Cancer gene expression profiles associated with clinical outcomes to chemotherapy treatments

Background Machine learning (ML) methods still have limited applicability in personalized oncology due to low numbers of available clinically annotated molecular profiles. This doesn’t allow sufficient training of ML classifiers that could be used for improving molecular diagnostics. Methods We reviewed published datasets of high throughput gene expression profiles corresponding to cancer patients with known responses on chemotherapy treatments. We browsed Gene Expression Omnibus (GEO), The Cancer Genome Atlas (TCGA) and Tumor Alterations Relevant for GEnomics-driven Therapy (TARGET) repositories. Results We identified data collections suitable to build ML models for predicting responses on certain chemotherapeutic schemes. We identified 26 datasets, ranging from 41 till 508 cases per dataset. All the datasets identified were checked for ML applicability and robustness with leave-one-out cross validation. Twenty-three datasets were found suitable for using ML that had balanced numbers of treatment responder and non-responder cases. Conclusions We collected a database of gene expression profiles associated with clinical responses on chemotherapy for 2786 individual cancer cases. Among them seven datasets included RNA sequencing data (for 645 cases) and the others – microarray expression profiles. The cases represented breast cancer, lung cancer, low-grade glioma, endothelial carcinoma, multiple myeloma, adult leukemia, pediatric leukemia and kidney tumors. Chemotherapeutics included taxanes, bortezomib, vincristine, trastuzumab, letrozole, tipifarnib, temozolomide, busulfan and cyclophosphamide.


Background
Personalized approach provides important advantages in clinical oncology in terms of improved patient survival and lower drug toxicities [1,2]. However, so far it can only cover a minor fraction of cancer patients [3,4] due to lack of robust prognostic biomarkers for most of the treatments [5]. The proportion of patients eligible for personalized oncology slightly grows. For example, the percentage of US patients with cancer estimated to benefit from personalized prescriptions of targeted therapeutics was only 0.7% in 2006, and it had increased tõ 5% in 2018 [4]. However, this progress could be more significant if more companion diagnostic tests would be available for the standardly used cancer drugs. In this regard, gene expression data, either obtained by RNA sequencing [1] or using microarrays [6], frequently provide an advantage over genomic tests. Several trials and clinical case reports were published recently evidencing high efficiency of gene expression-based prescriptions of cancer chemotherapeutics. Cancer gene expression data can be used per se or can be normalized on the available profiles of healthy human tissues [7].
High-quality training and validation datasets are required to run both types of the above models. Nowadays there is a shortage of clinically annotated molecular data that would help developing ML-assisted diagnostic tools. The datasets available are usually considered too small for applying ML [23,25,26,[29][30][31][32][33]. Indeed, the figure of dozens or hundreds of annotated biosamples is negligible in comparison with~20,000 protein coding genes measured in transcriptomic assays. Intelligent data filtering is, therefore, needed to reduce dimensionality of data [8]. However, a recent approach using dynamic feature extraction, or flexible data trimming, can significantly improve performances of ML-based methods for the real-world datasets [15,25].
This study was performed to review available clinically annotated datasets of cancer transcriptomic profiles that may be suitable for applications in ML models. To our knowledge, this is the largest published collection of processed gene expression data coupled with case history excerpts indicating positive or negative response to certain treatment protocols for cancer patients. This manually curated collection of molecular datasets will be helpful for those working with the ML or artificial intelligence applications in oncology, as well as for the fundamental research and development of cancer biomarkers.

