1 Introduction

Worldwide, cancer is a major cause of death and a significant barrier to life expectancy [1]. The World Health Organization (WHO) identified cancer as the primary or secondary cause of death for people under the age of 70 in 112 out of 183 nations and as the third or fourth leading cause of death in a further 23 countries [2]. In 2022, the most commonly diagnosed cancer was lung cancer, accounting for 12.4% of all cancers worldwide, followed by cancers of the female breast, colorectum, prostate, and stomach [3]. The International Agency for Research on Cancer (IARC), estimated that in a lifetime, approximately 20% of people develop cancer, whereas around one in nine men and one in 12 women die from it [4]. For all cancers combined, despite differences between countries, regions, ethnicities, and lifestyles, there is a gender gap, with higher incidence and mortality in men than in women [4]. It can also affect people of all ages and, although more common in adults, affects around 400,000 children each year. Cancer is a complex disease caused by multiple factors, including genetics, environment, and behavior. These causes can range from exposure to toxic substances and hereditary genetic predispositions to viral infections, lifestyle choices like smoking and diet, and aging.Footnote 1

Cancer is characterized by the uncontrolled growth and abnormal spread of abnormal cells in the body. This abnormal cell proliferation can form masses of tissue that disrupt the normal functions of organs and body systems [5]. Genetic changes are necessary for the development and progression of cancer. However, they are not sufficient because cancer is not just a genetic disorder but a complex ecosystem involving multiple non-cancerous cells and their interactions within the tumor [6]. Although cancer research has made significant progress in recent years, achieving a complete understanding of the disease and developing effective therapies remain major challenges. The inherent complexity of cancer continues to hinder our ability to optimally prevent, diagnose, and treat the disease.

Artificial Intelligence (AI) is increasingly important in cancer research, offering powerful tools to tackle complex challenges and accelerate scientific progress [7]. This interdisciplinary field combines computer science, mathematics, and psychology to create machines that perform tasks requiring human intelligence. AI techniques include rule-based systems, natural language processing, and computer vision. Machine learning (ML), a subset of AI, involves the development of algorithms that allow computers to learn from data and make predictions or decisions based on it. ML allows systems to learn and improve from experience without requiring explicit programming, unlike traditional programming where explicit instructions are given. AI can help develop more precise and effective diagnostic and therapeutic approaches, improving the accuracy of tumor diagnosis and guiding personalized therapies. This personalization relies on the identification of unique molecular features of a patient’s cancer, enabling targeted and optimized treatments.

Feature selection in deep learning and biomarker identification are crucial to this aim. Unlike traditional machine learning methods that rely on explicit feature selection, deep learning models can extract complex patterns and representations from data without predefined features. In addition, AI is making a significant contribution to data management and analysis. It can speed up the analysis of results, identify patterns of response to treatments, and help to select the most appropriate participants for trials [8]. AI offers many promises and opportunities in the context of patient care and the evolution of medical practice [9] and is emerging as a powerful tool in cancer research, particularly for its ability to analyze vast amounts of incredibly complex genomic data.

Advances in technological tools to enable more efficient and cost-effective sequencing, coupled with robust computational algorithms to extract biologically meaningful information from vast amounts of data, played a critical role in the success of international cancer genome sequencing projects [10, 11]. Genomic data are an invaluable source of knowledge for understanding cancer at the cellular level, acting as a comprehensive guide, revealing the genetic make-up and providing detailed information about the genetic profile of cancer cells. AI algorithms can analyze massive amounts of genomic data and identify subtle patterns and relationships that are difficult to detect using traditional methods and may be missed by humans. These patterns can reveal biomarkers, driver mutations and drug resistance mechanisms, enabling earlier and more accurate diagnosis and opening the door to targeted therapies that can overcome resistance. Genomics plays a crucial role in assessing cancer risk by studying genetic mutations, gene expression, protein activity and epigenetic changes. The ability to profile each patient for a comprehensive set of clinically actionable genomic alterations opens new perspectives in the exploration of new therapeutic targets [12] and underpins the emerging vision of personalised cancer medicine [13]. In the area of treatment personalization, AI enables the analysis of extensive patient genomic and clinical data to identify specific patterns and biomarkers that guide treatment selection. This means that patients could receive targeted therapies based on the specific genetic characteristics of their tumor, maximizing efficacy and minimizing side effects. More accurate and earlier diagnoses are also possible; AI models can analyze genomic information to detect early signs of cancer or genetic disorders. This advanced diagnostic capability could enable more effective therapeutic interventions in the early stages of disease. Identifying complex and non-linear correlations in data reveals new insights that may not have been possible with traditional approaches. Integrating AI into clinical decision-making processes can therefore improve the efficiency and accuracy of medical practice.

However, despite the exciting progress being made, there are challenges to the application of AI in cancer research. The need to ensure the interpretability of algorithms, address ethical issues and resolve biases in data are critical aspects that require ongoing attention [14]. This review examines AI applications to genomic data in cancer research, focusing on five key areas:

  1. 1.

    Types of genomic data used with AI, including DNA sequences, RNA-seq, DNA methylation, and proteomics.

  2. 2.

    Common machine learning (ML) and deep learning (DL) methodologies in genomic data analysis.

  3. 3.

    Recent advances and successful AI applications in oncology, with examples.

  4. 4.

    Ethical and technical challenges in using AI for cancer research.

