Prognostic and Predictive Factors in Advanced Urothelial Carcinoma Treated with Immune Checkpoint Inhibitors: A Review of the Current Evidence

Simple Summary Despite immune checkpoint inhibitors’ (ICIs) improved overall survival in urothelial carcinoma patients, only a minority of them benefit from immunotherapy. Therefore, there is an unmet clinical need to identify biomarkers which are useful to select the patients who are most likely to respond to ICIs. This review describes the prognostic and predictive role, and potential clinical applicability, of patient- and tumour-related factors. These factors include new molecular classes, tumour mutational burden, mutational signatures, circulating tumour DNA, programmed death-ligand 1, inflammatory indices and clinical characteristics. This summary may help clinicians to assess patients who are considered for ICI treatment, and may drive further prospective research on these biomarkers. Abstract In recent years, the treatment landscape of urothelial carcinoma has significantly changed due to the introduction of immune checkpoint inhibitors (ICIs), which are the standard of care for second-line treatment and first-line platinum-ineligible patients with advanced disease. Despite the overall survival improvement, only a minority of patients benefit from this immunotherapy. Therefore, there is an unmet need to identify prognostic and predictive biomarkers or models to select patients who will benefit from ICIs, especially in view of novel therapeutic agents. This review describes the prognostic and predictive role, and clinical readiness, of clinical and tumour factors, including new molecular classes, tumour mutational burden, mutational signatures, circulating tumour DNA, programmed death-ligand 1, inflammatory indices and clinical characteristics for patients with urothelial cancer treated with ICIs. A classification of these factors according to the levels of evidence and grades of recommendation currently indicates both a prognostic and predictive value for ctDNA and a prognostic relevance only for concomitant medications and patients’ characteristics.


Introduction
Worldwide, urothelial carcinoma (UC) represents the seventh most common cancer and the ninth most deadly tumour, with about 212,000 related deaths [1].
For a long time, the only effective treatment of metastatic UC (mUC) was platinumbased chemotherapy, which is still the standard of care in the first-line setting [2]. In recent years, the treatment landscape of mUC has been changed profoundly by the introduction of immune-checkpoint inhibitors (ICIs) [3] and, even more recently, antibody-drug conjugates (anti-nectin 4 and anti-Trop2) [4] and FGFR inhibitors [5].
The introduction of these new drugs urges the identification of potential biomarkers which are able to select the patients most likely to respond to immunotherapy. Promising prognostic and predictive factors in patients with mUC treated with ICIs include clinical features, new tumour molecular classes, the tumour mutational burden (TMB), mutational signatures, circulating tumour DNA (ctDNA) and programmed death-ligand 1 (PD-L1). Despite the increasing number of biomarkers under investigation, these factors still need validation for their application in clinical practice.
In this review, we summarize the landscape of clinical, molecular and genomic determinants of the prognosis and response to ICIs in patients with mUC. A classification of these factors by their prognostic and predictive value according to levels of evidence and grades of recommendation [7] based on the available evidence has been attempted.

Tumour Mutational Burden (TMB)
TMB can be defined as the total number of non-synonymous mutations per coding area of a tumour genome [18]. These mutations can be transcribed and translated to generate neoantigens displayed on the cell surface; T-lymphocytes can recognise some of these neoantigens and promote the apoptosis of tumour cells [18]. Tumours with a high mutational load are more likely to express neoantigens, and to induce a strong immune reaction [19]. Several studies have demonstrated an association between a high TMB and the response to immunotherapy in different locally advanced or metastatic solid tumour types [20].
In June 2020, the FDA approved pembrolizumab for the treatment of adult and paediatric patients with unresectable or metastatic solid cancers and a high TMB [>10 mutations/ megabase (mut/Mb)] [21]. This approval was based on the efficacy data from the phase II KEYNOTE-158 trial, which demonstrated an association between a high TMB and the tumour response to immunotherapy, with a durable response (>2 years) rarely being observed in heavily pre-treated metastatic cancers [22]. UC, along with melanoma and lung cancer, is characterized by a high number of somatic mutations and, therefore, a significant genomic instability [8]. The exploratory analyses of the phase II IMvigor210 trial included the quantification of the mutational load and a correlation with the clinical outcomes [23]. As expected, the TMB was significantly higher in patients who responded to atezolizumab than in non-responders. Moreover, patients with a higher mutational load had significantly longer overall survival than patients with a lower load [23].
However, the TMB status alone was not able to stratify the patients according to the survival benefit. In this regard, exploratory analyses of the JAVELIN Bladder 100 trial demonstrated that some DNA mutational signatures, involving certain base pair alterations, were associated with a survival benefit from avelumab, while other mutations were not [24]. This suggests that the type and the location of mutations may influence the predictive role of TMB assessment more accurately than the degree of the mutational load.
The predictive value of TMB for the pathological response to immunotherapy has also been explored in the neoadjuvant setting [25,26]. In patients with MIBC, a high TMB was generally observed in responders versus non-responders to neoadjuvant pembrolizumab, irrespective of the histological subtypes [27].
The predictive role of combining TMB with PD-L1 was also investigated. Patients with high TMB and PD-L1-positive tumours were more likely to derive a survival benefit from avelumab maintenance therapy [24]. Similar findings were observed in the randomised phase III IMvigor130 study [28]. The association of high PD-L-1 (>5% of immune cells) and a high TMB (>10 mut/MB) was associated with improved OS in the atezolizumab monotherapy arm, as compared to the chemotherapy arm [28]. Nevertheless, similar outcomes were not seen with the combination of atezolizumab and chemotherapy compared to the chemotherapy alone, suggesting a potentially distinct biology driving the benefit from atezolizumab and its combination with chemotherapy, and eventually highlighting the predictive inconsistency of this biomarker combination [28].
As for other tumour types, many factors hinder the clinical application of TMB as a biomarker, including the variability and the lack of a validated cut-off, a clear prognostic value, the differences related to the sequencing platforms used for its assessment, and the high scoring failure rate due to the quantity and quality of tumour tissue analysed [29]. As an example for the mUC, in the IMvigor210 study, the cut-off of the TMB varied between the two different cohorts of cisplatin-ineligible mUC treated with first-line atezolizumab and the platinum-treated patients [23]. Moreover, a high tumour mutational load may not be a specific biomarker for immunotherapy, as it has also been associated with the tumour response to neoadjuvant chemotherapy [30,31].

