLearner combining Random Projection (RP) for dimensionality reduction and stacking ensemble learning to accurately predict lung cancer subtypes. In detail, by leveraging multiple independent RP matrices and fusing these reduced features, RPSLearner can capture essential sample-to-sample distances 16 , 17 . Subsequently, RPSLearner uses stacking learning framework to integrate diverse heterogeneous base classifiers with a deep linear network as a meta-model 18 , 19 . The stacking framework not only consolidates complementary predictive insights but also yields improved performance metrics over conventional methods. Ultimately, in this study, we demonstrated that RPSLearner significantly enhance NSCLC subtype classification and offered a versatile framework that may be adapted for other cancer subtyping applications. Results The design of RPSLearner To accurately identify subtypes of NSCLC, we presented RPSLearner, a novel computational framework that integrates RP with stacking ensemble learning. The whole process of the method was shown in Fig. 1a . RPSLearner leverages RP’s capabilities of dimensionality reduction to deal with the high-dimensional nature of transcriptomic data 20 . Specifically, it applied multiple random projections to patients’ transcriptomic profiles to extract meaningful hidden information, significantly decreasing the feature space while preserving the intrinsic distances between samples. This approach facilitated the handling of large-scale omics datasets, ensuring computational efficiency without compromising the integrity of the underlying biological signals. Download figure Open in new tab Figure 1. The overview of lung cancer TCGA datasets and the RPSLearner algorithm. ( a ) The computational pipeline of the RPSLearner. ( b ) Low-dimension visualization of original data and RP-transformed data through diverse techniques. ( c ) The correlative analysis across diverse dimensionality reduction methods. ( d ) The performance of score average strategy against feature fusion. Ori.: Original, PCA: Principle Component Analysis, tSNE: t-distributed Stochastic Neighbor Embedding, UMAP: Uniform Manifold Approximation and Projection. RP: Random Projection. Statistical significance level denoted as: ns (not significant) or P ≥ 0.05, * for P < 0.05, ** for P < 0.01, *** for P < 0.001. RPSLearner efficiently preserved sample-to-sample distances To demonstrate the effectiveness of RP, we applied PCA 21 , t-SNE 22 , and UMAP 23 to visualize the RP-transformed NSCLC RNA-seq data and compared them against original data ( Fig. 1b ). Across diverse techniques, the RP-transformation, fusing multiple RP dimension reduction results, exhibited clear subtype identities with distinct cluster groups, suggesting that effectively extract meaningful features in lung cancer subtype prediction and simplify the prediction task. To further validate the performance of RP for dimensionality reduction, we next evaluated RP against PCA, t-SNE, and UMAP for NSCLC RNA-seq data. Fig. 2b compared Pearson correlation coefficients and distance-fitting curve between original and projected feature vectors. Download figure Open in new tab Figure 2. The performance of RPSLearner in lung cancer subtype prediction. ( a ) The comparison of performance of diverse individual base model and stacked model, RPSLearner. ( b ) The comparison of performance of various combination of base models in stacking learning, top k selection was based on results of previous individual base model comparison, ALL represented using all ten models. KNN-D: KNN in distance metrics, KNN-U: KNN in uniform metrics, RF-G: RF in Gini metrics, RF-E: RF in entropy metrics, ET-G: ExtraTree in Gini metrics, ET-E ExtraTree in entropy metrics, LGBM: LightGBM, CB: CatBoost, XGB: XGBoost, NN: neural network. Statistical significance level denoted as: ns (not significant) or P ≥ 0.05, * for P < 0.05, ** for P < 0.01, *** for P < 0.001. RP consistently maintained higher correlations across various dimensions up to 0.94 ( Supplementary Fig. 2 , more effectively preserving sample-to-sample relationships. UMAP showed the second-best correlation (0.84) but lagged RP. PCA and t-SNE were less effective in preserving pairwise distances for this dataset. These findings underscore RP’s suitability for retaining meaningful data patterns in a reduced-dimension space. To demonstrate the effectiveness of the stacking learning method in RPSLearner, we compared two ways of leveraging multiply RPs, feature fusion and score ensemble. Conventionally, methods applied score ensemble, which averaging the prediction probabilities from independent runs as the final score. Feature fusion concatenates RP-transformed features for each randomly generated feature and feed them into stacking learning pipeline. We compared the performance between modified RanBALL 16 , which originally used for leukemia subtypes prediction applied ensemble strategy, and RPSLearner, using feature fusion. Through 20 times 5-fold CV, we demonstrated feature fusion yields higher accuracy, F1 score, MCC and AUC than score ensemble ( Fig. 1d ), suggesting that integrating projected features directly exploits complementary representations more efficiently than simple probability averaging. Stacking meta learning outperforms individual base learning models We next tested RPSLearner using different projected dimension sizes (from 100 to 1,400) to find optimal performance ( Supplement Fig 3a ). With 10 times 5-fold cross validation (CV), it demonstrated that the accuracy and other metrics, improved from 100 to 400 dimensions and then achieved stabilized. Subsequently, we explored to use multiple independent RP matrices ( Supplement Fig. 3b ). With the same setting using 10 times 5-fold CV, it showed that accuracy and other metrics steadily rose as the number of RPs used up to around 25, reaching 0.961 in accuracy, 0.96 in F1, 0.92 in MCC, and 0.986 in AUC on average. Beyond 25 RPs, the performance gains were marginal, suggesting that moderate RP ensembles sufficiently capture complex signals. To evaluate the performance of stacking learning, we compared the performance among each individual base model and NN-based meta-model. We assessed them using 10 times 5-fold CV with 20 individual 400-dimensional RP matrices. Fig. 2a showed meta-model achieved the best performance across all metrics. The NN base classifier performed the second best, but stacking meta-model delivered a consistent and significant boost in accuracy, F1, and MCC. For AUC, individual strong learners attained high scores around 0.985, leaving less room for stacking gains to show significant improvement. Nevertheless, the stacking approach never degraded performance. To deeper examine the effectiveness of stacking learning, we compared the performance across various composite of base models in used. We assessed with the same RP setting as above, and Fig 2b exhibited that the performance of using all base models obtained smallest variance while was no less than the best performance of base model combination using top 3. In other words, it enhanced the robustness of the model prediction. RPSLearner outperformed state-of-the-art approaches for lung cancer subtype identification To demonstrate that RPSLearner could outperform state-of-the-art methods, we benchmarked RPSLearner against previously published methods for NSCLC subtype prediction ( Fig. 3a ). Approaches includes MCFS + support vector machine (SVM), MCFS + RF, Lasso + SVM, Lasso + RF, ANOVA + SVM 24 , ANOVA + RF and a Lasso-based developed by Huang et al. 11 obtained accuracy ranging from around 0.93 to 0.95. By contrast, RPSLearner achieved around 0.96 accuracy, which is