  5. 5.

    Future perspectives on AI in cancer research and its potential impact on patient care.

The paper is structured as follows: Sect. 1 introduced the topic of cancer research using AI in the field of genomics. Section 2 presented the existing literature investigations on the relevant topic. Section 3 defines the inclusion and exclusion criteria for constructing our dataset, and describes the results obtained from the Scopus databases. In Sect. 4, we propose seven research questions. Specifically, in the subsections 4.1, 4.3, 4.4, we provide a quantitative analysis of the state-of-the-art concerning the application of AI techniques in cancer research using genetic data. We focus on the main publication channels and the countries where the most active research centers are located. In addition, we analyse the selected articles in 4.4. In the subsections 4.5, 4.6, 4.7, we discuss the application domains of the proposed techniques, provide technical insights into the problem formalization, discussed te performance metrics used for evaluation, and finally, considere the challenges addressed by each article. In Sect. 5, we discuss the selected papers. In the concluding Sect. 6, we illustrate our study’s key findings and implications.

2 Related work

In recent years, oncology research has increasingly embraced the multidisciplinary approach of the omics sciences. This integrated approach, encompassing disciplines such as genomics, transcriptomics, proteomics, and metabolomics, has proven to be crucial in gaining a comprehensive and detailed understanding of the underlying mechanisms of cancer. This section provides a comprehensive review of related research, focusing on the importance of the field of omics in the context of oncology.

Zhang et al. [15] provide a comprehensive review of the current applications of AI in oncology. The authors explore various AI applications, focusing on how this technology can be used to analyze omics data, improve the accuracy of early diagnosis, and support the personalization of oncological treatments. The research involves three major medical databases (MEDLINE, CENTRAL, and Embase), examining publications up to November 2023 and focusing on the last 2 years of relevant research. A total of 254 articles are included in the review. This article highlights that the power of AI extends to prognosis, where predictive models based on omics data provide a detailed overview of disease progression and future risks for the patient.

Arjmand et al. [16] discuss the importance of integrating omics sciences in cancer research, focusing on genomics, transcriptomics, proteomics, metabolomics, and others. The authors emphasize the importance of epigenetic regulation in cancer research, highlighting mechanisms such as DNA methylation, histone modifications, and microRNAs that affect genomics function without altering the DNA sequence. The discussion focuses on how these epigenetic modifications may influence the tumor microenvironment and contribute to tumorigenesis. The omics sciences are then extended to include transcriptomics, proteomics, and metabolomics, providing further insights into biological processes at the RNA, protein, and metabolite levels. The article highlights how these multi-omics approaches can lead to discoveries in biomarkers, treatment responses, and the overall understanding of cancer. ML in omics sciences integrates data from multiple sources to improve understanding, accelerate research, and optimize the analysis of large datasets.

Dixit et al. [17] analyze the integration of CRISPR-based genome editing technologies with AI and their impact on medicine and human health. The paper describes how AI is being used to design optimal guide RNAs (gRNAs) for CRISPR-Cas systems, thereby increasing the precision of genomics modifications. AI models, such as DeepCRISPR and CRISTA take into account various factors to predict effective gRNAs, helping to optimize advanced genome editing technologies. Combined with genome editing and precision medicine, AI enables personalized treatments based on patient’s genetic profiles, identifying mutations associated with diseases such as cancer. However, the authors highlight challenges such as high costs, accuracy of off-target modifications and methods of delivering CRISPR payloads. The review also explores the future potential of AI in CRISPR genome editing, opening up new perspectives for genetics, biomedicine and human health.

Tran et al. [18] highlight the increasing adoption of deep learning in healthcare in recent years, in particular in well-characterized cancer datasets, which have accelerated research into its utility in analyzing the complex biology of cancer. This review provides an overview of emerging DL techniques and their applications in oncology, focusing on omics data types (genomic, methylation and transcriptomic) and pathology-based genomic inference. The review also highlights several challenges to widespread clinical implementation. Key challenges included data variability, a lack of phenotypically characterized datasets, the importance of the interpretability of AI, and the quantification of uncertainty in predictions. Data variability and the scarcity of high-quality, phenotypically characterized datasets are barriers, but interpretability methods can facilitate model adaptation for clinical use. In addition, the interpretability of AI and the ability to handle prediction uncertainty are critical for clinical acceptance and regulatory compliance.

Quazi [19] explore the application of AI and ML in precision medicine and genomics. The article highlights the ability of AI to analyze multi-omics data, integrating epidemiological, demographic, clinical and imaging information to improve diagnosis and personalized treatment. Precision medicine is described in terms of four approaches: predictive, preventive, personalized and participatory. The importance of a future-oriented medical environment is emphasized, enabling physicians to have a clear view of the patient by integrating phenotypic details, lifestyle and non-medical factors into medical decisions.

Dias et al. [20] look at the application of AI in precision medicine and genomics, highlighting both progress and challenges. They explain how AI was used in different clinical contexts, addressing issues such as computer vision, time series analysis, speech recognition and natural language processing. In clinical genomics, AI addresses variant calling, annotation, classification of genomic variants, and phenotype-genotype mapping. Challenges related to interpretability, data bias and regulatory issues are explored, emphasizing the need for transparency and fairness.