Mulecular Signatures
The increasing interest in tumour molecular features and new omics technologies has led to the discovery of different molecular signatures. These involve genes, messenger ribonucleic acids (mRNAs) and proteins which are studied as biomarkers to better predict clinical outcomes, or to better understand the cancerogenesis process [32,33].
Furthermore, because the TMB and/or the PD-L1 expression did not precisely identify patients who were more likely to derive benefit from immunotherapy [34,35], the assessment of tumour-related or immune-related gene signatures was actively investigated.
The DNA and RNA sequencing analyses of the IMvigor130 trial investigated the apolipoprotein B editing catalytic polypeptide (APOBEC) signature [28]. The APOBEC enzymes are a family of cytidine deaminases involved in the DNA repair processes, and are responsible for a mutation signature (TCW > T/G) frequently observed in MIBCs [36,37]. Patients with a high APOBEC mutational signature had a longer survival, whether in the atezolizumab monotherapy or the combination arm with chemotherapy compared to chemotherapy alone, whereas a high TGFβ signature predicted a worse OS with the immunotherapy alone [28].
RNA-based immune-gene expression profiling has the advantage of providing information from many cancer cells and immune cells, identifying more accurately the inflammatory status. The Checkmate-275 study investigated an interferon-gamma (IFN-γ) expression signature, and found a significant correlation with the response to nivolumab in the metastatic setting [38].
At the ESMO 2020 meeting, an exploratory analysis of the Javelin Bladder 100 study on tumour biomarkers was presented. While neither PD-L1 and TMB alone nor in combination predicted the response to immunotherapy and survival benefit, the expression of the immune-related genes of both the innate and adaptive immune system (i.e., CD8, IFNG, LAG3, TIGIT and CXCL9) and the number of alleles encoding high-affinity Fc gamma receptor variants predicted the survival benefit from avelumab first-line maintenance [24]. The application of the T-cell-inflamed and JAVELIN-Immuno signatures seemed to correlate with the treatment response with HRs of 0.55 and 0.49, respectively. Furthermore, the JAVELIN-Immuno high signature showed enrichment in some signalling pathways (Notch, Hedgehog, TGFbeta) which were likely to have an increased antitumour response to immunotherapy [24].
In the PURE-01 study, investigating the efficacy of pembrolizumab as a neoadjuvant therapy, a T-cell-inflamed signature was able to identify those patients who achieved pathological T0 downstaging. In a subsequent analysis, other RNA-based immune signatures were evaluated for their association with pCR and a high Immune190 signature, as well as hallmark signatures for interferon gamma and interferon-alpha, which were significantly associated with pCR and progression-free survival (PFS) after pembrolizumab in the PURE01 study [27]. Similarly, a correlation between the tumour response to atezolizumab and the transcriptional signature of eight genes (IFNG, CXCL9, CD8A, GZMA, GZMB, CXCL10, PRF1 and TBX21) representing interferon signalling, and the presence of CD8+ effector T cells, namely the tGE8, was reported within the ABACUS neoadjuvant study [39].
Moreover, a strong immune-mediated adaptive resistance was observed in nonresponding tumours, suggesting the investigation of ICI combinations to counteract the expression of negative regulators of the immune response [40].
The BISCAY study combined durvalumab (anti-PD-L1) with different targeted therapies depending on tumour gene alterations determined by NGS [41]. Despite the overall negative results, this study suggested a negative predictive role for the FGFR expression with regards to immunotherapy. In fact, FGFR-mutated tumours did not have a high expression of immune-active T-cell signatures, and the addition of durvalumab did not result in enhanced activity from the targeting agent only [9,41].
Although the above data look promising, there is a need to standardize molecular assays and find a molecular panel which is applicable to daily clinical practice.

ctDNA
The sequencing of the tumour fraction of the cell-free DNA (ctDNA) is an emerging and sensitive method to detect residual disease, anticipate relapse, and monitor the therapeutic efficacy in patients with several cancer types [42,43]. Furthermore, it provides a basis for clinical studies evaluating early therapeutic interventions [42,43].
In UCs, proof-of-concept data documented that ctDNA is detectable in plasma and urine, and could be a prognostic factor [44,45].
Vandekerkhove et al. demonstrated that ctDNA profiling may also identify putative biomarkers for therapy response in bladder cancer (such as FGFR3, ERCC2, ERBB2 and TMB), and represents a cost-effective and minimally invasive method for their identification and patient stratification [46]. Furthermore, the serial monitoring of ctDNA can optimize the use of therapies, as ctDNA fractions are expected to decrease in patients responding to treatment [46].
In a longitudinal analysis, the presence of ctDNA-assessed by whole-exome sequencing (WES) in 68 patients with localised advanced bladder cancer at diagnosis-during chemotherapy, before and after cystectomy and during surveillance, resulted in being Cancers 2021, 13, 5517 6 of 17 highly prognostic at diagnosis (HR 29.1) [47]. The presence of ctDNA identified all of the patients with metastatic relapse (with 100% sensitivity and 98% specificity), with a median of 96 days of diagnostic anticipation compared to the radiographic imaging [47]. Furthermore, the dynamics of ctDNA during chemotherapy, for those patients with positive ctDNA before or during the treatment, were associated with disease recurrence. It is worth noting that pathological downstaging, including related mutational signatures, was not associated with disease recurrence [47].
A post-surgical ctDNA detection was also associated with a higher risk for recurrence and death in the randomised phase III IMvigor010 study with adjuvant atezolizumab versus observation following cystectomy for patients with MIBC [48]. Moreover, the presence of ctDNA was predictive, as it was able to identify the patients who were likely to benefit from adjuvant atezolizumab. The IMvigor010 study did not meet his primary endpoint of disease-free survival (DFS) in the overall population of 809 enrolled patients [48]. The ctDNA was assessed by WES at a median of 11 weeks post-cystectomy in 581 patients with evaluable samples, comprising 72% of the intent-to-treat population of the IMvigor010 study. While the detection of ctDNA (in 214 patients, 37%) was associated with worse DFS and OS compared to the absence of ctDNA (in 367 patients, 63%), a significant difference in DFS (HR 0.58, 95% CI 0.43-0.79) and OS (0.59, 95% CI 0.41-0.86, p = 0.0059) was observed in favour of the patients with detectable ctDNA treated with atezolizumab compared to the observational arm. Furthermore, in patients with the presence of ctDNA, the rate of ctDNA clearance from cycle 1 to cycle 3 was significantly higher in the atezolizumab arm versus the observational one, with ctDNA clearance rates of 18.2% and 3.8%, respectively (p = 0.0048) [48].