Methods
We curated GEO [34], TARGET [35] and TCGA [36] repositories to extract cancer gene expression profiles associated with the clinical outcomes of chemotherapeutic treatments. We attempted to build a knowledgebase of molecular datasets suitable for building ML classifiers of clinical responses on chemotherapy treatments ( Table 1, Additional file 1). Every included dataset met the following criteria: at least 40 gene expression profiles present; data obtained for the same cancer type and using the same experimental platform every profile is linked with the case clinical history all cancers treated with at least one common drug or chemotherapy regimen treatment outcomes are available enabling to classify every case as either responder or non-responder.
We used different approaches to discriminate between the treatment responders and non-responders. Where available, e.g. for the datasets extracted from the GEO repository, we used the responder/non-responder marks assigned by the authors of the original communications publishing these data. In many instances, the number of response groups was more than two and included groups like "partial responders". However, most frequently binary ML-assisted drug response classifiers are needed that classify patients in only two classes: either responders or non-responders [8,23,25,29,30].
For the TCGA profiles, namely for the low-grade glioma (TCGA-LGG), lung cancer (TCGA-LC), and uterine corpus endothelial carcinoma (TCGA-UEC) datasets, and for the acute myeloid leukemia dataset GSE5122  [56], stable disease cases can be most probably classified as the responders whereas progressive disease casesas the non-responders. For the multiple myeloma dataset GSE9782 [42], the classification can be used as defined by the authors, where patents with complete and partial response were annotated as the responders, and with no change and progressive diseaseas the non-responders. For three other multiple myeloma datasets, namely GSE39753 [43], GSE68871 [44], and GSE55145 [45], complete, near-complete and very good partial response groups can be most likely considered as the responders, whereas partial, minor and worse response groupsas the non-responders. Classification of the TARGET repository profiles was more sophisticated as no responder classification was given by the authors. This was the case for the datasets of pediatric Wilms kidney tumor (TARGET-50), acute myeloid leukemia (TARGET-20) and acute lymphoblastic leukemia (TARGET-10) extracted from the gene expression repository of National Cancer Institute [35]. However, these latter clinical cases were annotated by the time of event-free survival. Distributions of the event-free survival time enabled us to identify for every dataset two different modes of survival with different slopes (Fig. 1), that can be recognized as either responders or non-responders.

Results
For raw gene expression data, the number of features i.e. interrogated genes, usually exceeds the number of tumor cases by roughly two orders of magnitude. Therefore, for robust application of ML the dimensionality of data must be reduced to make the number of selected features lower than the number of tumor cases or at least comparable to it (Fig. 2a). To reduce dimensionality, the gene expression data can be aggregated into the higher-order molecular markers like activation profiles of molecular pathways [23,29,30,60,61]. Alternatively, the most informative fraction of the initial data can be selected that can distinguish between the responder and non-responder classes. For selection of such marker features, several approaches have been proposed, e.g. Pearson chi-squared test [62], correlation test [27,62], variance thresholding, genetic algorithms [63], univariate feature selection, recursive feature elimination, principal component analysis [27], CUR matrix [64], decomposition [65] and covariate regression [66].
In the current research, we applied the following leave-one-out-based method for finding robust marker features [25] (Fig. 2c). Imagine that we have a gene expression dataset that embraces N clinical cases, each with corresponding expression profile. For each clinical case i = 1, … N, we determine the top Q marker genes that distinguish responding and non-responding cases in a sub-dataset that contains all samples but i. In other words, for all N sub-datasets each having N-1 cases, we interrogate each gene taken one by one and retrieve the top Q set of genes that showed the highest ROC AUC values for the difference between responder and nonresponder profiles. The quality metric area under the ROC curve (AUC) is the universal metric of a biomarker robustness that depends on its sensitivity and specificity Table 1 Overview of selected transcriptomic datasets of responders/non-responders to cancer chemotherapy, responders (R) vs non-responders (NR) (Continued) [54] GSE37946  Fig. 1 Distribution of event-free survival time for the patients with (a) childhood kidney Wilms tumor from TARGET-50 dataset, (b) childhood ALL from TARGET-10 dataset and (c) childhood AML from TARGET-20 dataset [35]. Patients on the left from vertical threshold can be considered as the non-responders, and on the rightas the responders to the treatment [67]. It positively correlates with the quality of a biomarker and varies from 0.5 till 1. The standard discrimination threshold is 0.7 and the entries with higher AUC are considered high-quality biomarkers, and vice versa [68]. AUC is broadly used for detection of biomarkers in oncology [69][70][71][72][73].
To provide trobust feature selection, the number Q shouldn't exceed the number of cases N. In the current application, we took Q equal to 30 because all tdatasets under consideration had more than 40 cases. The final list of core marker genes was obtained by intersecting top Q gene sets for all N sub-datasets.
We applied this procedure to all the clinically annotated cancer transcriptomic datasets under consideration and identified for them core marker genes (Table 1). Twentythree out of 26 datasets investigated provided 7-20 core marker gene features for further ML applications ( Table 1).
The remaining three datasets, namely GSE22358 [52], GSE76360 [58] and TCGA-UEC [36], were poorly balanced because the numbers of responders greatly exceeded the respective numbers of non-responders, or vice versa. For these three instances we were unable to generate robust core marker gene sets for ML applications because the number of such genes was too low (two-three per dataset, Table 1). Fig. 2 Possible scenarios of using ML to build classifiers based on gene expression datasets. a Methods data dimensionality reduction; b approaches to merging and enlarging of gene expression datasets for ML application; c general workflow for a core marker set determination

