The first biomarkers associated with the liquid biopsy concept were the CTCs, but their discovery is not state of the art. Since 1869, when Thomas Ashworth first reported the presence of a subpopulation of cells in the bloodstream of metastatic cancer patients, cells that shared similar features to the primary tumor where they detached from (62), numerous successive studies reported the existence of CTCs and their role in the metastatic cascade (63–65). However, since the number of CTCs is extremely low, the promotion of these circulating biomarkers as promising tools in oncological clinical practice was dependent on the development of novel analysis methods with superior sensitivity that emerged from the recent progress made in the biomedical field (66). The frequency of CTCs found in circulation is directly related to the tumor burden, but even in metastatic CRC (mCRC) patients, only 1-10 CTCs/mL of blood can be detected in peripheral blood samples (67). Considered initiators of the metastatic cascade, CTCs are cells of epithelial origin that shed into the blood and lymphatic systems, exit routes that allow them to colonize distant tissues and generate secondary cancer lesions. These cells depart from solid tumors as single intact cells or cell clusters, a feature that impacts their metastatic potential and half-life in circulation (68). CTCs can agglomerate and form homotypic cell clusters or can interact with macrophages, platelets, and/or stromal cells to form heterotypic clusters (68). Of note, in mCRC, CTCs originate both from primary and metastatic lesions, showing the effectiveness of these biomarkers to capture the global landscape of CRC tumor heterogeneity (69, 70). However, independent of the departure point, the capacity of CTCs to cause metastatic lesions is dependent on their stability in circulation, as most of CTCs die as a result of their inability to surpass physical and oxidative stress, anoikis, and absence of the tumor-specific microenvironment (71–73). Based on studies performed on preclinical models, it is estimated that of all detached CTCs, 0.1% remain viable in circulation, from which only 0.01 – 0.2% ultimately will grow into a metastatic lesion (74).

CTCs are characterized by the expression of the epithelial cell adhesion molecule (EpCAM) (75) and the cytokeratins (panCK) (76). However, there are serious concerns on using EpCAM only to capture and define CTCs as these cells undergo epithelial to mesenchymal transition (EMT) and down-regulate the expression of this marker (77, 78). Additional markers MUC-1 (79, 80) could be also relevant for targeting CRC cells. More, circulating cells of tumor origin undergo epithelial to mesenchymal transition (EMT) a process that was first identified in embryogenesis and that refers to epithelial cells reprogramming by acquiring a mesenchymal phenotype. EMT is involved in crucial physiological processes such as development and wound healing, but it is also involved in malignant progression. Upon activation of this reprogramming process, CTCs undergo a series of physical changes that enable their long-distance dissemination, invasion, and survival in the bloodstream (81).

Due to the rarity of CTCs in circulation, the identification, capture, and analysis of these cells are challenging issues. In this view, to improve the success of CTCs detection, the process is preceded by an enrichment technique to unmask these low-frequency cells from the myriad of blood cells: white blood cells (5-10 × 10⁶/mL), red blood cells (5-9 × 10⁹/mL), and platelets (2.5-4 × 10⁸/mL) (67). The enrichment strategies for CTCs detection take advantage of their specific physical and biological characteristics in terms of size, density, deformability, or specific cell surface markers (82). Among all the enrichment strategies explored, the most popular choice remains immunocapturing which can be performed either for CTCs collection (positive enrichment) or removal of blood cell subpopulations (negative enrichment).

The CellSearch^® platform is so far the only system that has been approved by the Food and Drug Administration (FDA) for CTCs identification, isolation, and enumeration in whole-blood samples collected from metastatic colorectal (83), breast (84, 85), and prostate cancer (86) patients. The technology uses a positive enrichment strategy based on selection and immunomagnetic capture of Epithelial Cell Adhesion molecule (EpCam) positive cells (87) through ferrofluidic nanoparticles functionalized with the EpCam antigen that select and capture a subpopulation of cells which are subsequently discriminated from residual hematopoietic cells by CD45 negative expression after an immunofluorescence assay. Using the CellSearch^® system, EpCam positive cells that are cytokeratin⁺/DAPI⁺/CD45^- are selected as CTCs and counted for cancer prognosis (88, 89). The platform was successfully employed in various studies that aimed to quantify CTCs in CRC patients and revealed the importance of these as independent prognostic markers (66, 90, 91). However, despite being the only system FDA approved, this platform presents several disadvantages such as low sensitivity and specificity and most important, the failure of detecting CTCs that lose their epithelial features as a result of EMT (92, 93). As a result of undergoing EMT, CTCs acquire a mesenchymal phenotype and thus lack specific epithelial markers such as EpCam and cytokeratins (94), pillar molecules for numerous CTCs enrichment and detection strategies, including the FDA-approved CellSearch^® System. From this point of view, Flow Cytometry which is extensively used in hematology could be successfully employed in CTCs detection and count as a versatile approach considering that the instruments work with full customizable protocols. Fluorescence-activated Cell Sorting (FACS) which is also a flow cytometry assay allows also the isolation of the CTCs sorted based on their specific markers expression. Other methods for CTCs isolation, divided by the enrichment strategy, are presented in Table 1 .