3 Research methodology

This section describes the research methodology used in this study, which consists of five stages that can be summarised as follows: i) Definition of research questions, ii) Preliminary data analysis, iii) Definition of inclusion and exclusion criteria, iv) Identification of relevant studies based on inclusion and exclusion criteria, and v) Data extraction and analysis. For each document we consider the problem addressed by the authors, its formalization, the approach used and the challenges faced. The first step is to define the research questions. Inspired by other survey articles such as [21] and [22], we decided to consider the following ones:

  1. 1.

    RQ1: How many scientific studies on the application of AI to genomic data in cancer research have been published between 2013 and 2024?

  2. 2.

    RQ2: What were the main publication channels?

  3. 3.

    RQ3: Which countries had the most active research?

  4. 4.

    RQ4: Which applications and methods have been used most?

  5. 5.

    RQ5: Which were the most commonly used algorithms?

  6. 6.

    RQ6: What parameters were used to assess performance?

  7. 7.

    RQ7: What challenges have been faced?

The database used to collect the papers is Scopus, a comprehensive academic research database that provides abstracts and citations across a range of disciplines. The database offers options for setting alerts, advanced search features and journal analysis [23].

To limit the scope of our research, we used the following search string: "(("Artificial Intelligence" OR "AI" OR "Machine Learning" OR "ML" OR "Deep Learning" OR "DL" ) AND ( "cancer" OR "tumour" OR "neoplasia" OR "neoplasm" OR "malignancy") AND ("genomic data" OR "genomics" OR "gene expression"))". To refine the results of our analysis, we used the following inclusion and exclusion criteria:

  • Inclusion criteria: The papers eligible for analysis were limited to those:

    • written in English;

    • written between 2013 and 2024;

    • published in a journal article;

    • clearly focused on Artificial Intelligence;

    • related to the medical field.

  • Exclusion criteria: In the case of duplicate articles, less recent versions were not included in the analysis.

Our analysis returned 147.586 results. We then examined the more common subject areas, as shown in Fig. 1 and selected the 58.159 papers related to the medical field.

Fig. 1
figure 1

Documents by subject area for the Scopus database

Full size image

4 Research questions

The analysis of the initial set of selected papers was based on the search strings outlined in Sect. 3.

4.1 RQ1: How many scientific studies on the application of AI to genomic data in cancer research have been published between 2013 and 2024?

This research question aims to quantify the interest of the international scientific community in applying AI methods to genomic data in medicine over the last 10 years.

Fig. 2
figure 2

Academic studies published from 2013 to 2023 in Scopus database

Full size image

As shown in Fig. 2, the number of publications remained relatively low until 2013, with less than 2.362 publications per year. Since 2018, we observed a rapid exponential growth of interest in this topic, reaching 9.876 in 2022, with subsequent stabilization of growth in 2023 with 9.974 articles.

4.2 RQ2: What were the main publication channels?

We aim at identifying and analyzing the primary mediums or platforms through which research findings, articles, papers, are disseminated. The Fig. 3 shows the results of our analysis.

Fig. 3
figure 3

Documents per year by source from 2013 to 2023 in Scopus database

Full size image

Academic journals that published the highest number of documents are Frontiers In Immunology with 1.599 papers, Frontiers In Genetics with 1.430 papers, Frontiers In Oncology with 1.242 papers, Cancers with 1.122 papers, and Oncotarget with 1.004 papers. As shown in Fig. 3, there was a peak in publications in 2022, followed by a decline in 2023.

4.3 RQ3: Which countries had the most active research?

This research question focused on countries whose research centers contribute to the study of the application of AI in cancer research with genomic data. Figure 4 shows the main countries that contribute to this research development in this field and, as we can easily see, the largest number of articles are from China (2.1856 articles), followed by the United States (1.5410 articles ), Germany (3.475 articles), United Kingdom (3.352 articles), Italy (2.456 articles), Japan (2.386 articles), South Korea (2.116 articles), Canada (2.108 articles), India (2.012 articles) and Iran (1.866 articles).

Fig. 4
figure 4

Number of publications per Country on the application of AI in cancer research using genomic data

Full size image

We also carried out an analysis of the institutions that have made significant contributions to this field, and as shown in Fig. 5, the main institutions include: Ministry of Education of the People’s Republic of China with 1.552 documents; Harvard Medical School with 1.196 documents; Fudan University with 902 documents; Chinese Academy of Sciences with 807 documents; Shanghai Jiao Tong University School of Medicine with 804 documents; Chinese Academy of Medical Sciences & Peking Union Medical College with 760 documents; Inserm with 706 documents; National Institutes of Health NIH with 705 documents; Central South University with 675 documents and The University of Texas MD Anderson Cancer Centre with 671 documents.

Fig. 5
figure 5

Document by affiliation of the application of AI in cancer research with genomic data

Full size image

4.4 RQ4: Which applications and methods have been used?

In our analysis and review, we focus on documents published between 2022 and 2023. This particular period was characterized by a significant increase in publications and represents the peak of research output. By targeting these documents, we collected a total of 19.850 articles, to capture the most recent and productive contributions to the field during this period. This deliberate approach allows us to provide a comprehensive and up-to-date analysis, highlighting the latest advances and insights in the field.

To investigate the application domains and AI techniques used in cancer research, we used genomic data. Our methodology involved performing a clustering analysis of scientific article titles, which helped us identify potential thematic clusters within our dataset. The process was divided into several key stages:

  • Document creation: We started the process by creating a text document from the titles column of our dataset. This document served as the basis for the clustering analysis.