significantly higher than SOTA method. The consistent and significant improvement also shown in other metrics. Besides, we further identify differential expression genes (DEGs) based on RPSLearner prediction ( Fig. 3b, d ) There are many well-established biomarkers of LUAD and LUSC included in the identified DEGs. For example, KRT5, KRT6A, KRT14, and DSG3 in LUSC 25 , and NKX2.1 26 , SFTA2 27 in LUAD have been evidently proved directly correlated to the lung cancer diagnosis and progression. Additionally, we compared the DEGs identified with prediction of other methods and ground-truth ( Fig. 3c ). Through RPSLearner prediction, there is a larger overlap between RPSLearner with ground-truth, and RPSLearner could find more DEGs than ground-truth. These findings highlight the accuracy of the prediction, as well as potential novel molecular signatures identified through RPSLearner prediction. Download figure Open in new tab Figure 3. RPSLearner outperforms state-of-the-art methods with interpretable results. (a ) The benchmark of methods for lung cancer subtype prediction. ( b ) The volcano plot illustrates the significance of the DE genes of each subtype. ( c ) Comparison among differential expression genes (DEGs) identified based on RPSLearner and ground-truth. ( d ) Expression of the DEGs extracted from RPSLearner prediction. Statistical significance level denoted as: ns (not significant) or P ≥ 0.05, * for P < 0.05, ** for P < 0.01, *** for P < 0.001. MCFS: Monte Carlo Feature Selection. Discussion In this study, we introduced RPSLearner, a novel approach that combines RP for dimensionality reduction and stacking learning framework to aggregate predictions from diverse base models. RPSLearner utilized RP to reduce the dimensionality of transcriptomics data. Unlike other dimensionality reduction approaches that distorted intrinsic sample-to-sample relationships, RP applied in RPSLearner was able to capture meaningful and accurate latent features from the transcriptomics data, efficiently reduce dimensionality while preserve intrinsic distances among samples. This was crucial for maintaining the structural integrity of the data, which is essential for accurate subtype prediction. RPSLearner applied feature fusion which concatenated multiple independent RP reduced features for a richer representation learning. Compared to score ensemble strategies used in conventional RP-based algorithms that merely integrated the model scores from multiple RP reduced results, feature fusion could represent a richer aspect of hidden information since it integrated multi-view latent features, and each reduced feature was independently generated from the original transcriptomics data. These results suggested that it could efficiently exploit variances from multiple RP matrices, achieving a more representative latent features for lung cancer subtype prediction. These results demonstrated exploiting multi-scale feature space could enhance the effectiveness and generalization ability of the model. Another unique advantage of RPSLeaner is the stacking framework. RPSLearner leverages a meta-learner, specifically a deep linear neural network to learn the fused features combined multiple RP reduced results and diverse conventional machine learning models. The stacking framework leveraged the complementary strengths of diverse base models as well as the insights from the meta-learner. The meta-leaner could learn the utilize the knowledge from other leading to improved performance across multiple metrics. Our results emphasized that the predictions of meta-model are better than that from each individual base model. The use of stacking learning framework improves the robustness and capacity of the prediction model through leveraging various base models with heterogeneous ability to explore the same data. RPSLearner