The low frequency of CTCs in blood severely impacts the clinical utility of these biomarkers for CRC screening and early detection as CTCs detection is almost impossible in early-stage cases (118). In this case, their detection has been a challenging issue that is reflected by the reduced number of studies available regarding the use of CTCs as CRC screening and early diagnosis tools. However, promising results for employing CTCs in screening protocols were shown by Yang et al. (119) that used a negative enrichment strategy for CTCs detection by fluorescence in situ hybridization in blood samples of patients with colorectal polyps and non-metastatic CRC. This study revealed the presence of CTCs in both benign and malign disease, with a significant increase of CTCs in the metastatic setup, where the CTCs count was modulated by the tumor anatomical location and degree of tissue differentiation. More, Tsai et al. employed the CellMax biomimetic platform to detect and enumerate CTCs in subjects with adenoma, polyps, or stage I-IV CRC and revealed an 88% accuracy of this approach for detection of CTCs in all tumor stages, including precancerous lesions (120). The CellMax platform’s unique design enables a superior detection method of CTCs in terms of sensitivity and purity and leads to recovery of viable CTCs that can be employed in downstream applications, making this approach appealing not only for early detection of CRC (121). Another approach to overcome the low frequency of CTCs for early CRC cancer could be repositioning diagnosis leukapheresis (DLA) as a CTCs enrichment tool based on the similar density these cells have with the mononuclear cells from peripheral blood (116, 122, 123), a method that has been successfully employed by Soya and colleagues for CRC patients (124).

Nevertheless, the most common clinical utility of CTCs remains cancer prognosis, a high count of CTCs being associated with a worse prognosis. A prospective study on preoperative patients with stage I-IV CRC showed that CTCs frequency, quantified using the CellSearch^® system, was increased in mCRC patients as compared with non-mCRC patients. For non-mCRC, CTCs detection represented a strong and independent prognostic marker and was associated with worse overall survival (OS) and progression-free survival (PFS) compared to CTCs negative patients (90). Using CTC for mCRC patients exclusively revealed that the quantification of CTCs represents an independent predictor of PFS and OS as both indicators were shorter for patients with ≥3 CTCs/7.5mL as compared with those of patients with <3 CTCs (83).

Regarding the clinical implications of CTCs analysis with impact on the post-operative therapy regimen setting in early and mCRC, Guadagni et al. (125) showed that CTCs from 79% of patient exhibited moderate to high sensitivity to mitomycin as compared to only 7% for irinotecan, 30% for oxaliplatin, 2.3% for 5-fuorouracil, 14% for raltitrexed, 4.6% for alkeran, and 4.6% for carboplatin. More, the same group used CTCs to select drug regimens and subsequent target therapy that was subsequent successfully administered by hepatic artery infusion to patients with colorectal cancer liver metastasis (126).

The limited clinical applicability of CTCs for CRC diagnosis and treatment personalization has led to a shift of liquid biopsy studies in recent years to employ preferentially cell-free circulating nucleic acids (cfNAs) in CRC patient management, cfNAs including cell-free circulating DNA (cfDNA) and cell-free circulating RNA (cfRNA). Indeed, analyzing cfNAs by molecular technologies can lead to real-time genomic profiling of the tumor, an extremely useful approach for the modern trends in CRC management that are directed towards precision medicine and patient-oriented treatment. As the tumor molecular profile dynamically changes in response to the immune system and administered drugs, it is mandatory to implement a novel analysis tool that can be used on demand for ensuring the appropriate treatment regimens for CRC patients (127).