Programmed Death Ligand-1 (PD-L1)
In the first-line treatment of mUC, a high PD-L1 combined positive score (CPS) of ≥10%, defined as the percentage of tumour cells (TC) and immune cells (IC)/the total tumour cells given by the 22C3 Dako assay, was associated with a prolonged median OS (mOS) in patients ineligible for cisplatin treatment with first-line pembrolizumab in the single-arm phase II KEYNOTE 052 study. The mOS was 18.3 versus 9 months in patients with PD-L1 CPS ≥ 10% and <10%, respectively, and it was 11.3 months in the overall population [49]. As the KEYNOTE 052 was not a controlled phase III study, a favourable prognostic value only for the high PD-L1 expression could not be rule out. Indeed, in the following KEYNOTE 361, IMvigor130 and Danube phase III randomised trials, the anti-PD1/PD-L1 agents either alone or in combination with chemotherapy were not superior to the chemotherapy according to OS in patients with high PD-L1 tumours, as well as in the overall population [50][51][52]. Only in the Danube trial did the combination of the anti-PD-L1 inhibitor durvalumab and anti-CTLA4 tremelimumab show a superior OS compared to chemotherapy in patients with high PD-L1. Therefore, the predictive value of high PD-L1 remains uncertain [52]. In contrast, a decreased survival with first-line pembrolizumab or atezolizumab compared to platinum-based chemotherapy was reported by the KEYNOTE 361 and IMvigor130 phase III trials for patients with low PD-L1 tumour expression, suggesting a negative predictive value for low PD-L1 expression [50,51]. A low PD-L1 expression was defined in the KEYNOTE 361 and IMvigor130 studies as a CPS < 10% or positive IC < 5% by the immunohistochemistry (IHC) 22C3 pharmDx and SP142 Ventana assays, respectively [50,51].
In the maintenance setting of mUC, following first-line chemotherapy, the Javelin-100 phase III trial showed an advantage in OS from the anti-PD-L1 avelumab compared to the best supportive care irrespective of the positive PD-L1 expression, which was assessed by a SP263 Ventana assay and classified as positive if at least one of the following three criteria were met: at least 25% TC staining for PD-L1, at least 25% IC staining for PD-L1 if more than 1% of the tumour area contained IC, or 100% ICs staining for PD-L1 if no more than 1% of the tumour area contained ICs [53].
In the adjuvant setting of UC, the IMvigor 010 with the anti-PD-L1 atezolizumab did not show a significant advantage in DFS compared to observation either in patients with positive PD-L1 tumours IC according to the SP142 Ventana assay, or in the overall population [54], while the Checkmate274 showed a significant DFS advantage in favour of nivolumab versus observation in the intention-to-treat population with a better HR in patients with a PD-L1 expression ≥ 1% (HR 0.55 vs. 0.70 for the overall population), as assessed on TC by a PD-L1 IHC 28-8 pharmDx assay [55].
In the second-line of mUC, in the IMvigor 211 phase III study with anti-PD-L1 atezolizumab compared to chemotherapy, the OS was not significantly different in patients with a high PD-L1 defined by IC 2/3 (or ≥5%/10% of PD-L1 positive IC according to a SP142 Ventana assay) [56]. In contrast, in the KEYNOTE 045 study, a significant difference in OS was observed in favour of the anti-PD-1 pembrolizumab versus chemotherapy in patients with high PD-L1 tumours defined as CPS according the 22C3 Dako assay [57]. Although possible differences in efficacy between the anti-PD1 and anti-PD-L1 agents cannot completely be ruled out [58], plausible explanations may rely on the different tests and scores used to assess the PD-L1 expression, as well as on the tumoural site and timing of the biopsy in respect to the disease history. In this regard, it is noteworthy that the mOS duration in the high PD-L1 subgroup of patients differed between those two trials despite the similar populations enrolled and the mOS observed in the overall population of the chemotherapy control arm (of 8.4 and 7.3 months, for the IMvigor211 and the Keynote-045 study, respectively) [56,57]. In the KEYNOTE 045 study, the mOS of patients with high CPS (≥10%) treated with pembrolizumab was of 8.0 months, which was significantly longer than 5.2 months with chemotherapy (p = 0.0048) [57]; on the other hand, in the IMvigor211 study, the patients with high IC 2/3 had an mOS of 11.1 months with atezolizumab, which was not significantly longer than 10.6 months with chemotherapy (p = 0.41) [56]. The OS difference seen in the chemotherapy control arms of these two studies, with a different definition of high PD-L1 tumours, might suggest a positive prognostic value which is marginally predictive for the high PD-L1 by the IC 2/3, and a negative prognostic-although predictive-value by the CPS.
In the neoadjuvant setting, only phase II studies are currently available. However, the Pure-01 study demonstrated the dynamic characteristic of the PD-L1 by showing its significantly increased expression as CPS following three administrations of the anti-PD1 pembrolizumab [40].
In conclusion, the current evidence on the prognostic and predictive value of PD-L1 expression is limited by the different diagnostic assays used in the clinical trials for each anti-PD1 or anti-PD-L1 agent, and remains inconclusive. PD-L1 may have a prognostic role, the value of which may also change depending on the type of tumour cells scored (i.e., negative if CPS, or positive if IC), and may be predictive only for the low expression.
The molecular predictive and prognostic factors in UCs are summarized in Table 1. Ba/Sq is associated with shorter OS, NE-like is associated with worse prognosis (LumP as reference) [9]. LUmNS, LumU, NE-like are more responsive to ICIs [9].
High TMB predicts OS benefit from avelumab maintenance therapy [23] and improved OS with atezolizumab when compared to CT [28]. TMB was higher in responder to neoadjuvant pembrolizumab [27].
Issues: variability, lack of a validated cut-off, differences related to the sequencing platforms
Need to standardize molecular assays and find a molecular panel applicable to daily clinical practice ctDNA Quantitative and qualitative analysis of circulating tumoral DNA detected on blood samples.
ctDNA detection is associated with worse prognosis in early stages and could identify metastatic relapses before imaging [47]. ctDNA detection is predictive for adjuvant atezolizumab (both DFS and OS) [48]. ctDNA profiling may be predictive for response to specific therapies [46].

