Implementation of a Clinical Decision Support System for Precision Oncology across an Academic Network


 There is a growing need for systems that efficiently support the work of medical teams at the precision oncology point-of-care. Here we present the implementation of the Molecular Tumor Board Portal (MTBP), an academic clinical decision support system that creates a unified legal, scientific and technological platform to share and harness next-generation sequencing data across the Cancer Core Europe network. Automating the interpretation and reporting of sequencing results decreased drastically the need for manual procedures that are time consuming and prone to errors. In addition, the adoption of an expert-agreed process to systematically link tumor molecular profiles with clinical actions promoted consistent decision-making and structured data capture across centers. Finally, the use of information-rich patient reports with interactive content facilitated collaborative discussion of complex cases during virtual molecular tumor board meetings. Overall, we believe that streamlined digital systems like the MTBP are crucial to better address the challenges brought by precision oncology and accelerate the use of emerging biomarkers.


Introduction
Next-generation sequencing (NGS) assays are a key component of the modern oncology work ow.
Beyond on-label drug prescriptions, tumor sequencing results can guide clinical trial enrollment and identify investigational drug opportunities in individual patients. NGS data can also reveal other events of clinical relevance, such as germline pathogenic variants, pharmacogenomics ndings and clonal hematopoiesis drivers, which should be recognized and acted upon. However, clinical interpretation of NGS results often relies on manual procedures, which poses considerable challenges to the medical teams undertaking this task. First, variant annotation bene ts from numerous resources developed by medical, biological and bioinformatics domains that are not easy to integrate. Second, agreeing on annotation criteria and rules to prioritize actionable ndings is critical for consistent clinical decisionmaking. And third, in the absence of on-label treatment options, patients must be matched with the speci c portfolio of investigational therapies and clinical trials available in each hospital (or hospital network), which are subject to continuous changes. Failure to address these issues or the inability to perform them in a clinically acceptable time frame can impair the outcome of individual patients and precision cancer medicine initiatives.
Clinical decision support systems (CDSS) can tackle these challenges by implementing e cient data analysis and reporting processes. Several commercial CDSS software are currently available, but inhouse solutions are often used to better accommodate speci c needs of each center. In fact, we believe that the capacity of academic institutions to develop custom CDSS accelerates the use of emerging biomarkers, promotes precision medicine across healthcare professionals and catalyzes clinical research. We have therefore developed the Molecular Tumor Board Portal (MTBP), a CDSS that creates a uni ed framework to interpret sequencing results across the seven comprehensive cancer centers that form Cancer Core Europe (CCE)1 at present. Importantly, the portal is seamlessly integrated in the CCE clinical work ows and provides a single platform to distribute the results and support shared discussions at scale2. Seamless communication among clinical investigators is essential to leverage the collective expertise of the community in this era of rapidly changing precision oncology landscape. To our knowledge, this is an unprecedented effort for co-developing new anti-cancer drugs and biomarkers in Europe. Here, we describe our approach and discuss the results of using the MTBP in a consecutive cohort of 500 advanced solid tumors evaluated from January 2019 to January 2021 in the context of the Basket of Baskets (NCT03767075), an ongoing CCE multi-basket phase II clinical trial matching molecular biomarkers with immunotherapies and targeted drugs.