cfDNA is comprised of multiple degraded fragments of DNA that are released in circulation by cell apoptosis, necrosis, or active release mechanisms (128, 129). Apoptosis generates short fragments of cfDNA (< 1000 bp) while following necrosis longer fragments are released (130). The presence of cfDNA in circulation is not exclusively associated with cancer, and this observation was made since its discovery in the blood plasma harvested from healthy subjects (131). However, in cancer patients, the levels of cfDNA are increased due to the supplementary cfDNA released by tumor cells. This small fraction (<1%) of cfDNA, known as cell-free circulating tumor DNA (ctDNA), originates from primary tumors, metastases, or CTCs and harbors specific tumor-related alterations that mirror the genomic status of CRC tumors and epigenetic alterations, thus unraveling the tumor molecular make-up (130, 132). More, cfDNA levels are directly dependent on the cancer stage, mCRC patients showing significantly increased levels of cfDNA in comparison with non-mCRC patients, highlighting the prospective use of cfDNA analysis for early detection of CRC (132–134). More, the ctDNA concentrations are correlated not only with CRC stage, but also with tumor size (135), and could be also employed as a prognostic tool since high levels of cfDNA reveal a shorter OS (136).

In early CRC, ctDNA can be used for investigating DNA methylation, a feature that is closely related to initiating and sustaining the transition of colon polyps to CRC (137). This alteration can be monitored through the FDA-approved Epi proColon^® assay, designed for identifying septin-9 (SEPT9) methylation in CRC patients. SEPT9 gene is a tumor suppressor gene that loses its role as a result of hypermethylation of the CpG island in the promoter region, a molecular aberration that promotes the development of CRC (138). The Epi proColon^® test aids the early detection of CRC and showed good sensitivity and specificity when compared with the other available non-invasive tests: stool-based guaiac fecal occult blood tests (FOBT) and fecal immunochemical tests (FIT) and blood-based CEA test (139, 140). A study conducted by Tóth and colleagues (141) that included both left- and right-sided CRC cases revealed that this assay is a reliable screening method in both cases, showing superior sensitivity and specificity than FOBT and CEA. Superior performance was observed also comparing SEPT9 assay with FIT in terms of sensitivity (73.3% vs. 68%), while the specificity was significantly better for FIT (97.4% vs. 81.5%). However, the combinatorically approach of using both tests improved these numbers to 88.7% sensitivity and 78.8% specificity (142).

At the moment, for ctDNA analysis, PCR-based and next-generation sequencing (NGS)-based approaches are explored. While PCR remains the backbone for ctDNA analysis, this approach is limited to CRC-specific driver mutations or requires prior knowledge of genetic alterations associated with the investigated CRC tumor. However, PCR-derived techniques are cost-affordable compared with NGS and allow the identification of therapeutically actionable targets or acquired mutations responsible for resistance to therapy. By qPCR-derived techniques, therapeutic vulnerabilities of CRC patients can be easily identified, allowing to discriminate between patients that are suitable for a particular treatment regimen or to readjust treatment according to acquired mutations. Currently, for mCRC patients, fluoropyrimidines remain the backbone of more intensive treatment strategies, including doublet (FOLFOX or FOLFIRI) or triplet combinations (FOLFOXIRI) with leucovorin, oxaliplatin (OXP), irinotecan (IRI), and capecitabine (CAPE) (143–145), added to target agents such as bevacizumab and/or anti-EGFR (146). With this respect, pivotal phase 3 randomized studyAVF2107g proves that the addition of bevacizumab to FOLFIRI significantly improved the overall survival. More, based on promising results obtained in phase 2 studies on the combination of bevacizumab with a FOLFOXIRI, a phase 3 randomized study (TRIBE) showed that this approach displays improved progression-free survival among patients with mCRC as compared with FOLFIRI and bevacizumab (hazard ratio for progression, 0.75; 95% CI, 0.62 to 0.90; P = 0.003) and and reached a median OS of 37.1 months in the RAS wild-type (147). More, the addition of cetuximab (CET) and panitumumab (PAN), which are anti-EGFR drugs, to FOLFOX or FOLFIRI chemotherapy in mCRC patients with KRAS exon 2 wild type and overexpression of EGFR led to improved results with respect to objective response rate, progression-free survival and OS (148). The main concern in the association of anti-EGFR drugs with triplet therapy regimens is the significant toxicity rate in therms of diarrhea, asthenia, mucositis and neutropenia. However, previous analysis of some pharmacogenomics markers such as 5-FUDR for 5-FU degradation rate or and single nucleotide polymorphisms (SNPs) in ABCB1, CYP3A4, DYPD, UGT1A1 which encode for enzymes responsible for 5-FU metabolism, are usefull to predict chemotherapy induced gastrointestinal toxicity (148). Consequently, a shift towards the administration of biological therapies is observed based on the success of these to improve the OS of mCRC patients, among which the best outcomes were observed following treatment with human vascular endothelial growth factor (VEGF) monoclonal antibody bevacizumab and the EGFR monoclonal antibodies cetuximab and panitumumab (149, 150). Independent of the therapeutic choice, the success of the therapy is based on the correct identification of the existing patients’ genetic abnormalities. In contrast to tissue biopsy, identifying genetic alteration by ctDNAs analysis can be employed on-demand establishing, therefore, a patient’s real-time treatment sensitivity pattern that helps avoid the administration of drugs that lack their antitumor effects and exert unnecessary toxicity. More, in absence of tumor tissue for CRC patient molecular characterization, blood can be used for ctDNA mutation screening, therefore providing a correct and efficient targeted therapy scheme (151).