PD-L1
Expression of the ligand of PD1 receptor has been widely studied as predictive biomarker of response to anti-PD1 and anti-PD-L1 therapies across human cancers.
Issues: different diagnostic assays used in clinical trials for each anti-PD1 or anti-PD-L1 agent, discordant efficacy of these agents between studies Abbreviations: UC, urothelial carcinoma; ICIs, immune checkpoint inhibitors; TMB, tumor mutational burden; OS, overall survival; CT, chemotherapy; ctDNA, circulating tumor DNA.

Patient's Characteristics
Several clinical factors have been reported to be prognostic in patients with mUC treated with chemotherapy. The most important ones include performance status (PS), metastatic sites, haemoglobin levels, and time from prior chemotherapy. Bajorin et al. showed that, in untreated patients, a poor Karnofsky Performance Status (KPS) < 80% and the presence of visceral (liver, lung, and bone) metastases were independent prognostic factors associated with a worse OS [59]. Bellmunt et al. reported that a poor Eastern Cooperative Oncology Group (ECOG) PS ≥ 2, low haemoglobin levels (<10 g/dL) and the presence of liver metastasis were independent poor prognostic factors in patients who failed platinum-based chemotherapy [60]. Moreover, a shorter time from prior chemotherapy Cancers 2021, 13, 5517 9 of 17 (<3 months) has been reported to enhance the prognostic classification of patients receiving second-line therapy [61].

Concomitant Medications
Several retrospective analyses and meta-analyses have recently been reported in the literature about the impact of concomitant medications on the clinical outcomes of patients with cancer treated with ICIs [73][74][75][76][77][78]. These studies mainly included patients with nonsmall cell lung cancer (NSCLC), melanoma or renal cell carcinoma (RCC). Commonly used drugs in clinical practice-such as corticosteroids, antibiotics and proton pump inhibitors (PPIs)-have been reported to negatively affect the activity of immunotherapy through immune-modulatory effects. In fact, these drugs may induce a detrimental effect on the immune system and gut microbiota, which is a well-known regulator of immune homeostasis [79]. A drug-based prognostic score developed by Buti et al.-including the cumulative exposure to high-dose steroid therapy (a dose of ≥ 10 mg prednisoneequivalents per day), antibiotics and PPIs-can help to stratify patients treated with ICIs in routine practice and clinical trials [73].
The post-hoc analyses of the single-arm phase II IMvigor210 trial and the randomised phase III IMvigor211 trial confirmed the negative predictive role of PPI and antibiotic use in patients with mUC treated with ICI, but not in the case of treatment with chemotherapy [80,81].

Inflammatory Indices
An inflammatory condition in patients with cancer has been associated with worse outcomes and a lower therapeutic response across different tumour types [82]. Inflammatory indices from peripheral blood have been investigated as potential biomarkers in different tumours, settings and therapies [83][84][85], including patients with UC treated with surgery or chemotherapy [64,[86][87][88].
Most of these analyses are derived from retrospective studies, except for the two prognostic scores developed by Fornarini et al. and Sonpavde et al., and the machine learning (ML) model developed by Abuhelwa et al. [104][105][106]. The two prognostic models were developed by the retrospective analysis of the phase IIIb SAUL trial [104] and the two phase I/II trials [105], both evaluating second-line ICI monotherapy. The ML model was built using the atezolizumab cohort of the IMvigor210 trial as the training cohort, and the IMvigor211 trial population as the external validation [106].
Most of these prognostic models (Table 2) need external validation with prospective studies before being incorporated into clinical practice; only one of them has shown a predictive value for immunotherapy but not for chemotherapy in a retrospective analysis [96]. Use of antibiotics and PPIs have a negative predictive role in mUC patients treated with ICIs but not in those treated with CT [80,81].

Inflammatory indices
Inflammatory indices, like NLR, CMP and albumin, have been investigated in several human cancer types as prognostic tools, especially regarding ICIs.

Combined prognostic tools
Combinations of inflammatory indices and clinical factors have been investigated in several human cancer types as prognostic tools, especially regarding ICIs.
Two prognostic models were developed in second-line ICIs mono-therapy [104,105]. A machine learning model was built from IMvigor210 atezolizumab arm and validated in IMvigor211 atezolizumab arm [106].

Radiomics
Due to the growing need to identify useful biomarkers to select the patients who are most likely to benefit from ICIs, the quantitative analysis of imaging features by artificial intelligence algorithms, namely radiomics, has recently been investigated as a possible surrogate marker to predict the outcome of patients treated with immunotherapy [107,108]. Radiomics has been reported as a promising approach to predict the response and survival outcomes in patients with NSCLC and melanoma receiving ICIs [109,110]. Three retrospective analyses investigated radiomics from baseline contrast-enhanced computed tomography (CT) images in patients with mUC receiving anti-PDL1 or anti-PD1 monotherapy [111][112][113].
Artificial intelligence algorithms may also allow us to combine the information obtained by the image features with other clinical and laboratory prognostic factors, thus increasing the diagnostic accuracy of predictive models [112,113].
Artificial intelligence and deep machine learning-based models may provide helpful decision tools to clinicians for the selection of patients for ICIs in the near future. However, none of these radiomics-based models have been assessed in patients with UC treated with chemotherapy yet; further studies on larger populations and external validation are still needed to confirm their predictive value.

Conclusions
In the current era, in which ICIs have been tested for several tumour types, it is crucial to identify prognostic and predictive biomarkers or models which are able to select UC patients who will benefit from ICIs, especially as novel therapeutic agents could soon become available in clinical practice for the treatment of UC. In addition, adjuvant ICIs have recently shown a benefit in terms of disease-free survival in patients with radically resected muscle-invasive urothelial carcinoma [112]. With more follow-up, a survival benefit is awaited, opening new methods for biomarker identification even in this setting.
By reviewing the available evidence and attempting a classification of clinical and tumour factors according to levels of evidence and grades of recommendation (see Table 3), both a prognostic and a predictive value is currently suggested for ctDNA, while only a prognostic relevance seems to apply to concomitant medications and patient's characteristics. increasing the diagnostic accuracy of predictive models [112,113]. Artificial intelligence and deep machine learning-based models may provide helpful decision tools to clinicians for the selection of patients for ICIs in the near future. However, none of these radiomics-based models have been assessed in patients with UC treated with chemotherapy yet; further studies on larger populations and external validation are still needed to confirm their predictive value.