demonstrated superior performance compared to state-of-the-art lung cancer subtype prediction methods. We tested the other methods, such as MCFS + SVM from Huang et al. 12 which applied MCFS for feature selection, and SVM for cancer subtype prediction. Comparative analysis revealed that RPSLearner consistently showed higher performance across various metrics. A reasonable explanation to RPSLearner’s superior performance was that the application of fused features and stacking learning to leverage both RP transformed features and predictions from diverse base models rather than simply averaging the prediction results. RPSLearner exhibited biological interpretable to identify characteristic biomarkers from NSCLC subtypes. RPSLearner not only extract meaningful hidden features from transcriptomics data but also offers insights about the detection of novel potential biomarkers. These results highlighted its potential for clinical application. Besides, the method could be extended to more cancer types that each cancer type may present distinct molecular and genomic characteristics, making is essential to evaluate the effectiveness in universal cancer subtypes prediction 28 . Moreover, the method could also be extended to other tasks related to the cancer diagnosis such as cancer progression prediction, where the goal shifts from classifying subtypes to predict tumor growth or patient prognosis over time 29 . By incorporating longitudinal data, such as multiple time points of gene expression data, RPSLearner could provide valuable insights into disease progression, ultimately informing personalized treatment strategies 30 . In addition, we also note that there are potential future research directions that can be explored to improve the performance of RPSLearner. First, we need to expand the dataset usage to validate the efficacy of the model, which is necessary for translating these methods into clinical practice. This plan remains challenging due to the high dimensionality of data and the need for models that are both accurate and interpretable 31 , and considering the high computational cost of stacking framework 32 . It requires us to develop more efficient approaches. To solve this concern, we need to fully utilize high-performance computational resources such as GPU to accelerate the computation processing. Conclusion In this paper, to address the concerns in NSCLC subtype prediction, we developed RPSLearner which combines RP and stacking learning for effective and accurate classification. It effectively reduced the dimensionality while preserving sample-to-sample distances through RP and integrated fused features and predictions from diverse models through stacking learning. RPSLearner succeeds in boosting classification prediction with higher accuracy, F1 and AUC metrics than conventional machine learning models and state-of-the-art methods. RPSLearner utilized feature fusion strategy which exhibited better performance than score ensemble approaches in subtype prediction. RPSLearner’s results are interpretable that the expression of DEGs aligns well with the published literature, which also offering insights about potential novel biomarkers. This framework could be potentially extended to subtype identification of other cancers. Methods Dataset This study collected RNA-seq data for LUAD and LUSC retrieved from the NCI Genomic Data Commons (GDC) database 33 with 9 different data sources ( Supplement Fig. 1 ). We excluded samples with incomplete expression profiles and yielded a final dataset of 1,333 cases, with 731 LUAD and 602 LUSC cases, respectively. The raw read counts of all RNA-seq samples were normalized to Transcripts Per Million (TPM). RPSLearner Specifically, considering a data matrix X of size n × d , where n was the number of samples and d was the original dimensionality. RP found