Several genes alterations are screened for aiding CRC treatment, most usual for identifying anti-EGFR-resistant-CRC (152). EGFR is a transmembrane receptor tyrosine kinase with a crucial role in CRC development and progression, being, therefore, an important target for therapy by administration of EGFR monoclonal antibodies or tyrosine kinase inhibitors (153). Following administration of EGFR monoclonal antibodies, these bind to the extracellular domain of EGFR and prevent the activation of the receptor tyrosine kinase and multiple downstream signaling pathways related to cell survival, proliferation, and migration, such as RAS-RAF-MAPK, PI3K-PTEN-AKT, and JAK/STAT pathways (154). Half of the CRC patients harbor RAS mutation, a particularity that renders the anti-EGFR drugs ineffective (155). While KRAS-mutated patients are not exclusively associated with EGFR resistance, mutations in KRAS codons 12 and 13 at exon 2 are clear indicators of EGFR therapy resistance (156, 157), together with mutations in exons 3 and 4 that can indicate a poor response to anti-EGFR (149, 158–160). Another mutation that can overdrive the MAPK pathway, thus promoting cell survival and proliferation, is BRAF identified in 5% - 10% of CRC cases, the commonest being the BRAF^V600E mutation (>95%) (161). The presence of BRAF mutation is an indicator of poor prognosis in patients with mCRC and a marker of resistance to anti-EGFR therapy, and it was considered to be a signature of patients not harboring RAS mutations (154, 162, 163). However, recent studies showed the presence of concomitant KRAS and BRAF mutations (164–167). Other EGFR-resistant related mutations are identified with significantly low frequency than KRAS and BRAF mutation, making it difficult to unravel their role in anti-EGFR therapy resistance. The phosphatidylinositol-4,5-bisphosphate 3-kinase (PI3K)-Phosphatase and Tensin Homolog (PTEN) pathway is an oncogenic signaling network that is deregulated in CRC by aberrant activation of the prooncogenic gene PI3K or through loss of function of the tumor suppressor PTEN (43). PI3K mutations are identified in 10-20% mCRC in exon 9 or exon 20 and lead to constitutive activation of the p110a protein kinase and its downstream pathway, which promotes tumor cell proliferation and survival (168, 169). Both PI3K mutations and loss of PTEN function have been associated with EGFR blockade resistance (170, 171). More, although HER2 amplification was considered relevant for breast and gastric cancer, it appears that it represents also a negative predictor for anti-EGFR therapy (172). With a lower rate of incidence (2% - 4%), MET amplification can determine resistance to anti-EGFR drugs (173).