  • Text pre-processing: We performed several pre-processing operations to standardize the titles and facilitate the analysis. This included converting the text to lowercase, removing punctuation, numbers, and stop words, and cleaning up white spaces.

  • Creation of Document-Term Matrix (DTM): We used the document-term matrix to represent the frequency of words in the article titles. This matrix was crucial for the application of the clustering algorithm.

  • Application of K-means algorithm: We chose the K-means algorithm to perform document clustering. The number of clusters (k=3) was selected based on the nature of our dataset and the analysis requirements.

  • Dimensionality reduction with t-SNE: To visualize the clusters effectively, we used the t-SNE algorithm to reduce the dimensionality of the document-term matrix to two dimensions.

  • Assignment of clustering labels: We assigned clustering labels to the documents based on the assignments generated by the K-means algorithm. The papers were clustered exclusively into clusters 1 and 3, with no articles assigned to cluster 2.

  • Principal Component Analysis (PCA) execution: We performed Principal Component Analysis (PCA), by standardizing the data and specifying the desired number of principal components.

  • Manipulation of principal components: We extracted the principal component scores obtained from the PCA.

  • Dataset ordering: We ordered the dataset in descending order based on the first principal component.

  • Calculation and visualization of explained variance: We calculated the eigenvalues, and the percentage of explained variance, and set a threshold of 0.05. We extracted only the principal components that exceeded the threshold.

  • Creation of a new dataset with relevant principal components: We created a new dataset containing identifiers, relevant principal components, and columns from the original dataset.

  • Ordering the dataset based on the relevance of principal components: We ordered the dataset based on the relevance of the principal components, using indices obtained from eigenvalues, and extracted only the first 10 most relevant articles.

  • Cluster visualization: We added the a new column, labelled by cluster, to the new dataset and visualized its distribution as shown in Fig. 6.

Fig. 6
figure 6

Distribution of clusters within the dataset

Full size image

Dimensionality reduction with t-SNE (t-distributed Stochastic Neighbor Embedding) was used to visualize high-dimensional data in a two- or three-dimensional space. The main goal of t-SNE was to preserve the similarity relations between the original instances during the projection into a lower dimensional space.

It computed the similarities between all pairs of instances in the original dataset, computed a second probability distribution in the new (reduced-dimensional) space so that the distances in the reduced space reflected the similarities computed in the previous step, and minimized the Kullback–Leibler divergence between the two probability distributions. This process optimized the positions of the points in the reduced dimensional space to reflect the original similarities.

PCA, or Principal Component Analysis, is a method used to analyze and represent the variability present in data by reducing its dimensions. The primary goal of PCA is to transform a collection of correlated variables into a new set of uncorrelated variables, called "principal components". In our experiment, the principal components we derived were used primarily for two purposes: reducing the dimensionality of the data and selecting relevant features.

The tasks consider the items presented in Cluster 1 and shown in Table 1.

Table 1 Article Titles and Corresponding Principal Component Values (PCA) related to Cluster 1
Full size table

Although these articles cover different topics within the field of cancer research, they share some key elements:

  1. 1.

    Focus on cancer: All articles focused explicitly on different aspects of cancer research, whether it is understanding disease mechanisms, developing new treatments, or improving diagnosis and treatment.

  2. 2.

    Utilisation of knowledge from genetics and/or genomics:

    • The first article in the list above [24] discussed PD-L1, TMB, and other potential predictors of immunotherapy response in hepatocellular carcinoma (HCC). This article directly used genetic markers such as PD-L1 and genomic data such as TMB to predict immunotherapy response.

    • The second article [25] analysed the impact of TP53 genomics alterations in large B-cell lymphoma treated with CD19 antigen chimeric receptor T-cell therapy. This study investigated the impact of specific genetic alterations (TP53) on the effectiveness of CAR-T therapy.

    • The third article [26] proposed a molecularly integrated grade for meningioma. This article presented a new grading system for meningioma based on genetic information, including karyotype and gene expression analysis.

    • The fourth article [27] discussed the diagnosis and management of AML in adults. This article highlighted the increasing importance of genetic data (including next-generation sequencing) in the diagnosis and management of AML.

    • The fifth article [28] discussed the spatial landscape of progression and immunoediting in primary melanoma at single-cell resolution. Although not directly using genetic markers, this study used spatial transcriptomics to provide information on single-cell gene expression within tumor tissue.

    • The sixth article [29] discussed the validation of a new highly sensitive multi-target blood test for early-stage HCC. This article used a blood test based in part on DNA methylation analysis, a genetic alteration associated with cancer development.

  3. 3.

    Advancement of cancer research and applications: Each article contributed in a different way to the advancement of cancer research and its practical applications:

    • The first [24] and the sixth articles [29] aimed to improve diagnosis and treatment selection by identifying potential biomarkers for immunotherapy and early detection of cancer.

    • The second [25]) and the fifth articles [28] explored the effectiveness of specific therapies (CAR-T cell therapy and single-cell analysis) concerning specific genetic factors.

    • The third [26] and the fourth articles [27] proposed new classification systems and management strategies based on the integration of genetic information with other clinical data.

The task considered the elements presented in Cluster 3 and shown in Table 2

Table 2 Article titles and corresponding Principal Component Values (PCA) related to Cluster 3
Full size table

These latest articles also focused on different aspects of cancer research but shared some key elements:

  1. 1.