a random matrix R ∈ ℝ d×k , with k « d , such that the transformed data matrix Y = X · R , ∈ ℝ n ×k , captured the hidden geometry of the original data X in the reduced dimension k . One common choice of R is a Gaussian RP matrix 34 . Each entry r ij in R is independently sampled from a Gaussian distribution 𝒩(0, σ 2 ), where σ 2 is set to , ensuring that the resulting projected vectors are suitable to Johnson-Lindenstrauss Lemma. In other words, Then for a data point x ∈ ℝ d , its - dimensional projection is y = x · R . With high probability, the Euclidean distances among samples in the projected space ℝ k will be close to those in original space ℝ d . Following RP-based dimensionality reduction, the transformed feature vectors were fed into diverse machine-learning base models, each generating unique predictive labels. These base models encompassed a range of algorithms, including K-Nearest Neighbors (KNN) 35 , Random Forest (RF) 36 , LightGBM 37 , Extra Trees 38 , Catboost 39 , XGBoost 40 , Neural Networks 41 , thereby capturing a wide spectrum of data patterns and interactions. Then, RPSLearner aggregated the predictions from these base models with transformed feature vectors using a stacking meta-learner, specifically a multi-layer perceptron (MLP), which synthesized the individual predictions and features to enhance overall classification performance. This stacking strategy not only mitigated the biases inherent in individual models but also leveraged their complementary strengths, resulting in a more robust and accurate subtype prediction. In our stacking framework, we employed a diverse set of base classifiers, each generating its own prediction. Let x ∈ ℝ d denote an input feature vector from the dataset and let the set of possible class labels be ℓ = {1,2,… C }. Each base classifier f i (·) returns its predicted label . The stacking method required a meta learner model to integrate the prediction from various base models. Let there be B base models { f 1 , f 1 ,… f N }. Each model f N could output a predicted label .. For numerical stability and richer information, we concatenated of all base model predictions, appending the original input x to form a composite feature . The final prediction ŷ could be formulated as: Where g ( z ( x )) represented the stacking meta-learning model and argmax function aims to get the prediction label with highest probability. Competing Interests The authors declare no conflict of interest. Authors’ Contributions X.W.: data collecting, data preprocessing, model development, data analysis and interpretation, manuscript preparation, editing, and review. J.W.: manuscript editing and review. S.W.: Study concept and design, manuscript editing and review. Data availability The RNA-seq data of lung cancer samples can be publicly accessed from the NCI GDC database: https://portal.gdc.cancer.gov . Code availability The RPSLearner Python package can be accessed at https://github.com/wan-mlab/RPSLearner . Funding information Research reported in this publication was supported by the Office Of The Director, National Institutes Of Health of the National Institutes of Health under Award Number R03OD038391, and by the National Cancer Institute of the National Institutes of Health under Award Number P30CA036727. This work was supported by the American Cancer Society under award number IRG-22-146-07-IRG, and by the Buffett Cancer Center, which is supported by the National Cancer Institute under award number CA036727. This work was also partially supported by the National Institute of General Medical Sciences under Award Numbers P20GM103427. This study was in part financially supported by the Child Health Research Institute at UNMC/Children’s Nebraska. This work was also partially supported by the University of Nebraska Collaboration Initiative Grant from the Nebraska