KIT proto-oncogene receptor tyrosine kinase and platelet-derived growth factor receptor (PDGFR) mutations play a key role in the prognosis of gastrointestinal stromal tumors (4), while ALK receptor tyrosine kinase fusion genes (174) and beta-catenin 1 (CTNNB1) mutations (175, 176) are crucial in inflammatory myofibroblastic tumors and desmoid fibromatosis, respectively. More, the key feature of the CRC occurring neuroendocrine neoplasms (NENs) classification is the discrimination between well-differentiated neuroendocrine tumors (NETs), the carcinoid tumors, and the poorly differentiated neuroendocrine carcinomas (NECs). Despite that NETs and NECs share the expression of neuroendocrine markers, they display significant genetic differences, and therefore, they are not considered to be akin neoplasms (177–180). The main genetic differences between them rely on the Retinoblastoma (RB) and Tumor Protein p53 (TP53) mutations that are commonly displayed by NECs. High-grade colorectal NECs are one of the most lethal types of neuroendocrine neoplasm (181). This poor prognosis pushes the exploration of more effective treatments. Late studies aimed to unreveal colorectal NECs genomic fingerprint. Considering the low incidence of colorectal NECs, only a few studies included highly significant amounts of patients. However, some noteworthy findings show that BRAF gene is frequently altered (182–186). Chen et al. (187) explored the genomic characteristics, and potentially targetable gene alterations in colorectal NEC and compared these characteristics with those of colorectal adenocarcinomas and gastrointestinal NETs using the American Association of Cancer Research (AACR) Project Genomics, Evidence, Neoplasia, Information, Exchange (GENIE) public NGS database (188). They discovered that the genetic characteristics of colorectal NEC are more similar to colorectal adenocarcinomas than gastrointestinal NETs and that the Wnt, MAPK, and PI3K signaling pathways as well as the cell cycle regulation are frequently aberrant in colorectal NECs. More, they confirmed that a significant group of colorectal NEC patients was bearing potentially targetable alterations such as BRAF^V600E (187).

There is no doubt that the identification of actionable mutations in mCRC has significantly contributed to the expansion of the therapeutic options available for these patients, but unfortunately very few of the drugs used are at the moment FDA-approved. However, besides the use of monoclonal antibodies to target EGFR, anti-angiogenic and anti-kinase drugs are also used for mCRC management, and studies have shown that these can significantly increase OS and PFS of mCRC patients. Generally, anti-angiogenic and anti-kinase agents are combinatorically administrated with fluoropyrimidine-based chemotherapy plus OXP and/or IRI (189, 190). Among these, the following are clinically approved and used in mCRC treatment regimens: bevacizumab (VEGF inhibitor), aflibercept (blocks activation of VEGF-A, VEGF-B, and PIGF), ramucirumab (anti-VEGFR-2 monoclonal antibody), and regorafenib (191) with approval and reimburse differently applied in the countries. This last one is a multikinase inhibitor that not only acts on angiogenic protein kinases (VEGFR-1, VEGFR-2, VEGFR-3, TIE-2), but as well on proteins involved in oncogenesis (KIT, RET, RAF-1, BRAF, BRAF^V600), metastasis (PDGFR, FGFR), and tumor immunity (CSF1R) (192, 193). Therefore, as in CRC cases mutations in genes encoding protein kinases are frequently identified, these drugs can be used to target and block the activity of these overexpressed proteins, improving, therefore, the OS and PFS of mCRC patients (194).

Based on the low content of ctDNA, its small fragment size, and limited high-life, qPCR-based methods were insufficient to detect mutations in ctDNA due to their limited sensitivity. Therefore, new techniques such as droplet digital PCR (ddPCR) or its high-throughput version BEAMing (beads, emulsion, amplification, and magnetics) showed superior efficiency in ctDNA analysis. The main advantage of ddPCR is the use of nanoliter-sized water-in-oil emulsion droplet technology to partition the cfDNA sample into numerous independent PCR sub-reaction (195), the target sequence being therefore concentrated between the aqueous droplets. In this way, after detection by fluorescence of the amplified targeted sequences and comparison of the signals obtained, rare mutations are discriminated from the wild-type sequences. ddPCR principle by which multiple PCR individual microreactors are generated, favors the quantification of low-abundance point mutations in cfDNA from a background of wild-type sequences, with high sensitivity ranging from 0.001% to 0.1% (173, 196).