Conclusions
In the current era, in which ICIs have been tested for several tumour types, it is crucial to identify prognostic and predictive biomarkers or models which are able to select UC patients who will benefit from ICIs, especially as novel therapeutic agents could soon become available in clinical practice for the treatment of UC. In addition, adjuvant ICIs have recently shown a benefit in terms of disease-free survival in patients with radically resected muscle-invasive urothelial carcinoma [112]. With more follow-up, a survival benefit is awaited, opening new methods for biomarker identification even in this setting.
By reviewing the available evidence and attempting a classification of clinical and tumour factors according to levels of evidence and grades of recommendation (see Table  3), both a prognostic and a predictive value is currently suggested for ctDNA, while only a prognostic relevance seems to apply to concomitant medications and patient's characteristics. increasing the diagnostic accuracy of predictive mod Artificial intelligence and deep machine learning decision tools to clinicians for the selection of patients none of these radiomics-based models have been a with chemotherapy yet; further studies on larger pop still needed to confirm their predictive value.

Conclusions
In the current era, in which ICIs have been tested to identify prognostic and predictive biomarkers or patients who will benefit from ICIs, especially as nov come available in clinical practice for the treatment of recently shown a benefit in terms of disease-free su sected muscle-invasive urothelial carcinoma [112]. W fit is awaited, opening new methods for biomarker id By reviewing the available evidence and attem tumour factors according to levels of evidence and g 3), both a prognostic and a predictive value is curren a prognostic relevance seems to apply to concomitan teristics. tained by the image features with other clinical and laboratory prognostic factors, thus increasing the diagnostic accuracy of predictive models [112,113]. Artificial intelligence and deep machine learning-based models may provide helpful decision tools to clinicians for the selection of patients for ICIs in the near future. However, none of these radiomics-based models have been assessed in patients with UC treated with chemotherapy yet; further studies on larger populations and external validation are still needed to confirm their predictive value.

Conclusions
In the current era, in which ICIs have been tested for several tumour types, it is crucial to identify prognostic and predictive biomarkers or models which are able to select UC patients who will benefit from ICIs, especially as novel therapeutic agents could soon become available in clinical practice for the treatment of UC. In addition, adjuvant ICIs have recently shown a benefit in terms of disease-free survival in patients with radically resected muscle-invasive urothelial carcinoma [112]. With more follow-up, a survival benefit is awaited, opening new methods for biomarker identification even in this setting.
By reviewing the available evidence and attempting a classification of clinical and tumour factors according to levels of evidence and grades of recommendation (see Table  3), both a prognostic and a predictive value is currently suggested for ctDNA, while only a prognostic relevance seems to apply to concomitant medications and patient's characteristics. tained by the image features with other clinical and increasing the diagnostic accuracy of predictive mod Artificial intelligence and deep machine learning decision tools to clinicians for the selection of patients none of these radiomics-based models have been a with chemotherapy yet; further studies on larger pop still needed to confirm their predictive value.

Conclusions
In the current era, in which ICIs have been tested to identify prognostic and predictive biomarkers or patients who will benefit from ICIs, especially as nov come available in clinical practice for the treatment of recently shown a benefit in terms of disease-free su sected muscle-invasive urothelial carcinoma [112]. W fit is awaited, opening new methods for biomarker id By reviewing the available evidence and attem tumour factors according to levels of evidence and g 3), both a prognostic and a predictive value is curren a prognostic relevance seems to apply to concomitan teristics. PD1 monotherapy [111][112][113]. Artificial intelligence algorithms may also allow us to combine the information obtained by the image features with other clinical and laboratory prognostic factors, thus increasing the diagnostic accuracy of predictive models [112,113].
Artificial intelligence and deep machine learning-based models may provide helpful decision tools to clinicians for the selection of patients for ICIs in the near future. However, none of these radiomics-based models have been assessed in patients with UC treated with chemotherapy yet; further studies on larger populations and external validation are still needed to confirm their predictive value.

Conclusions
In the current era, in which ICIs have been tested for several tumour types, it is crucial to identify prognostic and predictive biomarkers or models which are able to select UC patients who will benefit from ICIs, especially as novel therapeutic agents could soon become available in clinical practice for the treatment of UC. In addition, adjuvant ICIs have recently shown a benefit in terms of disease-free survival in patients with radically resected muscle-invasive urothelial carcinoma [112]. With more follow-up, a survival benefit is awaited, opening new methods for biomarker identification even in this setting.
By reviewing the available evidence and attempting a classification of clinical and tumour factors according to levels of evidence and grades of recommendation (see Table  3), both a prognostic and a predictive value is currently suggested for ctDNA, while only a prognostic relevance seems to apply to concomitant medications and patient's characteristics.  [111][112][113]. Artificial intelligence algorithms may also allow tained by the image features with other clinical and increasing the diagnostic accuracy of predictive mod Artificial intelligence and deep machine learning decision tools to clinicians for the selection of patients none of these radiomics-based models have been a with chemotherapy yet; further studies on larger pop still needed to confirm their predictive value.

Conclusions
In the current era, in which ICIs have been tested to identify prognostic and predictive biomarkers or patients who will benefit from ICIs, especially as nov come available in clinical practice for the treatment of recently shown a benefit in terms of disease-free su sected muscle-invasive urothelial carcinoma [112]. W fit is awaited, opening new methods for biomarker id By reviewing the available evidence and attem tumour factors according to levels of evidence and g 3), both a prognostic and a predictive value is curren a prognostic relevance seems to apply to concomitan teristics. Artificial intelligence and deep machine learning-based models may provide helpful decision tools to clinicians for the selection of patients for ICIs in the near future. However, none of these radiomics-based models have been assessed in patients with UC treated with chemotherapy yet; further studies on larger populations and external validation are still needed to confirm their predictive value.

Conclusions
In the current era, in which ICIs have been tested for several tumour types, it is crucial to identify prognostic and predictive biomarkers or models which are able to select UC patients who will benefit from ICIs, especially as novel therapeutic agents could soon become available in clinical practice for the treatment of UC. In addition, adjuvant ICIs have recently shown a benefit in terms of disease-free survival in patients with radically resected muscle-invasive urothelial carcinoma [112]. With more follow-up, a survival benefit is awaited, opening new methods for biomarker identification even in this setting.
By reviewing the available evidence and attempting a classification of clinical and tumour factors according to levels of evidence and grades of recommendation (see Table  3), both a prognostic and a predictive value is currently suggested for ctDNA, while only a prognostic relevance seems to apply to concomitant medications and patient's characteristics. Artificial intelligence and deep machine learning decision tools to clinicians for the selection of patients none of these radiomics-based models have been as with chemotherapy yet; further studies on larger pop still needed to confirm their predictive value.