Research Initiative (NRI) . The content is solely the responsibility of the authors and does not necessarily represent the official views from the funding organizations. Funder Information Declared Office Of The Director, National Institutes Of Health of the National Institutes of Health , R03OD038391 National Cancer Institute of the National Institutes of Health , P30CA036727 American Cancer Society , IRG-22-146-07-IRG Buffett Cancer Center , CA036727 National Institute of General Medical Sciences , P20GM103427 Child Health Research Institute at UNMC/Children’s Nebraska University of Nebraska Collaboration Initiative Grant from the Nebraska Research Initiative (NRI) Reference 1. ↵ Thandra KC , Barsouk A , Saginala K , Aluru JS , Barsouk A. Epidemiology of lung cancer . Contemp Oncol . 2021 ; 25 ( 1 ): 45 – 52 . doi: 10.5114/wo.2021.103829 OpenUrl CrossRef PubMed 2. ↵ Herbst RS , Heymach JV , Lippman SM . Lung Cancer . N Engl J Med . 2008 ; 359 ( 13 ): 1367 – 1380 . doi: 10.1056/NEJMra0802714 OpenUrl CrossRef PubMed Web of Science 3. ↵ Herbst RS , Morgensztern D , Boshoff C. The biology and management of non-small cell lung cancer . Nature . 2018 ; 553 ( 7689 ): 446 – 454 . doi: 10.1038/nature25183 OpenUrl CrossRef PubMed 4. ↵ Reck M , Rabe KF . Precision Diagnosis and Treatment for Advanced Non–Small-Cell Lung Cancer . N Engl J Med . 2017 ; 377 ( 9 ): 849 – 861 . doi: 10.1056/NEJMra1703413 OpenUrl CrossRef PubMed 5. ↵ Franklin WA . Pathology of lung cancer . J Thorac Imaging . 2000 ; 15 ( 1 ): 3 – 12 . doi: 10.1097/00005382-200001000-00003 OpenUrl CrossRef PubMed Web of Science 6. ↵ Travis WD , Brambilla E , Nicholson AG , et al. The 2015 World Health Organization Classification of Lung Tumors: Impact of Genetic, Clinical and Radiologic Advances Since the 2004 Classification . J Thorac Oncol . 2015 ; 10 ( 9 ): 1243 – 1260 . doi: 10.1097/JTO.0000000000000630 OpenUrl CrossRef PubMed 7. ↵ Huh J , Hwang W. The Role of Anesthetic Management in Lung Cancer Recurrence and Metastasis: A Comprehensive Review . J Clin Med . 2024 ; 13 ( 22 ): 6681 . doi: 10.3390/jcm13226681 OpenUrl CrossRef PubMed 8. ↵ Collisson EA , Campbell JD , Brooks AN , et al. Comprehensive molecular profiling of lung adenocarcinoma . Nature . 2014 ; 511 ( 7511 ): 543 – 550 . doi: 10.1038/nature13385 OpenUrl CrossRef PubMed Web of Science 9. ↵ Wang Z , Gerstein M , Snyder M. RNA-Seq: a revolutionary tool for transcriptomics . Nat Rev Genet . 2009 ; 10 ( 1 ): 57 – 63 . doi: 10.1038/nrg2484 OpenUrl CrossRef PubMed Web of Science 10. ↵ Guyon I , Elisseeff A. An introduction to variable and feature selection . J Mach Learn Res . 2003 ; 3 ( null ): 1157 - 1182 . OpenUrl CrossRef 11. ↵ Huang Z , Chen L , Wang C. Classifying Lung Adenocarcinoma and Squamous Cell Carcinoma using RNA-Seq Data . Cancer Stud Mol Med – Open J . 2017 ; 3 ( 2 ): 27 – 31 . doi: 10.17140/CSMMOJ-3-120 OpenUrl CrossRef 12. ↵ Yuan F , Lu L , Zou Q. Analysis of gene expression profiles of lung cancer subtypes with machine learning algorithms . Biochim Biophys Acta BBA - Mol Basis Dis . 2020 ; 1866 ( 8 ): 165822 . doi: 10.1016/j.bbadis.2020.165822 OpenUrl CrossRef PubMed 13. ↵ Zhang J , Li Y , Dai W , et al. Molecular classification reveals the sensitivity of lung adenocarcinoma to radiotherapy and immunotherapy: multi-omics clustering based on similarity network fusion . Cancer Immunol Immunother . 2024 ; 73 ( 4 ): 71 . doi: 10.1007/s00262-024-03657-x OpenUrl CrossRef PubMed 14. ↵ Carrillo-Perez F , Morales JC , Castillo-Secilla D , Gevaert O , Rojas I , Herrera LJ . Machine-Learning-Based Late Fusion on Multi-Omics and Multi-Scale Data for Non-Small-Cell Lung Cancer Diagnosis . J Pers Med . 2022 ; 12 ( 4 ): 601 . doi: 10.3390/jpm12040601 OpenUrl CrossRef PubMed 15. ↵ de Margerie-Mellon C , Chassagnon G. Artificial intelligence: A critical review of applications for lung nodule and lung cancer . Diagn Interv Imaging . 