In contrast, the modern NGS platform opens new opportunities for CRC patients by employing massively parallel sequencing techniques for genome-wide assessment and screen for multiple mutations with greater sensitivity without prior knowledge of the patient`s existing genetic abnormalities and can lead to the discovery of novel therapeutic targets (197). Undoubtedly, the main advantage of NGS analysis compared with the targeted approaches is the possibility of investigating a large range of markers reaching up to several hundred genes in a single panel. Also, a single test provides valuable information on single nucleotide polymorphisms, short indels, copy number variations, gene fusions, TMB (tumor mutational burden), or MSI (198). New NGS panels have been developed to detect the pathological variants associated with cancer at both cfDNA and cfRNA levels (e.g., Oncomine Pan-Cancer Cell-Free Assay - Thermo Fisher, Waltham, MA, USA). NGS approaches are based in fact on massively parallel sequencing technologies that offer high throughput, reproducibility, and speed. High-throughput sequencing, which includes Next-Generation “short-read” and “long-read” Sequencing methods allows scientists to perform a wide variety of applications in terms of genomics such as genome and exome de novo sequencing and resequencing, study of DNA-protein interactions, and epigenome characterization. Nowadays, there are at least 9 different NGS platforms using different template preparation and sequencing chemistry. However, all NGS platforms perform sequencing of millions of small fragments of DNA in parallel producing thousands or millions of sequences concurrently. Bioinformatics analyses are used to piece together these fragments by mapping the individual reads to a reference. NGS platforms allow deeply sequence target regions, analyze epigenetic factors such as genome-wide DNA methylation and DNA-protein interactions, and also sequencing cancer samples to study rare somatic variants or tumor subclones. Many methods are applying NGS to target panels for specific and highly sensitive detection of targeted ctDNA mutations (Tagged-Amplicon deep sequencing - TAm-seq; Safe-Sequencing System - Safe-SeqS or CAncer Personalized Profiling by deep sequencing - CAPP-Seq) (199). However, the use of NGS for detecting rare mutations provides a lower technical sensitivity than ddPCR (200), highlighting that PCR-based methods and NGS could be employed as complementary techniques for clinical applications.

Various commercial platforms are available for identifying actionable mutations in CRC such as Idylla or OncoBEAM, Idylla platform being used also for MSI detection (201–203). Vidal and colleagues (204) revealed a 93% overall agreement between the mutational status of RAS when comparing tissue and plasma samples. From 55 patients presenting RAS mutations in investigated tumor tissue, 53 had RAS mutational status confirmed by ctDNA analysis using OncoBEAM. Poor results were obtained by García-Foncillas et al. that showed just 89% agreement between results obtained by analyzing the mutational status of KRAS from tissue vs. plasma (201). For detecting MRD, Tie et al. (205) successfully used massively parallel sequencing-based assays to predict recurrence in patients with resected stage II CRC by analyzing ctDNA from postoperatively blood samples and can measure in real-time the effectiveness of treatment for patients receiving adjuvant chemotherapy.

Other valuable biomarkers in the liquid biopsy are ctRNAs such as micro-RNA (miRNA), messenger RNA (mRNA), and noncoding RNA (ncRNA) that are usually released in circulation within extracellular vesicles or associated with RNA-binding proteins (206). ncRNAs for example, are involved in all CRC stages of tumorigenesis and progression influencing key signaling pathways such as: WNT/-catenin, phosphoinositide-3-kinase (PI3K)/protein kinase B (Akt), epidermal growth factor receptor (EGFR), NOTCH1, mechanistic target of rapamycin (mTOR) and TP53 (207). Although an incomplete domain, in the last years many gene fusions have been identified in colorectal cancer only due to the possibility of high throughput NGS. The most frequent gene rearrangements detected in CRC with prognostic value and treatment perspectives are: NTRK fusions, ALK and ROS1 rearrangements, RET fusions and BRAF translocations. NTRK fusions were detected especially in women, elder people with RAS and BRAF wild-type genes and were presumably associated with primary resistance to EGFR-targeted treatment. Conventionally gene fusions are investigated by FISH, IHC or RT-PCR. Unfortunately, these methods can test for only one fusion gene at a time, being low-throughput and sometimes costly. The great advantage NGS brings in clinics is the large amount of information obtained. In just one test hundreds of gene rearrangements, known and unknown, frequent or rare can be detected. It can also investigate very small rearrangements that would be impossible to be evaluated using classical methods (208).

Nowadays we are all surrounded and widely interacting in our daily lives with AI in the form of consumer technology, whether we think about smartphones, wearable devices, search engines, social media channels, personalized advertising, facial recognition, autonomous vehicles, energy-efficient buildings, smart toys and many more. The concept of AI is not new (209), but only in the past few years, it has raised interest due to its potential as a virtual assistant able to serve various domains of human activity, especially the biomedical field. These AI experiences aim to improve human lives by increasing efficiency and by tailoring solutions for each individual. From this point of view, AI may be defined as a combination of theories, algorithms, and computing frameworks enabling a machine to independently reproduce intellectual processes associated with human cognition to decide on an action in response to its perceived environment to achieve a predetermined goal. Consequently, AI empowers humans to become faster in analyzing big amounts of data and smarter in decision-making by augmenting, however without replacement, human intelligence, and intuition (210). At its core, AI is a branch of computer science that has gradually changed not only our daily lives but also the landscape of healthcare and biomedical research for medical image analysis, intraoperative imaging, and genomics.