Conclusions
In the current era, in which ICIs have been tested to identify prognostic and predictive biomarkers or patients who will benefit from ICIs, especially as nov come available in clinical practice for the treatment of recently shown a benefit in terms of disease-free su sected muscle-invasive urothelial carcinoma [112]. W fit is awaited, opening new methods for biomarker id By reviewing the available evidence and attem tumour factors according to levels of evidence and g 3), both a prognostic and a predictive value is curren a prognostic relevance seems to apply to concomitan teristics. Radiomics-based models 7 IV, B ORR, PFS, OS PFS, progress-free survival; OS, overall survival; ORR, overall response rate; DFS, di gated. 1 DNA and RNA signatures such as the APOBEC, TGFbeta, IFN-γ and IFN-alph
Artificial intelligence algorithms may also allow us to combine the information ob-tained by the image features with other clinical and laboratory prognostic factors, thus increasing the diagnostic accuracy of predictive models [112,113].
Artificial intelligence and deep machine learning-based models may provide helpful decision tools to clinicians for the selection of patients for ICIs in the near future. However, none of these radiomics-based models have been assessed in patients with UC treated with chemotherapy yet; further studies on larger populations and external validation are still needed to confirm their predictive value.

Conclusions
In the current era, in which ICIs have been tested for several tumour types, it is crucial to identify prognostic and predictive biomarkers or models which are able to select UC patients who will benefit from ICIs, especially as novel therapeutic agents could soon be-come available in clinical practice for the treatment of UC. In addition, adjuvant ICIs have recently shown a benefit in terms of disease-free survival in patients with radically re-sected muscle-invasive urothelial carcinoma [112]. With more follow-up, a survival bene-fit is awaited, opening new methods for biomarker identification even in this setting.
By reviewing the available evidence and attempting a classification of clinical and tumour factors according to levels of evidence and grades of recommendation (see Table   3), both a prognostic and a predictive value is currently suggested for ctDNA, while only a prognostic relevance seems to apply to concomitant medications and patient's charac-teristics.

Conclusions
In the current era, in which ICIs have been tested to identify prognostic and predictive biomarkers or patients who will benefit from ICIs, especially as nov come available in clinical practice for the treatment of recently shown a benefit in terms of disease-free su sected muscle-invasive urothelial carcinoma [112]. W fit is awaited, opening new methods for biomarker id By reviewing the available evidence and attem tumour factors according to levels of evidence and g 3), both a prognostic and a predictive value is curren a prognostic relevance seems to apply to concomitan teristics.

Clinical factors
Patient's characteristics 3 PD1 monotherapy [111][112][113]. Artificial intelligence algorithms may also allow us to combine the information obtained by the image features with other clinical and laboratory prognostic factors, thus increasing the diagnostic accuracy of predictive models [112,113].
Artificial intelligence and deep machine learning-based models may provide helpful decision tools to clinicians for the selection of patients for ICIs in the near future. However, none of these radiomics-based models have been assessed in patients with UC treated with chemotherapy yet; further studies on larger populations and external validation are still needed to confirm their predictive value.

Conclusions
In the current era, in which ICIs have been tested for several tumour types, it is crucial to identify prognostic and predictive biomarkers or models which are able to select UC patients who will benefit from ICIs, especially as novel therapeutic agents could soon become available in clinical practice for the treatment of UC. In addition, adjuvant ICIs have recently shown a benefit in terms of disease-free survival in patients with radically resected muscle-invasive urothelial carcinoma [112]. With more follow-up, a survival benefit is awaited, opening new methods for biomarker identification even in this setting.
By reviewing the available evidence and attempting a classification of clinical and tumour factors according to levels of evidence and grades of recommendation (see Table  3), both a prognostic and a predictive value is currently suggested for ctDNA, while only a prognostic relevance seems to apply to concomitant medications and patient's characteristics. Artificial intelligence and deep machine learning-based models may provide helpful decision tools to clinicians for the selection of patients for ICIs in the near future. However, none of these radiomics-based models have been assessed in patients with UC treated with chemotherapy yet; further studies on larger populations and external validation are still needed to confirm their predictive value.

Conclusions
In the current era, in which ICIs have been tested for several tumour types, it is crucial to identify prognostic and predictive biomarkers or models which are able to select UC patients who will benefit from ICIs, especially as novel therapeutic agents could soon become available in clinical practice for the treatment of UC. In addition, adjuvant ICIs have recently shown a benefit in terms of disease-free survival in patients with radically resected muscle-invasive urothelial carcinoma [112]. With more follow-up, a survival benefit is awaited, opening new methods for biomarker identification even in this setting.
By reviewing the available evidence and attempting a classification of clinical and tumour factors according to levels of evidence and grades of recommendation (see Table  3), both a prognostic and a predictive value is currently suggested for ctDNA, while only a prognostic relevance seems to apply to concomitant medications and patient's characteristics. computed tomography (CT) images in patients with mUC receiving anti-PDL1 or anti-PD1 monotherapy [111][112][113]. Artificial intelligence algorithms may also allow us to combine the information obtained by the image features with other clinical and laboratory prognostic factors, thus increasing the diagnostic accuracy of predictive models [112,113].
Artificial intelligence and deep machine learning-based models may provide helpful decision tools to clinicians for the selection of patients for ICIs in the near future. However, none of these radiomics-based models have been assessed in patients with UC treated with chemotherapy yet; further studies on larger populations and external validation are still needed to confirm their predictive value.

Conclusions
In the current era, in which ICIs have been tested for several tumour types, it is crucial to identify prognostic and predictive biomarkers or models which are able to select UC patients who will benefit from ICIs, especially as novel therapeutic agents could soon become available in clinical practice for the treatment of UC. In addition, adjuvant ICIs have recently shown a benefit in terms of disease-free survival in patients with radically resected muscle-invasive urothelial carcinoma [112]. With more follow-up, a survival benefit is awaited, opening new methods for biomarker identification even in this setting.
By reviewing the available evidence and attempting a classification of clinical and tumour factors according to levels of evidence and grades of recommendation (see Table  3), both a prognostic and a predictive value is currently suggested for ctDNA, while only a prognostic relevance seems to apply to concomitant medications and patient's characteristics.  [107,108]. Radiomics has been reported as a promisi and survival outcomes in patients with NSCLC and Three retrospective analyses investigated radiomic computed tomography (CT) images in patients with PD1 monotherapy [111][112][113].
Artificial intelligence algorithms may also allow tained by the image features with other clinical and increasing the diagnostic accuracy of predictive mod Artificial intelligence and deep machine learning decision tools to clinicians for the selection of patients none of these radiomics-based models have been a with chemotherapy yet; further studies on larger pop still needed to confirm their predictive value.