2023 ; 104 ( 1 ): 11 – 17 . doi: 10.1016/j.diii.2022.11.007 OpenUrl CrossRef PubMed 16. ↵ Li L , Xiao H , Wu X , et al. RanBALL: An Ensemble Random Projection Model for Identifying Subtypes of B-Cell Acute Lymphoblastic Leukemia . BioRxiv Prepr Serv Biol . Published online January 2, 2025:2024.09.24.614777. doi: 10.1101/2024.09.24.614777 OpenUrl Abstract / FREE Full Text 17. ↵ Cannings TI , Samworth RJ . Random-projection Ensemble Classification . J R Stat Soc Ser B Stat Methodol . 2017 ; 79 ( 4 ): 959 – 1035 . doi: 10.1111/rssb.12228 OpenUrl CrossRef 18. ↵ Pavlyshenko B. Using Stacking Approaches for Machine Learning Models . In: 2018 IEEE Second International Conference on Data Stream Mining & Processing (DSMP) .; 2018 : 255 – 258 . doi: 10.1109/DSMP.2018.8478522 OpenUrl CrossRef 19. ↵ Erickson N , Mueller J , Shirkov A , et al. AutoGluon-Tabular: Robust and Accurate AutoML for Structured Data . Published online March 13, 2020. doi: 10.48550/arXiv.2003.06505 OpenUrl CrossRef 20. ↵ Wan S , Kim J , Won KJ . SHARP: hyperfast and accurate processing of single-cell RNA-seq data via ensemble random projection . Genome Res . 2020 ; 30 ( 2 ): 205 – 213 . doi: 10.1101/gr.254557.119 OpenUrl Abstract / FREE Full Text 21. ↵ Abdi H , Williams LJ . Principal component analysis . WIREs Comput Stat . 2010 ; 2 ( 4 ): 433 – 459 . doi: 10.1002/wics.101 OpenUrl CrossRef 22. ↵ Maaten L van der , Hinton G. Visualizing Data using t-SNE . J Mach Learn Res . 2008 ; 9 ( 86 ): 2579 – 2605 . OpenUrl CrossRef PubMed Web of Science 23. ↵ McInnes L , Healy J , Melville J. UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction . Published online September 18, 2020. doi: 10.48550/arXiv.1802.03426 OpenUrl CrossRef 24. ↵ Hearst MA , Dumais ST , Osuna E , Platt J , Scholkopf B. Support vector machines . IEEE Intell Syst Their Appl . 1998 ; 13 ( 4 ): 18 – 28 . doi: 10.1109/5254.708428 OpenUrl CrossRef 25. ↵ Chen JW , Dhahbi J. Lung adenocarcinoma and lung squamous cell carcinoma cancer classification, biomarker identification, and gene expression analysis using overlapping feature selection methods . Sci Rep . 2021 ; 11 ( 1 ): 13323 . doi: 10.1038/s41598-021-92725-8 OpenUrl CrossRef PubMed 26. ↵ Winslow MM , Dayton TL , Verhaak RGW , et al. Suppression of Lung Adenocarcinoma Progression by Nkx2-1 . Nature . 2011 ; 473 ( 7345 ): 101 – 104 . doi: 10.1038/nature09881 OpenUrl CrossRef PubMed Web of Science 27. ↵ Li N , Zhai Z , Chen Y , Li X. Transcriptomic and immunologic implications of the epithelial-mesenchymal transition model reveal a novel role of SFTA2 in prognosis of non-small-cell lung carcinoma . Front Genet . 2022 ; 13 : 911801 . doi: 10.3389/fgene.2022.911801 OpenUrl CrossRef PubMed 28. ↵ Hollmann N , Müller S , Purucker L , et al. Accurate predictions on small data with a tabular foundation model . Nature . 2025 ; 637 ( 8045 ): 319 – 326 . doi: 10.1038/s41586-024-08328-6 OpenUrl CrossRef PubMed 29. ↵ Kourou K , Exarchos TP , Exarchos KP , Karamouzis MV , Fotiadis DI . Machine learning applications in cancer prognosis and prediction . Comput Struct Biotechnol J . 2015 ; 13 : 8 – 17 . doi: 10.1016/j.csbj.2014.11.005 OpenUrl CrossRef PubMed 30. ↵ Cook DJ , Kallus J , Jörnsten R , Nielsen J. Molecular natural history of breast cancer: Leveraging transcriptomics to predict breast cancer progression and aggressiveness . Cancer Med . 2020 ; 9 ( 10 ): 3551 – 3562 . doi: 10.1002/cam4.2996 OpenUrl CrossRef PubMed 31. ↵ Cruz JA , Wishart DS . Applications of Machine Learning in Cancer Prediction and Prognosis . Cancer Inform . 2006 ; 2 : 117693510600200030 . doi: 10.1177/117693510600200030 OpenUrl CrossRef 32. ↵ Džeroski S , Ženko B. Is Combining Classifiers with Stacking Better than Selecting the Best One? Mach Learn . 