    Attention to the tumor microenvironment (TME): All articles explored the Tumour Microenvironment (TME), the ecosystem that surrounds tumour cells and plays a critical role in tumor growth, invasion and response to treatment:

    • The first article [30] investigated the study of the functional heterogeneity of Carcinoma-Associated Fibroblasts (CAFs) within the pancreatic cancer tumor microenvironment (TME) and their role in IL-6-mediated therapy resistance.

    • The second article [31] discussed the potential of modulating the intestinal microbiota to enhance the effectiveness of cancer therapy, highlighting the role of the intestinal microbiota as part of the broader tumor microenvironment.

    • The third article[32] analyzed the interaction between Cancer-Associated Fibroblasts (CAFs) and Lipid-Associated Macrophages (LAM) in breast cancer, highlighting their role in creating immunosuppressive tumor microenvironment.

  2. 2.

    Exploration of new therapeutic strategies: while the specific approaches differ, each article explored new avenues to improve cancer treatment:

    • The first article [30] proposed targeting specific subpopulations of Cancer-Associated Fibroblasts (CAFs) based on their IL-6 production to overcome therapeutic resistance.

    • The second article [31] proposed the modulation of the intestinal microbiota as a potential strategy to enhance the effectiveness of existing antitumor therapies.

    • The third article [32] identified the interaction between CAFs and LAM as a potential target for the development of new immunotherapies for breast cancer.

    • The fourth article [33] explored the novel mechanism of ENO1 in suppressing tumor cell ferroptosis, paving the way for potential therapeutic strategies.

  3. 3.

    Use of advanced research techniques: all articles used advanced research techniques to investigate their topics:

    • The first article [30] used techniques such as single-cell RNA sequencing (scRNA-seq), immunohistochemistry, and genetic mouse models.

    • The second article [31] highlighted the research based on techniques such as microbiota sequencing and its analysis.

    • The third article [32] used techniques such as immunohistochemistry, cell culture experiments, and possibly RNA analysis for gene expression.

    • The fourth article [33] used techniques such as cell culture experiments, gene expression analysis and possibly biochemical assays to investigate the role of ENO1 in ferroptosis.

4.5 RQ5: Which were the most commonly used algorithms?

This section analyzed the frequency and use of algorithms in ten carefully selected research papers. Algorithms are the cornerstone for advancing research and development across various fields, making their identification critical for understanding current trends and potential areas for innovation. This concise review aimed to identify the most commonly used algorithms, highlight their applications and discuss their significance within the current research landscape.

Table 3 Most used algorithms in the selected articles
Full size table

Table 3 shows the most frequently used algorithms in the selected articles. The study conducted by Chalasani et al. [29] focused on the development and validation of a new multitarget blood test for the early diagnosis of hepatocellular carcinoma. The mt-HBT algorithm underwent a thorough analysis using three types of models: LASSO regression (Least Absolute Shrinkage and Selection Operator) [34], Random Forest [35], and Logistic Regression [36]. Each model explored different combinations of biomarkers and demographic parameters. The evaluation criteria were based on accuracy and robustness of performance. The logistic regression model, which included three methylation markers (HOXA1, TSPYL5, and B3GALT6), AFP, and patient gender, proved to be the most effective.

The article by Döhner et al. [27] used the Minimal Residual Disease (MRD) algorithm [37], which measures the residual presence of cancer cells in the body after treatment. In addition, the article mainly focused on updated recommendations for the diagnosis and management of acute myeloid leukemia (AML) in adults, taking into account the latest knowledge on the molecular pathogenesis of AML, advances in diagnostic technology, and new therapeutic agents. The article highlighted the revised European LeukemiaNet (ELN) genetic risk classification and treatment guidelines, which are critical for personalized patient management but does not explicitly mention the use of ML or statistical algorithms in this context.

The study by Driver et al. [26] exemplified the innovative use of ML in clinical research, using a combination of Random Survival Forest [38], Gradient Boosting [39], and Cox Proportional Hazards Model [40] to develop a novel, more predictive grading system for meningiomas. Using these algorithms, the research synergized genetic data with traditional clinical indicators, illustrating the transformative potential of advanced computational methods in improving diagnostic precision and therapeutic guidance.

McAndrews et al. in [30] investigated the functional heterogeneity of fibroblasts associated with pancreatic cancer, focusing on their role in therapeutic resistance. They used cell clustering within the Seurat R software, specifically the Uniform Manifold Approximation and Projection (UMAP) algorithm [41] for non-linear dimension reduction. In addition, pairwise differential expression analysis was performed by comparing individual cell clusters to identify marker genes within each cell cluster.

Nirmal et al. [28] used a multidisciplinary approach to analyze cutaneous melanoma, integrating highly complex imaging, high-resolution 3D microscopy, and spatially resolved transcriptomics. They used the Latent Dirichlet Allocation (LDA) model, a probabilistic technique in statistical data analysis, specifically for modeling themes in documents or relationships between variables. Additionally, they applied the Uniform Manifold Approximation and Projection (UMAP), a dimensionality reduction algorithm used to visualize complex data in lower-dimensional spaces, allowing a better understanding of the tumor microenvironment and the interplay between different cell populations within melanomas. This aligned with their broader objective of mapping the spatial landscape of melanoma progression and immune interactions, providing critical insights into the disease’s biological basis and potential therapeutic targets.