Functional interpretation of cancer variants
Interpretation of NGS data rst requires elucidating whether the variants observed in cancer genes can confer tumor-promoting traits, as not all of them have equal biological consequences. Besides the identi cation of the individual tumor genomic drivers, this analysis enables matching patients to biomarkers de ned by functional criteria, such as 'activating' mutations in a given oncogene or 'loss-offunction' alterations in a given tumor suppressor. Of note, close to one third of the cancer biomarkers reported at present rely on interpreting the functional effect of variants found in drug targets (Supp Figure   1). This number will likely continue to grow as genes involved in more cellular processes, such as DNA damage repair, epigenetics, metabolism and immune-regulation, become actionable.
The MTBP interprets the functional relevance of cancer variants under an allele-centric perspective ( Fig   1A). In other words, a given BRCA1 mutation known to disrupt the activity of the wild type allele will always be declared as functionally relevant (i.e. loss-of-function) regardless of context considerations such as the germline versus somatic origin of the variant, the status of the second allele and/or the cancer subtype in which it is observed, which are however contemplated in the actionability analysis (see next section). Multiple genomic knowledge resources can be integrated for a more comprehensive variants' functional annotation, but currently there are no well established guidelines on how to do it. Therefore, we agreed on criteria considered to provide strong or very strong supporting evidence (>90% and >99% of certainty, respectively) as extrapolated from previous work3 and based on three distinct sources of knowledge (Supp Figure 2A and 2B): 1. First, the MTBP inspects whether the gene variants observed in the tumor have an already wellreported effect. Note that different effect assertion types can be equally mapped to the contextagnostic notion of a variant being functionally relevant; for example, a given BRCA1 mutation can be considered to be a putative loss-of-function event both when reported to predispose to early breast/ovarian cancer or a known biomarker of PARP inhibitor response. Therefore, the MTBP queries a number of expert knowledge databases that continuously gather results of clinical, experimental and population genetic studies4-10 according to the standard procedure de ned for each of their respective scopes (such as the pathogenicity classi cation of germline variants11), and assertions compatible with the functional relevance (or lack thereof) of the variant are then matched as appropriate (Supp Figure 2A). The aggregation of these knowledgebases, which are not often used in combination, enables a better use of international curation efforts (Supp Figure 2C).
2. Second, if no variant effect is reported, or the information is inconclusive and/or supported by weaker evidence (Supp Figure 2C), the portal evaluates whether bona de biological assumptions (such as whether a given premature termination codon is likely to trigger nonsense-mediated decay) can be applied (Supp Figure 2D). These assumptions are largely based on accepted criteria to identify loss-of-function variants in Mendelian disease genes12. Of note, the MTBP re nes the use of some of these criteria by leveraging the aggregated content of the aforementioned knowledgebases, which for example help to delineate protein regions that are critical for a given tumor suppressor function (Supp Figure 2D). This exempli es the value of the MTBP for integrating the knowledge available in the community and developing ensemble bioinformatics models.
3. Third, if none of these bona de assumptions can be applied or ful lled, computational-based metrics are used as the lowest level of supporting evidence. For example, hotspots of somatic mutations observed across previously sequenced cancer cohorts point out mutated protein sites that are preferentially selected by tumors and thus relevant for the disease development in both oncogenes and tumor suppressors13. In order to reduce the number of false positives, the MTBP uses methods that consider underlying genomic mutational processes to declare the accumulation of mutations of a given consequence type as statistically signi cant14,15. In addition, functional impact predictions can be also used to estimate whether other variants drive loss-of-function events.
Among all the methods developed with that purpose, we decided to use deleteriousness score calculations16 with stringent thresholds exhibiting a 90% of predictive value, as required for strong supporting criteria3, based on the results of our own benchmarking (Supp. Figure 2E).
Variants that cannot be classi ed as functionally relevant or functionally neutral according to any of the aforementioned criteria appear as of unknown functional signi cance in the MTBP reports. For the CCE prospective cohort presented here, composed of 500 solid tumors pro led by NGS panels (including up to 350 cancer genes; Table 1), the MTBP identi ed a median of 3 (interquartile range 2-4) functionally relevant mutations (single nucleotide changes and/or small indels) per tumor. Overall, and after excluding mutations assumed to be non-relevant (such as those that do not alter the protein sequence or are known polymorphisms), a total of 26% of the tumor mutations were classi ed as (putative) functionally relevant, while 9% were classi ed as (putative) neutral ( Figure 1B). One fourth of these classi cations were solely based on bioinformatic predictions, which as discussed is the lowest level of supporting evidence. Therefore, and even with the comprehensive functional annotation provided by the MTBP, most (65%) of the tumor mutations observed in cancer genes remained of unknown functional signi cance (although this number largely varies across genes; Figure 1C). This illustrates our still limited ability for interpreting the biological relevance of the genomic alterations that occur in tumor cells. As drug prescriptions progressively move towards a more holistic consideration of the tumor genome (pathway and/or signature-centric), we stress the importance of using interpretation tools that are kept up-to-date with the knowledge provided by emerging capabilities, such as high throughput functional assays.