Starting with the ‘50 when the concept was first defined, there have been developed rule-based clinical decision support systems (211) that helped physicians to diagnose disease (212, 213) or decide the best treatment (214). Despite all these advantages and clear progress in the current practice, the rule-based systems displayed a series of inconveniences such as high building cost and system limitation by the difficult encoding of complex interactions and requirement of solid medical knowledge (215). Therefore, this first generation of AI systems has been improved by the integration of machine learning (ML) methods , which apply complex interactions to identify data patterns. ML is a subfield of AI that uses both mathematical and statistical approaches to improve the performance of computers. More specifically, ML resides in the development and implementation of dynamic algorithms that are able to take particular actions in response to particular inputs (environment), analyze the data to determine the actions (216, 217). These algorithms are able to self-improve (learn) as more data is available and the optimization process is called “training”. Depending on the tasks to be solved, basic ML algorithms can be divided into supervised, unsupervised (218), and reinforcement learning.

Supervised learning (SL) algorithms such as Naive Bayes classification (217), linear and logistic regression (219), support vector machines (SVMs) (216), or random forests (RF) (220) analyses the training data (input-output pairs) and map an input to output (221). A good example, in this case, is the histopathology grading of digitalized tissue biopsy slides. First, a set of images are labeled by the pathologists as positive or negative for a specific type of cancer and then SVMs for example learn to classify new, unlabeled images and tag them on a “probability map”. Zhi et al. (222) used five data sets containing metastatic and non-metastatic CRC samples to identify potential biomarkers for CRC metastasis. For this, the authors used a meta-analysis method to identify the metastatic and non-metastatic CRC differentially expressed genes and then, used the SVM classifier to select the top candidates having as reference a CRC dataset from The Cancer Genome Atlas database. The authors report a 100% precision of the SVM-classified 40 gene signatures and highlight the CREB1, CUL7, and SSR3 genes as biomarkers for the prognosis of metastatic CRC.

Unsupervised learning (UL) algorithms such as k-means clustering (217), principal component analysis (223), and autoencoders (224) identify patterns from untagged data by separating the items into different classes based on the training data features to find sub-clusters and outliers in the data. Bae et al. (225) proposed a feature selection method able to distinguish CRC patients from normal individuals using K-means clustering and the modified harmony search algorithm. To classify CRC using gene information the authors analyzed 6500 genes in 40 CRC tissue biopsies and 22 normal colonic tissue samples by a 4 step hybrid method consisting of: (i) Z-normalization of gene information values, (ii) Fisher score based reduction of redundant genes, (iii) K-means clustering of representative genes and (iv) Harmony Search (HS) algorithm based selection.

Reinforcement learning (RL), consists of a set of algorithms (agents) that operate sequentially by predicting the features of each step based on the past and present features while assigning a reward or a penalty on the prediction basis. The purpose of RL is for the agent to learn an optimal or nearly-optimal policy that maximizes the “reward function” (226).

Deep learning (DL) is a subfield of ML which involves training an artificial neural network (ANN) with many layers. DL recapitulates the biological neural network of the human brain and uses a layered structure of algorithms to analyze data, identify patterns, draw conclusions, and make decisions. The basic architecture of deep neural networks (DNNs) consists of an input and an output layer together with a variable number of hidden layers in between. In such a network, the input is provided to the input layer, which transfers its computed value towards the hidden layers that are finally linked to an output layer. DL can be further classified into the deep neural network (DNN), recurrent neural network (RNN), and convolutional neural network (CNN). CNNs are particularly important to identify patterns from unprocessed images (227). CNNs apply nonlinear transformations to structured data, for example the raw pixels within an image, to automatically learn relevant features. In this approach, it is mandatory to process accurately the images before analysis in order to reduce the risk of the model to learn from artifacts. CNNs basically uses two main models: the first one uses images from a large collection such as ImageNet to train the initial layers, and the second one is based on an auto encoder where the model learns background features from a subset of representative images (228). DL has been widely applied to empower medical procedures mainly involving analysis of images resulting from a wide range of procedures (229–232) and genomics (233).