In the study by Shouval et al. [25], univariable and multivariable logistic regression and Cox regression models played an important role in understanding the impact of TP53 genomics alterations on the success of CD19-chimeric antigen receptor T-cell (CAR-T) therapy in patients with large B-cell lymphoma (LBCL). The regression analysis was used to investigate the relationship between TP53 alterations and the likelihood of achieving a complete response (CR) after CD19-CAR-T therapy.

The study by Timperi et al. [32] employed unsupervised clustering analysis to identify and characterize myeloid cell heterogeneity in triple-negative breast cancer (TNBC) using single-cell RNA sequencing (scRNA-seq) technique. This analysis was projected using the Uniform Manifold Approximation and Projection (UMAP) algorithm, providing a visual representation of the distribution of individual cells in the tumor samples and surrounding tissues.

While many studies integrate advanced computational algorithms to dissect complex biological data, the research by Ting et al. (2022) [31] adopted a different methodological approach, focusing on the empirical investigation of the gut microbiota’s interaction with cancer therapies. This study did not utilize ML algorithms but provided critical insights into how the microbiota can influence the pharmacodynamics and pharmacokinetics of cancer treatments.

Similarly, the study by Zhang et al. [33] did not use specific ML or statistical learning algorithms, although it used correlation analysis. The study investigated the role of ENO1 as a promoter of hepatocellular carcinoma (HCC) by inhibiting the ferroptosis pathway. The study elucidated how ENO1 affects the stability and expression of IRP1 and Mfrn1 mRNAs, which subsequently affect mitochondrial iron homeostasis and ferroptosis in HCC. Using correlation analysis, they linked the ENO1/IRP1/Mfrn1 axis to patient clinicopathological characteristics and showed that altered expression of these molecules correlates with HCC progression and patient survival.

Finally, the study by Rizzo et al. [24], did not directly employ ML algorithms but focused on evaluating potential predictive biomarkers such as PD-L1 and TMB (tumor mutational burden) which could significantly influence the effectiveness of immunotherapies for HCC.

4.6 RQ6: What parameters were used to assess performance?

In the context of the research papers that used ML algorithms, various performance evaluation metrics were used to assess the efficacy and accuracy of these computational models. These metrics provided a comprehensive understanding of the algorithm’s ability to accurately process and analyze data and served as critical indicators of its reliability and applicability in real-world scenarios.

  • The study conducted by Chalasani et al. in [29] used logistic regression to evaluate the performance of the algorithm in detecting early-stage hepatocellular carcinoma. The focus of the evaluation was on parameters such as sensitivity and specificity, which are crucial metrics in understanding the algorithm’s ability to correctly identify true positive cases (sensitivity) and true negative cases (specificity) [42]. The study also used Receiver Operating Characteristic (ROC) curves and Area Under the Curve (AUC) as graphical tools to assess the performance of the algorithm. [43]. The ROC curve provides a visual representation of the trade-off between sensitivity and specificity at different test thresholds, while the AUC quantifies the model’s overall performance. Analyzing the ROC curve provided an understanding of how well the blood test distinguishes patients with HCC from those without it. By comparing the AUC of the multitarget blood test with established methods such as alpha-fetoprotein (AFP), the study aimed to assess whether the new test offered a significant improvement. A higher AUC indicates better discrimination between true positives and false positives. Therefore, the parameters evaluated in the study demonstrated the diagnostic accuracy of the model and its potential to outperform biomarker-based tests for early-stage HCC detection.

  • In Döhner et al. [27] the Minimal Residual Disease (MRD) algorithm was used to assess the residual presence of cancer cells in the body after treatment. This algorithm was evaluated using quantitative polymerase chain reaction (qPCR) [44], a molecular technique used to quantify the presence of specific genetic material. In leukemia or other hematological disorders, qPCR was often used to measure the number of leukemic cells remaining after treatment. In this particular case, the authors consider qPCR to be \(\ge\) 2%, indicating that 2% served as a critical threshold. MRD levels above this threshold may have significant implications for patient assessment and management.

  • In the paper by Driver et al. [26], the authors used several machine learning techniques, including Random Forest, Gradient Boosting, and the Cox Proportional Hazards Model. The models were internally validated on the discovery cohort using cross-validation techniques, bootstrap resampling and leave-one-out cross-validation. Performance evaluation is likely to focus on predictive accuracy, feature importance scores from random forest and gradient boosting, and survival analysis parameters such as Cox hazard ratios. These parameters were critical in assessing the ability of the models to accurately predict meningioma grading and provide insightful predictions.

  • In the articles by McAndrews et al. [30],Nirmal et al. [28] and Timperi et al. [32], the authors employed the Uniform Manifold Approximation and Projection (UMAP) algorithm [41] for high-dimensional data visualization and analysis. While UMAP is primarily used for data exploration and pattern recognition, the effectiveness of the algorithm in these studies would be assessed by its ability to reveal meaningful biological insights and clusters within complex datasets. Therefore, evaluation parameters might include the coherence and separability of identified clusters or the algorithm’s success in uncovering known versus novel biological patterns.

  • Shouval et al. [25] used logistic regression to analyse the impact of TP53 alterations on therapy response. The key performance metrics here would include sensitivity, specificity, and possibly the AUC-ROC, which provides a quantitative measure of the model’s ability to discriminate between different patient response categories based on genomics alterations.