Conclusions
In the current era, in which ICIs have been tested to identify prognostic and predictive biomarkers or patients who will benefit from ICIs, especially as nov come available in clinical practice for the treatment of recently shown a benefit in terms of disease-free su sected muscle-invasive urothelial carcinoma [112]. W fit is awaited, opening new methods for biomarker id By reviewing the available evidence and attem tumour factors according to levels of evidence and g 3), both a prognostic and a predictive value is curren a prognostic relevance seems to apply to concomitan teristics.    [107,108]. Radiomics has been reported as a promising approach to predict the response and survival outcomes in patients with NSCLC and melanoma receiving ICIs [109,110].
Artificial intelligence algorithms may also allow us to combine the information ob-tained by the image features with other clinical and laboratory prognostic factors, thus increasing the diagnostic accuracy of predictive models [112,113].
Artificial intelligence and deep machine learning-based models may provide helpful decision tools to clinicians for the selection of patients for ICIs in the near future. However, none of these radiomics-based models have been assessed in patients with UC treated with chemotherapy yet; further studies on larger populations and external validation are still needed to confirm their predictive value.

Conclusions
In the current era, in which ICIs have been tested for several tumour types, it is crucial to identify prognostic and predictive biomarkers or models which are able to select UC patients who will benefit from ICIs, especially as novel therapeutic agents could soon be-come available in clinical practice for the treatment of UC. In addition, adjuvant ICIs have recently shown a benefit in terms of disease-free survival in patients with radically re-sected muscle-invasive urothelial carcinoma [112]. With more follow-up, a survival bene-fit is awaited, opening new methods for biomarker identification even in this setting.
By reviewing the available evidence and attempting a classification of clinical and tumour factors according to levels of evidence and grades of recommendation (see Table   3), both a prognostic and a predictive value is currently suggested for ctDNA, while only a prognostic relevance seems to apply to concomitant medications and patient's charac-teristics. tained by the image features with other clinical and laboratory prognostic factors, thus increasing the diagnostic accuracy of predictive models [112,113]. Artificial intelligence and deep machine learning-based models may provide helpful decision tools to clinicians for the selection of patients for ICIs in the near future. However, none of these radiomics-based models have been assessed in patients with UC treated with chemotherapy yet; further studies on larger populations and external validation are still needed to confirm their predictive value.

Conclusions
In the current era, in which ICIs have been tested for several tumour types, it is crucial to identify prognostic and predictive biomarkers or models which are able to select UC patients who will benefit from ICIs, especially as novel therapeutic agents could soon become available in clinical practice for the treatment of UC. In addition, adjuvant ICIs have recently shown a benefit in terms of disease-free survival in patients with radically resected muscle-invasive urothelial carcinoma [112]. With more follow-up, a survival benefit is awaited, opening new methods for biomarker identification even in this setting.
By reviewing the available evidence and attempting a classification of clinical and tumour factors according to levels of evidence and grades of recommendation (see Table  3), both a prognostic and a predictive value is currently suggested for ctDNA, while only a prognostic relevance seems to apply to concomitant medications and patient's characteristics.  [107,108]. Radiomics has been reported as a promising approach to predict the response and survival outcomes in patients with NSCLC and melanoma receiving ICIs [109,110].
Artificial intelligence algorithms may also allow us to combine the information ob-tained by the image features with other clinical and laboratory prognostic factors, thus increasing the diagnostic accuracy of predictive models [112,113].
Artificial intelligence and deep machine learning-based models may provide helpful decision tools to clinicians for the selection of patients for ICIs in the near future. However, none of these radiomics-based models have been assessed in patients with UC treated with chemotherapy yet; further studies on larger populations and external validation are still needed to confirm their predictive value.

Conclusions
In the current era, in which ICIs have been tested for several tumour types, it is crucial to identify prognostic and predictive biomarkers or models which are able to select UC patients who will benefit from ICIs, especially as novel therapeutic agents could soon be-come available in clinical practice for the treatment of UC. In addition, adjuvant ICIs have recently shown a benefit in terms of disease-free survival in patients with radically re-sected muscle-invasive urothelial carcinoma [112]. With more follow-up, a survival bene-fit is awaited, opening new methods for biomarker identification even in this setting.
By reviewing the available evidence and attempting a classification of clinical and tumour factors according to levels of evidence and grades of recommendation (see Table   3), both a prognostic and a predictive value is currently suggested for ctDNA, while only a prognostic relevance seems to apply to concomitant medications and patient's charac-teristics.  [107,108]. Radiomics has been reported as a promisi and survival outcomes in patients with NSCLC and Three retrospective analyses investigated radiomic computed tomography (CT) images in patients with PD1 monotherapy [111][112][113].
Artificial intelligence algorithms may also allow tained by the image features with other clinical and increasing the diagnostic accuracy of predictive mod Artificial intelligence and deep machine learning decision tools to clinicians for the selection of patients none of these radiomics-based models have been a with chemotherapy yet; further studies on larger pop still needed to confirm their predictive value.

Conclusions
In the current era, in which ICIs have been tested to identify prognostic and predictive biomarkers or patients who will benefit from ICIs, especially as nov come available in clinical practice for the treatment of recently shown a benefit in terms of disease-free su sected muscle-invasive urothelial carcinoma [112]. W fit is awaited, opening new methods for biomarker id By reviewing the available evidence and attem tumour factors according to levels of evidence and g 3), both a prognostic and a predictive value is curren a prognostic relevance seems to apply to concomitan teristics.  [107,108]. Radiomics has been reported as a promising approach to predict the response and survival outcomes in patients with NSCLC and melanoma receiving ICIs [109,110].
Artificial intelligence algorithms may also allow us to combine the information ob-tained by the image features with other clinical and laboratory prognostic factors, thus increasing the diagnostic accuracy of predictive models [112,113].
Artificial intelligence and deep machine learning-based models may provide helpful decision tools to clinicians for the selection of patients for ICIs in the near future. However, none of these radiomics-based models have been assessed in patients with UC treated with chemotherapy yet; further studies on larger populations and external validation are still needed to confirm their predictive value.

