GI-PREDICT

GI-PREDICT is an observational platform study in which predictive biomarkers of sensitivity/resistance to anticancer treatments will be evaluated, or the prediction of molecular characteristics will be assessed by integrating radiomic and pathomic data using artificial intelligence (AI) models in gastrointestinal malignancies.

Background
The widespread availability of comprehensive genomic profiling (CGP) allows to concomitantly assess guidelines-recommended biomarkers and additional actionable drivers in cancer patients, especially in patients suffering from gastrointestinal (GI) tumors(1). In this setting, high-throughput genomic and transcriptomic profiling parallels the increasing availability of innovative drugs, including novel immmunotherapeutics or targeted agents.

For instance, microsatellite instability-high status (MSI-high) is recognized as an agnostic biomarker of benefit from immune checkpoint inhibitors, leading to the FDA and EMA approval of the immune-checkpoint inhibitor (ICI) Pembrolizumab for the treatment of patients with MSI-high cancers(2). Noteworthy, context specificity of MSI-high as a biomarker has been described in different tumors(3) and a clinically meaningful rate of primary resistance is observed in patients with MSI-high mCRC, gastro-esophageal (GEC), pancreatic (pancreatic ductal adenocarcinoma, PDAC) and biliary tract (BTCs) tumors receiving anti-PD-1-based therapies(2, 3). Additional biomarkers such as high expression of Programmed Death Ligand 1 (PD-L1) within the tumor microenvironment has also been associated with efficacy of ICIs in metastatic GEC and its assessment is now recommended by most of the healthcare systems(4). However, this biomarker has still several limitations, including the different assessment methods and thresholds adopted for treatment selection according to each ICIs in GECs. Moreover, ICIs have entered clinical practice also for the first-line treatment of patients with advanced BTCs, unselected for molecular features and without clear evidence of predictivity for PD-L1 expression(5). These issues highlight the unmet need of refining patients’ selection to ICIs-based treatments in GI cancers.

Beyond immunotherapy, molecularly-guided therapies are now recommended by international guidelines for the treatment of different GI tumors(5-8). Among these, targeting the human epidermal growth factor receptor 2 (HER2) pathways evidenced a benefit in terms of tumor response and survival outcomes in patients with mCRC, mGEC and BTCs with a high expression of HER2(9-13). Similarly, the introduction of other target agents, either as monotherapy or combined with chemotherapy, found different degrees of success in the treatment of GI tumors, including mCRC, BTCs and PDACs with KRAS(G12C)(14, 15) or BRAF(V600E)(16, 17) mutations, mGECs with CLDN18.2 overexpression(18), metastatic BTCs harboring FGFR2 fusions(19) or IDH1 mutations(20) and any tumor with NTRK 1-3 fusions(21).
However, these treatments are burdened by significant rates of primary and secondary resistance and only a small proportion of patients experience long-term benefit, highlighting the need for a more refined analysis of predictive biomarkers. For example, the study of common genomic alterations such as KRAS and BRAF mutations and more uncommon molecular features, analyzed through specific panels (such as the PRESSING panels), can identify a subset of mCRC refractory to anti-EGFR treatment(22).

Finally, the continuous research for new biomarkers and targeted agents could lead to future novel strategies in cancer types such as mPDACs, in which neither the current target therapies nor immunotherapy provided convincing results. For example, while germline mutations of the BRCA1/2 and PALB2 genes are known biomarkers of DNA homologous recombination deficiency (HRD) and thus sensitivity to platinum-based chemotherapy and PARP-inhibitors in patients affected by PDAC(23-25) these mutations are very rare (<5% of the Western PDAC patient population) and a larger subset of PDACs (approximately up to 25% of cases) have been shown to express HRD despite not harboring these mutations. Therefore, it is imperative to develop new reliable methods to define HRD status as a biomarker for PDAC patients’ treatment in clinical practice.

These considerations highlight the need to improve the evaluation of molecular biomarkers in patients suffering from GI tumors, improving their predictivity and identifying new ones. Moreover, given the fundamental role of molecular biomarkers in modern GI oncology, it is imperative to reduce costs, both in terms of time and resources, and increase overall access to molecular profiling for patients.