The specific focus and methodologies employed in the studies highlighted the nuanced approaches used to evaluate machine learning (ML) models in the context of biomedical research. These evaluation metrics reflected the diverse applications of ML algorithms, ranging from diagnostic tools to the prediction of therapeutic outcomes. This diversity underscores the importance of tailored evaluation strategies to accurately assess the performance and practical utility of each individual ML model.

4.7 RQ7: What challenges have been faced?

The implementation of ML algorithms in research presents a unique set of challenges that can significantly affect the results and interpretations of these studies. This section examined the challenges reported or inferred within the selected papers that used ML algorithms, providing insights into the common hurdles faced by researchers in this area.

  • Data quality and availability: a recurring challenge, highlighted in several studies such as those in [29] and [26], was the dependency on high-quality, comprehensive datasets. In medical research, the scarcity of large-scale, annotated datasets can limit the training and validation of ML models. Data heterogeneity, noise, and missing values can further complicate analysis and model performance.

  • Model complexity and interpretability: studies using sophisticated algorithms such as Gradient Boosting and Random Forest [26] face challenges related to model complexity and the interpretability of results. At the same time, these models can offer superior predictive power. However, their "black box" nature can hinder the understanding of model decisions, a critical aspect in medical contexts where explanatory power is crucial for clinical acceptance and decision-making.

  • Overfitting and generalization: ensuring that a model generalizes well to new, unseen data is a fundamental challenge in ML. There is an inherent risk of overfitting, especially when dealing with complex models and limited data, which could compromise the usefulness of the model in real-world applications. Studies need to implement robust validation techniques to mitigate this problem and confirm the generalizability of the models.

  • Algorithmic bias and fairness: The potential for algorithmic bias is a critical concern, particularly in healthcare applications where biases in the training data can lead to biased or unfair models. Studies such as those by Chalasani et al. [29] must carefully consider and address these challenges to ensure equitable and ethical application of their findings.

  • Computational resources: Advanced ML algorithms require significant computational resources, especially for large datasets or complex model architectures. Although not explicitly mentioned in the reviewed studies, the need for significant computational resources may be a limiting factor affecting the scalability and accessibility of ML-based research.

  • Integration with domain knowledge: The integration of ML with domain-specific expertise remains a key challenge, as highlighted in the studies by Nirmal et al. ([28] and Timperi et al. [32]. Effective collaboration between data scientists and domain experts is essential to guide the model development process, interpret results meaningfully, and ensure the algorithms align with domain-specific requirements and constraints.

  • Integration with domain knowledge: Integrating domain-specific knowledge into ML remains a crucial but challenging task, as demonstrated by studies such as [28] and [32]. Incorporating domain knowledge can significantly improve the interpretability and relevance of models in complex decision scenarios. According to Repetto et al. [45], domain knowledge not only improves the interpretability of deep learning models but also ensures compliance with regulatory standards. Effective collaboration between data scientists and domain experts is essential to guide the model development process, interpret results meaningfully and ensure that algorithms are aligned with domain-specific requirements and constraints.

  • Explainability of AI methods: The explainability of AI systems is now a legal requirement for high-risk applications, as stipulated by the European Union’s AI Act, which emphasizes the necessity for algorithms to ensure "sufficient transparency" for users[46]. This challenge is particularly significant as it relates not only to understanding the operational mechanics of models but also to providing meaningful insights that are understandable to non-experts, ensuring that the use of AI is responsible, ethical, and justifiable. This issue has been extensively discussed in recent studies, including the one by Fresz et al. [47], which highlights the difficulties in achieving truly transparent AI systems.

In summary, the challenges faced in applying ML algorithms are multifaceted, encompassing data-related issues, model complexity, ethical considerations, computational demands, and the critical need for accountability to meet evolving legal standards.

5 Discussion

The integration of Artificial Intelligence into cancer research fundamentally revolutionized the approach to this complex disease [48, 49]. AI’s ability to analyze large genomic datasets provided an unprecedented perspective on cancer biology, opening the door to novel diagnostic and therapeutic strategies and personalized therapies. Through advanced genomics analysis, AI-enabled more precise personalized medicine, leading to significant improvements in patient outcomes.

In the clinical context, advances in cancer genomics have been shaping oncology practice and patient care. Detailed analysis of clinical data, combined with machine learning, enabled the identification of subtle signals that might escape the human eye, with the potential for earlier diagnosis and improved treatment outcomes. In addition, AI is proving to be necessary for the systematic management of vast amounts of genomic and clinical data, capable of organizing, analyzing and interpreting the increasing volume of information in a quick and efficient manner.

The algorithms fueling the revolutionary advancements in cancer research employ a diverse array of methodologies. This reflects the variety and intricacy of the computational approaches being applied in this field. The ten articles selected, published during the pivotal years of 2022 and 2023, analyze different facets of this progress. They investigate the tumor microenvironment, propose new therapeutic perspectives, and explore various applications of these cutting-edge algorithms.