Discussion
By the current moment, ML hasn't made a revolution in biomedicine [12]. This may be partly connected with the relatively recent emergence of experimental methods generating big amounts of biomedical data combined with the developed IT infrastructure. Among these game-changing methods the major role was played by the next-generation sequencing (NGS) and novel massspectrometry approaches which made whole genome-, transcriptome-, proteome-and metabolome analyses relatively fast and cheap [74][75][76].
Yet further development of ML methods in personalized oncology is still strongly limited by the low number of clinically annotated cancer patient molecular datasets. A dataset suitable for ML should have all together enough number of high-throughput molecular profiles and also the associated clinical case history records featuring success of the therapeutic regimen used.
In this paper we reviewed three major repositories of omics data for the available responder/non-responder datasets including more the 40 cancer cases treated with the same chemotherapeutics. We identified 26 datasets with totally 2786 cases, ranging from 41 till 508 cases per dataset (Table 1). We checked the robustness of these datasets and their suitability for ML applications using our previous method of core maker feature determination [25]. According to this test, 23/26 datasets were suitable for ML, each having 7-20 core marker genes/features for further ML applications. Contrarily, the remaining three datasets produced only two or three features, which may seem insufficient for the ML. Poor performance of these three datasets was most likely due to unbalanced numbers of clinical responder/non-responder cases included.
To increase the number of cases (Fig. 2b), the datasets for the same disease or drug treatment conditions can be merged using cross-dataset harmonization. Different methods can be used to harmonize data obtained using the same [77,78] or two different experimental platforms [79,80], or even using multiple platforms [81] (Fig. 2b).
In addition, when the cases are deficient, transfer learning methods may be used for a certain disease or drug condition. Using this approach, the ML training process may be preformed on the multiple available molecular profiles corresponding to cell culture treated with certain drugs [82], whereas the ML classifier validation may be done on more rare patient cancer cases [23,29,30].

Conclusions
We identified 26 clinically annotated gene expression datasets ranging from 41 till 508 cases per dataset (Table 1). Collectively, they covered 2786 individual cancer cases. Among them seven datasets included RNA sequencing data (for 645 cases) and the othersmicroarray expression profiles. The datasets represented breast cancer, lung cancer, low-grade glioma, endothelial carcinoma, multiple myeloma, adult leukemia, pediatric leukemia and kidney tumors. Chemotherapeutics used included taxanes, bortezomib, vincristine, trastuzumab, letrozole, tipifarnib, temozolomide, busulfan and cyclophosphamide.
We hope that presented collection of clinically annotated transcriptomic profiles will be useful to those working with data analysis in oncology, as well as for the fundamental research and development of nextgeneration cancer biomarkers.
Additional file 1. Clinically annotated datasets and samples they contain.

About this supplement
This article has been published as part of BMC Medical Genomics Volume 13 Supplement 8, 2020: Selected Topics in "Systems Biology and Bioinformatics" -2019: medical genomics. The full contents of the supplement are available online at https://bmcmedgenomics.biomedcentral.com/articles/supplements/ volume-13-supplement-8.
Authors' contributions NB, MS and AB contributed conception and design of the study. NB, MS, VT and AG analyzed the data. NB and AB wrote the manuscript. All authors read and approved the final manuscript.

Funding
The work was funded by the Russian Science Foundation grant no. 18-15-00061. Publication costs have been funded by the Russian Science Foundation grant no. 18-15-00061. The funding body played no role in the design of the study and collection, analysis, and interpretation of data and in writing of the manuscript.

Availability of data and materials
All the data, including IDs of expression profiles, treatment methods and clinical response assessment, both done by the teams, who had worked with them, and our binary ("P-vs-N") response classifications, are provided in Additional file 1.
Ethics approval and consent to participate Current research did not involve any new human material. All the gene expression data that were used for research were taken from publicly available repositories GEO, TARGET and TCGA and had been previously anonymized by the teams, who had worked with them.

Consent for publication
Not applicable.