In brief, biomedical applications of AI can be divided into two main branches: virtual and physical. The virtual component consists of ML algorithms: supervised learning, unsupervised learning, and reinforcement learning, and DL as an ML subset ( Figure 2 ). CNN is the most prominent DL construction, representing a particular type of multilayer artificial neural network that is highly efficient for image classification (227). While the virtual branch of AI-powered biomedical applications refers to image analysis and genomic (big data) analysis, the physical component includes medical devices, surgery robots, and devices for automated therapy formulation and nanorobots for targeted drug delivery (234).

Virtual reality (VR) is a simulated experience that generates an immersive, artificial image and/or environment with real-time interaction. Applications of VR include entertainment (video games), education (virtual training in a wide range of fields including medicine), and even business (virtual meetings). In medicine, one of the earliest VR platforms developed was Minimally Invasive Surgical Trainer-Virtual Reality (MIST-VR) (Mentice Medical Simulation, Gothenburg, Sweden) for endoscopic training (235). VR 3D technology is being used for simulation-based training due to its major advantage of creating near true experiences overcoming the disadvantages of the current distance learning and 2D videos (236). Currently, using VR (237), previously segmented organs and structures can be highlighted by streaming a digital full image, often visualized as a 3D render in a virtual environment.

Augmented reality (AR) is an interactive experience consisting of the addition of artificial information to the objects in the real world through computer-generated perceptual information that enables the user to perform tasks easily. This can be achieved across multiple sensory modalities like overlapping images and generating video or computer models. The first headset equipped with an overlapping display was introduced by Ivan Sutherland in 1965 and served for military purposes (238). Some examples with applications in the biomedical field are AccuVein (AccuVein Inc., NY, USA), a projector-like device that displays a map of the vasculature on the skin surface (239), or Google Glass which is a head-mounted display with generated objects superimposed onto real-time images. Unlike the virtual environment provided with VR, AR has been used in real surgery for many years (240) as it can overlay “invisible” information such as pre-or intraoperatively obtained imaging findings in the real environment.

CRC is a highly preventable disease with very good survival rates when diagnosed in the early stages. This is why routine screening is a crucial step in lowering the incidence rates of this pathology (241) while offering patients the hope for disease-free. The premalignant alterations from normal to malignant lesions take almost over 15 years. Efficient screening methods are permanently developed to accurately diagnose abnormal modifications with potential pathologic signification (242, 243). Available CRC screening procedures include colonoscopy, capsule endoscopy, imaging examinations, and/or blood and stool tests (244, 245). ML algorithms may be used as non-invasive and cost-effective methods to screen the CRC risk in large populations using personal health data (246).

Along with chemotherapy administration, surgery remains the main curative procedure for CRC management. With the development of robotic surgery as a minimally invasive procedure, CRC management enters a new era. To date, the da Vinci System developed based on the physical branch of AI technologies, is the most used surgical robot worldwide. The major advantages of its use reside both on the surgeon’s and patient’s sites. Robot-assisted surgery needs smaller incisions that will conduct minimal scarring and a significant decrease in the risk of surgical site infections, faster recovery and decrease of hospitalization period. Postoperative pain and bleeding are significantly lower as compared to traditional open surgery (312, 313). On the other side, using computer-controlled devices surgeons benefit from enhanced visual field, more flexibility, increased precision, and minimal fatigue. Besides the above-mentioned advantages, the da Vinci dual-console also allows integrated teaching and supervising. Several retrospective studies on robotic-assisted CRC resections showed that the procedure significantly reduced the complications rate and resulted in less evident inflammatory response as compared with the open surgery (314, 315). Other robots for surgery such as The Senhance System (TransEnterix Surgical Inc., Morrisville, NC, USA) which is a laparoscopy-based system are developed. Overall, robot-assisted colorectal surgery has better performance in terms of both short- and long-term outcomes (316).

Concerning the actual preparation of the administered therapy for patients, manual handling of cytotoxic drugs is a high-risk activity due to the prolonged exposure to carcinogens (317) and also a major error generating point in the patient’s management (318). Therefore, the introduction of robots such as the APOTECAchemo automated system (319) improved the safety of cytotoxic drug preparation, almost excluding the errors associated with human handling (320). Regarding CRC therapy research, Martel et al. (321) developed the NamiRobot System that can deliver drugs to cancer cells and hypoxic regions based on the reduced oxygen levels caused by the proliferation of cancer cells. More, AI technology can also promote research on new drugs. For example, Cruz et al. (322) used machine learning to detect with a prediction accuracy of over 63% the half-maximal inhibitory concentration (IC50) of a new drug on the HCT116 colon cancer cell line.