2004 ; 54 ( 3 ): 255 – 273 . doi: 10.1023/B:MACH.0000015881.36452.6e OpenUrl CrossRef Web of Science 33. ↵ The NCI Genomic Data Commons | Nature Genetics . Accessed April 23, 2025 . https://www.nature.com/articles/s41588-021-00791-5 34. ↵ Li G , Gu Y. Restricted Isometry Property of Gaussian Random Projection for Finite Set of Subspaces . IEEE Trans Signal Process . 2018 ; 66 ( 7 ): 1705 – 1720 . doi: 10.1109/TSP.2017.2778685 OpenUrl CrossRef 35. ↵ Kramer O Kramer O. K-Nearest Neighbors . In: Kramer O , ed. Dimensionality Reduction with Unsupervised Nearest Neighbors . Springer ; 2013 : 13 – 23 . doi: 10.1007/978-3-642-38652-7_2 OpenUrl CrossRef 36. ↵ Rigatti SJ . Random Forest . J Insur Med . 2017 ; 47 ( 1 ): 31 – 39 . doi: 10.17849/insm-47-01-31-39.1 OpenUrl CrossRef PubMed 37. ↵ Ke G , Meng Q , Finley T , et al. LightGBM: A Highly Efficient Gradient Boosting Decision Tree . In: Advances in Neural Information Processing Systems . Vol 30 . Curran Associates, Inc .; 2017 . Accessed January 21, 2025 . https://proceedings.neurips.cc/paper/2017/hash/6449f44a102fde848669bdd9eb6b76fa-Abstract.html 38. ↵ Sharaff A , Gupta H. Extra-Tree Classifier with Metaheuristics Approach for Email Classification. In: Bhatia SK, Tiwari S, Mishra KK, Trivedi MC , eds. Advances in Computer Communication and Computational Sciences. Springer; 2019:189-197 . doi: 10.1007/978-981-13-6861-5_17 OpenUrl CrossRef 39. ↵ Prokhorenkova L , Gusev G , Vorobev A , Dorogush AV , Gulin A. CatBoost: unbiased boosting with categorical features . In: Advances in Neural Information Processing Systems . Vol 31 . Curran Associates, Inc .; 2018 . Accessed January 21, 2025 . https://proceedings.neurips.cc/paper/2018/hash/14491b756b3a51daac41c24863285549-Abstract.html 40. ↵ Chen T , Guestrin C. XGBoost: A Scalable Tree Boosting System . In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. KDD ‘16 . Association for Computing Machinery ; 2016 : 785 – 794 . doi: 10.1145/2939672.2939785 OpenUrl CrossRef 41. ↵ Hinton GE . Learning multiple layers of representation . Trends Cogn Sci . 2007 ; 11 ( 10 ): 428 – 434 . doi: 10.1016/j.tics.2007.09.004 OpenUrl CrossRef PubMed Web of Science View the discussion thread. Back to top Previous Next Posted May 07, 2025. Download PDF Email Thank you for your interest in spreading the word about bioRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following RPSLearner: A novel approach based on random projection and deep stacking learning for categorizing NSCLC Message Subject (Your Name) has forwarded a page to you from bioRxiv Message Body (Your Name) thought you would like to see this page from the bioRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share RPSLearner: A novel approach based on random projection and deep stacking learning for categorizing NSCLC Xinchao Wu , Jieqiong Wang , Shibiao Wan bioRxiv 2025.05.01.651699; doi: https://doi.org/10.1101/2025.05.01.651699 Share This Article: Copy Citation Tools RPSLearner: A novel approach based on random projection and deep stacking learning for categorizing NSCLC Xinchao Wu , Jieqiong Wang , Shibiao Wan bioRxiv 2025.05.01.651699; doi: https://doi.org/10.1101/2025.05.01.651699 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Bioinformatics Subject Areas All Articles Animal Behavior and Cognition (7624) Biochemistry (17650) Bioengineering (13871) Bioinformatics (41882) Biophysics (21424) Cancer Biology (18566) Cell Biology (25461) Clinical Trials (138) Developmental Biology (13365) Ecology (19867) Epidemiology (2067) Evolutionary Biology (24290) Genetics (15590) Genomics (22476) Immunology (17713) Microbiology (40331) Molecular Biology (17148) Neuroscience (88477) Paleontology (666) Pathology (2828) Pharmacology and Toxicology (4816) Physiology (7635) Plant Biology (15114) Scientific Communication and Education (2044) Synthetic Biology (4286) Systems Biology (9815) Zoology (2268)