AI applications in medicine are revolutionizing translational research and clinical practice, enabling the processing of large amounts of unstructured information to yield clinically directive information. In the last five years, exponential technical advancements have led to the application of AI methods also in oncology for several purposes, from the interpretation of radiological images to the analysis of histopathology whole-slide images (WSIs) for early tumor diagnosis, biomarkers identification, the choice of treatments and the prediction of patients’ prognosis(26). In particular, automated histopathology analyses utilizing deep learning (also defined as “pathomics”) focus on deriving clinically relevant prediction from WSIs, such as molecular alterations (gene mutational status, expression or complex biomarkers definition such as MSI or HRD status) and treatment response probability(27). Numerous studies have shown potential of pathomics in biomarkers prediction, making their evaluation possible even in cases where the tissue sample is insufficient for a traditional analysis(28, 29). Moreover, these techniques allow the evaluation of genotypic-phenotypic correlations and biological processes, including intratumor heterogeneity and composition(30), thus outperforming the biormaker-direct evaluation usually performed in clinical practice. Concerning the use of AI for the interpretation of radiological images, the term “radiomics” identifies the methodology that converts qualitative medical image data into quantitative and high-dimensional features by high-throughput feature extraction, thereby provide valuable information on the entire underlying intra-tumor heterogeneity and cancer phenotype(31). In recent years, several studies have proved the potential value of radiomics-based models to predict tumor extension and evolution (including early tumor diagnosis), tumor molecular features and patients outcomes, outperforming radiologists performance(32, 33). Despite this progress, dedicated studies are needed for calibration and validation of AI methods for narrower applications in specific patient and tumor settings, as is the case of molecularly selected GI tumors.

Based on these considerations, in this study we aim at collecting WSIs, radiological images and clinicopathologic tabular data in order to develop AI-based pathomics and radiomics models to reliably identify molecular biomarkers and predict patients’ prognosis in specific subsets of patients receiving treatments for GI tumors.

This study will evaluate approximately 700 patients with different GI malignancies and exposed to selected treatments according to the molecular profile at 7 Institutions.
For each patient an hematoxylin and eosin section from treatment-naïve tissue, the chest and abdominal CT scan performed prior to the specific treatment under investigation will be collected and clinical data will be collected for biomarker analysis.

The planned study populations include subjects who met the following key inclusion criteria:

• confirmed histological or cytological diagnosis of gastrointestinal cancer, including GECs, PDACs, BTCs and CRCs;
• eligible to systemic treatment or having received at least one cycle of systemic anticancer treatment, including chemotherapy, targeted therapies or immunotherapy;
• available clinical data with adequate follow up to assess treatment efficacy and survival;
• available histological slides (hematoxylin and eosin stained) from formalin-fixed paraffin-embedded (FFPE) tumor tissue collected before treatment initiation;
• available radiological images (CT scan preferred) at baseline of anticancer treatment of interest.

The primary objective of this study is to develop AI-based pathomics and radiomics models from WSIs, radiological images and clinical data of patients affected by molecularly-selected GI cancers, with the aim of predicting treatment efficacy and patients’ outcomes.

Secondary objectives will include:

• To compare the performance of AI-based models for treatment efficacy prediction with the performance of established biomarkers and risk-assessment models previously validated and/or adopted in clinical practice.
• The development of AI-based pathomics and radiomics models for molecular biomarkers prediction and treatment selection in patients affected by GI cancers.
• The investigation of novel biomarkers of sensitivity/resistance to anticancer therapies in molecularly-selected cohorts of patients affected by GI cancers.
• The use of AI-based pathomics and radiomics models for reclassification of tumors originating from the GI tract for which there is an uncertain histological diagnosis (e.g. mixed hepatocellular-cholangiocarcinoma) or which have an unknown primary origin (cancers of unknown primary).
• To evaluate the cost-benefit ratio, in terms of time and resources, of AI-based models compared with standard methods for predicting molecular biomarkers and treatment efficacy.