Conclusions
In the current era, in which ICIs have been tested for several tumour types, it is crucial to identify prognostic and predictive biomarkers or models which are able to select UC patients who will benefit from ICIs, especially as novel therapeutic agents could soon be-come available in clinical practice for the treatment of UC. In addition, adjuvant ICIs have recently shown a benefit in terms of disease-free survival in patients with radically re-sected muscle-invasive urothelial carcinoma [112]. With more follow-up, a survival bene-fit is awaited, opening new methods for biomarker identification even in this setting.
By reviewing the available evidence and attempting a classification of clinical and tumour factors according to levels of evidence and grades of recommendation (see Table   3), both a prognostic and a predictive value is currently suggested for ctDNA, while only a prognostic relevance seems to apply to concomitant medications and patient's charac-teristics. Three retrospective analyses investigated radiomics from baseline contrast-enhanced computed tomography (CT) images in patients with mUC receiving anti-PDL1 or anti-PD1 monotherapy [111][112][113]. Artificial intelligence algorithms may also allow us to combine the information ob-tained by the image features with other clinical and laboratory prognostic factors, thus increasing the diagnostic accuracy of predictive models [112,113].
Artificial intelligence and deep machine learning-based models may provide helpful decision tools to clinicians for the selection of patients for ICIs in the near future. However, none of these radiomics-based models have been assessed in patients with UC treated with chemotherapy yet; further studies on larger populations and external validation are still needed to confirm their predictive value.

Conclusions
In the current era, in which ICIs have been tested for several tumour types, it is crucial to identify prognostic and predictive biomarkers or models which are able to select UC patients who will benefit from ICIs, especially as novel therapeutic agents could soon become available in clinical practice for the treatment of UC. In addition, adjuvant ICIs have recently shown a benefit in terms of disease-free survival in patients with radically resected muscle-invasive urothelial carcinoma [112]. With more follow-up, a survival benefit is awaited, opening new methods for biomarker identification even in this setting.
By reviewing the available evidence and attempting a classification of clinical and tumour factors according to levels of evidence and grades of recommendation (see Table  3), both a prognostic and a predictive value is currently suggested for ctDNA, while only a prognostic relevance seems to apply to concomitant medications and patient's characteristics.  [107,108]. Radiomics has been reported as a promising approach to predi and survival outcomes in patients with NSCLC and melanoma receiving Three retrospective analyses investigated radiomics from baseline con computed tomography (CT) images in patients with mUC receiving anti PD1 monotherapy [111][112][113].
Artificial intelligence algorithms may also allow us to combine the i tained by the image features with other clinical and laboratory prognost increasing the diagnostic accuracy of predictive models [112,113].
Artificial intelligence and deep machine learning-based models may p decision tools to clinicians for the selection of patients for ICIs in the near fu none of these radiomics-based models have been assessed in patients w with chemotherapy yet; further studies on larger populations and externa still needed to confirm their predictive value.

Conclusions
In the current era, in which ICIs have been tested for several tumour ty to identify prognostic and predictive biomarkers or models which are ab patients who will benefit from ICIs, especially as novel therapeutic agents come available in clinical practice for the treatment of UC. In addition, adju recently shown a benefit in terms of disease-free survival in patients wi sected muscle-invasive urothelial carcinoma [112]. With more follow-up, a fit is awaited, opening new methods for biomarker identification even in t By reviewing the available evidence and attempting a classification tumour factors according to levels of evidence and grades of recommenda 3), both a prognostic and a predictive value is currently suggested for ctDN a prognostic relevance seems to apply to concomitant medications and p teristics.  [107,108]. Radiomics has been reported as a p and survival outcomes in patients with NSC Three retrospective analyses investigated ra computed tomography (CT) images in patien PD1 monotherapy [111][112][113].
Artificial intelligence algorithms may als tained by the image features with other clini increasing the diagnostic accuracy of predictiv Artificial intelligence and deep machine l decision tools to clinicians for the selection of p none of these radiomics-based models have with chemotherapy yet; further studies on lar still needed to confirm their predictive value.

Conclusions
In the current era, in which ICIs have been to identify prognostic and predictive biomar patients who will benefit from ICIs, especially come available in clinical practice for the treat recently shown a benefit in terms of disease sected muscle-invasive urothelial carcinoma [ fit is awaited, opening new methods for biom By reviewing the available evidence and tumour factors according to levels of evidenc 3), both a prognostic and a predictive value is a prognostic relevance seems to apply to con teristics.

of 18
been reported as a promising approach to predict the response patients with NSCLC and melanoma receiving ICIs [109,110]. yses investigated radiomics from baseline contrast-enhanced T) images in patients with mUC receiving anti-PDL1 or anti-13].
algorithms may also allow us to combine the information obres with other clinical and laboratory prognostic factors, thus accuracy of predictive models [112,113]. and deep machine learning-based models may provide helpful for the selection of patients for ICIs in the near future. However, based models have been assessed in patients with UC treated urther studies on larger populations and external validation are eir predictive value.
which ICIs have been tested for several tumour types, it is crucial d predictive biomarkers or models which are able to select UC from ICIs, especially as novel therapeutic agents could soon bepractice for the treatment of UC. In addition, adjuvant ICIs have in terms of disease-free survival in patients with radically reothelial carcinoma [112]. With more follow-up, a survival benew methods for biomarker identification even in this setting. ailable evidence and attempting a classification of clinical and to levels of evidence and grades of recommendation (see Table  a [7]. Levels of evidence: I, evidence from at least one large randomised, controlled trial of good methodological quality (low potential for bias) or meta-analyses of well-conducted randomised trials without heterogeneity; II, small randomised trials or large randomised trials with a suspicion of bias (lower methodological quality), or meta-analyses of such trials or of trials with demonstrated heterogeneity; III, prospective cohort studies; IV, retrospective cohort studies or case-control studies; V, studies without a control group, case reports, or expert opinions. Grades of recommendation: A, strong evidence for efficacy with a substantial clinical benefit, strongly recommended; B, strong or moderate evidence for efficacy, but with a limited clinical benefit, generally recommended; C, insufficient evidence for efficacy or benefit does not outweigh the risk or the disadvantages (adverse events, costs, etc.), optional; D, moderate evidence against efficacy or for adverse outcomes, generally not recommended; E, strong evidence against efficacy or for adverse outcomes, never recommended.
We think this information may be helpful to clinically assess mUCs patients who are considered for treatment with ICIs, and could also drive further prospective research on these biomarkers, either as single factors or within combined prognostic models implemented by artificial intelligence algorithms.