The analysis reveals that Uniform Manifold Approximation was the most commonly used algorithm across the ten selected articles from 2022–2023. This was followed in frequency by the use of Logistic Regression, Clustering techniques, and the Cox Proportional Hazards Model. In contrast, not a single one of the articles made any mention of Convolutional Neural Networks (CNNs), despite them being considered one of the most innovative and valuable algorithmic approaches in recent years in various domains. In fact, CNNs are proving to be extremely effective in analyzing complex genomic data, automatically analyzing DNA and RNA sequences, and identifying key features for a wide range of oncology applications. In addition, CNNs are also showing significant results in identifying mutations and genetic variants associated with cancer and, thanks to their machine learning capabilities, are demonstrating that they can accurately distinguish between somatic and germline mutations [50]. The final key application of CNNs is the classification of tumors based on their gene expression profiles [51]. This capability will increasingly allow doctors to classify tumors based on type, stage, and prognosis and tailor treatments to individual patients, maximizing efficacy and minimizing side effects.

The review does not emphasize the importance of gene regulatory networks (GRNs) in cancer research as well. GRNs represent a crucial interdisciplinary field that aims to understand the complex genetic and molecular interactions underlying the regulation of fundamental biological processes. GRNs are interpretable computational models of the regulation of gene expression in the form of networks [52] which are proving crucial for elucidating the molecular mechanisms underlying cancer development and progression and are essential for developing new therapies for the disease, designing new biological materials, and creating synthetic biological systems [53].

Understanding cancer complexity cannot be achieved without different types of networks; the abundance of multi-omics data and the advent of high-throughput technologies provide a unique opportunity to develop machine learning methods to study the underlying gene regulatory network [54]. Machine learning and artificial intelligence offer exciting possibilities for studying regulatory networks, but there are indeed challenges to consider, such as ensuring the accuracy and interpretability of models, addressing ethical concerns around data privacy and bias, and seamlessly integrating AI into clinical workflows.

Ethical challenges must be addressed with commitment and care to maximize the benefits of these revolutionary technologies. Protecting patient privacy, especially when it comes to genomic and medical information, is a major challenge when using AI in cancer research. Genetic and clinical data are highly sensitive and must be handled with the utmost care to avoid the risk of data breaches or misuse. Implementing rigorous data security policies and using anonymization and encryption measures are essential to protecting patient identity and confidentiality. It is equally important to ensure the transparency of AI algorithms used in cancer research. Understanding how algorithms make decisions is critical to assessing their reliability and consistency. Transparency builds confidence among users and healthcare professionals in the use of AI to inform clinical decisions.

Finally, addressing inequalities in access to advanced technologies is another critical issue to be mentioned. AI has the potential to revolutionize cancer research and treatment, but it is essential to ensure that all patients have equal access to these technologies. Socioeconomic and geographical inequalities may limit access to advanced AI-based health services, leading to health inequalities. It is important to develop policies and programs that reduce these disparities and promote equitable access to technological innovations in oncology.

6 Conclusion

The use of AI in cancer research is having a significant impact. It is opening up both exciting possibilities and complex challenges. AI is more than just a technological advancement in this field, as it is offering the hope of transformative changes in how cancer is diagnosed and treated. By enabling more precise diagnoses and personalized therapies, AI has the potential to greatly improve patient outcomes and the overall quality of cancer care. However, the integration of AI into cancer research and medicine is a complex process with various considerations and implications to navigate. The combination of available high-dimensional datasets advances in high-performance computing, and novel deep learning architectures has led to a significant increase in the application of AI in several areas of oncology research [7]. These include drug discovery, prediction of treatment response and resistance, and AI-enhanced diagnostics. Current research is making significant contributions to our understanding of cancer biology and treatment. Future research will focus on harnessing the strengths of AI to overcome current limitations. This includes developing models that can provide a more comprehensive and integrated understanding of cancer, and integrating not only large genomic datasets but also other omics data (such as transcriptomics, proteomics and metabolomics) alongside clinical information. Although current research is dedicated to creating strong data governance frameworks and reducing bias in AI models to ensure responsible and ethical application in cancer research and therapeutic settings, ethical considerations around data privacy and algorithmic bias need to be better addressed for responsible implementation in clinical settings. Future directions in AI-driven cancer research will increasingly involve addressing challenges related to data quality and bias, integration, interpretability and overall trustworthiness of these models.

Additionally, ongoing research efforts also focus on refining artificial intelligence models to improve the discover of new biomarkers. This is a fascinating and promising field in oncological research, but it presents several challenges, especially regarding their interpretability and clinical application. Identifying biomarkers involves discovering specific molecules that can indicate the presence or progression of cancer. However, cancer biology is extremely complex and variable among individuals. Each tumor can present different genetic and molecular alterations that influence its behavior. This heterogeneity makes it difficult to identify universal biomarkers that can be applied to all patients. Many biomarkers can be influenced by biological and technical variables, making it challenging to distinguish between significant signals and background noise. Furthermore, validating biomarkers requires thorough and reproducible studies, which can be costly and time-consuming. The interpretability issue with the AI models used to analyze biomarker data is also a major challenge. Many of the sophisticated AI models, like deep neural networks, operate as "black boxes" - their inner workings are opaque and difficult for clinicians to comprehend. This lack of transparency makes it hard for doctors to understand how these models arrive at their predictions. This limitation poses a significant barrier to the clinical adoption of such AI tools. Doctors need to have confidence in the decisions generated by these models and be able to explain the results to their patients. The black box nature of many advanced AI models undermines this requirement, hindering their integration into real-world medical practice. Achieving perfect integration of artificial intelligence technologies into clinical workflows and ensuring equitable access to AI-based tools will be essential to maximize the impact of artificial intelligence in cancer care.