Clinical interpretation of cancer variants
The nal objective of the MTBP is to help translate NGS results into the most appropriate therapeutic decisions according to state-of-the-art knowledge. Genomic alterations that in uence anticancer drug response (sensitivity or resistance) and that are of diagnostic or prognostic value are continuously reported in the literature and scienti c venues. Several international initiatives gather this information in speci c knowledgebases open for the access and feedback of the community6-8. However, these resources follow varying data models, and the accurate aggregation of their content requires an extensive harmonization of the lexicon, ontologies and genomic variant representation syntax employed by each.
The MTBP implements this process with a semi-automatic pipeline that combines a number of bioinformatic mapping tools17,18 and manually annotated dictionaries. The adoption of information exchange standards in the community is crucial to mitigate the need for these efforts and facilitate genomic knowledge sharing19-21.
The MTBP matches the cancer biomarkers aggregated across these knowledgebases with the variants observed in the tumor for (i) a speci c nucleotide and/or protein amino acid change (such as 'KIT:p.D572A'); (ii) a variant category (such as 'EGFR in-frame deletions in exon 19'); or (iii) a functional entity (such as 'FLT3 oncogenic mutations', as guided by the MTBP functional interpretation) (Supp Figure 3A). However, as previously mentioned, variant actionability must also take into account additional tumor context considerations beyond the mere variant match, such as the concordance between the biomarker' and patient's cancer type (or a subtype thereof), the presence of co-occurring alterations that can in uence the biomarker effect and the level of evidence that supports its clinical utility at present (Supp Figure 3B). The MTBP pipeline factors in these considerations as appropriate and automatically reports the results following the European Society for Medical Oncology (ESMO) Clinical Actionability of Molecular Targets (ESCAT) scale22, which is an extension of that previously presented by American expert associations23 (Supp Figure 3C).
The highest level of actionability corresponds to genomic alterations matching on-label prescriptions or clinical expert group recommendations, and thus ready to be used in routine clinical practice (ESCAT Level I; Figure 1D). However, in the context of CCE initiatives, we mostly pro le tumors of patients without standard-of-care therapeutic options available. Consequently, investigational and off-label drug opportunities based on preliminary clinical (ESCAT Levels II and III) or even preclinical (ESCAT Level IV) evidence are also considered. In these cases, we prioritize the allocation of patients to clinical trials, and one key feature of the MTBP is therefore to detect cases eligible for those open for recruitment across the connected centers, as detailed next. Of note, the CCE network hosts the Basket of Baskets (BoB), an European multi-arm phase II basket trial for advanced tumors selected according to prede ned molecular pro les (NCT03767075). Overall, 36% of the patients of the CCE cohort presented here were recommended for one of the BoB treatment arms available at the moment of the molecular tumor board discussion (Figure 2A). These were mostly associated with the use of immune checkpoint inhibitors in the presence of putative loss-of-function events in DNA damage repair genes, as estimated by the MTBP variant interpretation. However, the majority (60%) of these cases were not nally enrolled in the BoB trial, mostly due to the patient's clinical deterioration or subsequent screening failures. This further emphasizes the importance of deploying systems that can support an e cient and rapid trial recruitment at the point of clinical decision-making.
The MTBP retrieves the trials' eligibility criteria from in-house databases gathering up-to-date clinical, pathologic and molecular requisites. These requisites are stated following an ad-hoc syntax adapted to the growing complexity of cancer biomarkers, which are de ned by the presence (or absence) of a given combination of genomic alterations and/or genomic signatures. Trials database also includes prioritization rules in case that the allocation to multiple trials is possible, as well as variant interpretation nuances to be used by the MTBP analysis. One example of the latter is the evidence required for considering variants in a given actionable gene as functionally relevant; in general, we only match clinical trials with tumor variants whose effect is based on well-curated studies or bona de biological assumptions, but lower evidence such as bioinformatic predictions are also considered for emerging biomarkers associated with less characterized genes. Upfront agreement on these details enables the MTBP to re ne the actionability ags issued in the reports, which facilitates e cient discussions during the molecular tumor board meetings and increases the consistency of the clinical decision-making.
Importantly, the MTBP actionability ags can be used to automate the detection of other events of potential clinical relevance. For example, and in collaboration with the CCE genetic counseling task force, we have established uni ed criteria to ag germline variants requiring genetic specialist referral (Appendix 1). As a result, the MTBP issued genetic counseling alerts for 49 germline variants in 48 individuals (57±13 years), which represent 13% of the CCE cohort with paired tumor/normal samples sequencing available (with a cancer type distribution similar to that of the overall cohort, see Table 1). Of note, three (6%) of these variants showed a low variant allele frequency in the tumor sample, and thus they would have been not contemplated as of potential germline origin with tumor-only sequencing data, as per published criteria24. Genetic counselor review deemed these ndings as of clinical signi cance in all the cases, although close to half (44%) of the variant carriers did not meet personal criteria for clinical germline testing25,26 ( Figure 2B). Moreover, a considerable (18%) proportion of these pathogenic germline variants were found in genes not associated with the patient's index cancer, which further complicates their discovery via standard guidelines-directed genetic testing. Overall, these results illustrate the importance of the oncology setting for screening hereditary cancer susceptibility variants27, and the value of the MTBP to streamline that task.

The MTBP technology
The MTBP provides a single uni ed framework for sharing and harnessing NGS data across CCE centers.
De-identi ed patient clinical and pathological information is fetched from a centralized electronic case report forms system, while sequencing data les are retrieved from different institutional and external laboratories. Data transfer, storage and access is implemented by a set of technical measures in accordance with the European legal framework under compliance with data protection regulation. After data capture, the system triggers a number of pipelines for data integrity control, format harmonization and variant interpretation as appropriate. This ultimately creates the corresponding patient reports, which are immediately shared with the clinical investigators. The whole process is fully automated and thus performed in a negligible amount of time, which dramatically reduces the efforts required for case preparation. At present, our turnaround time from biopsy collection to sequencing report availability is less than 14 days, as required in patients whose clinical condition can rapidly deteriorate28.
The MTBP patient reports are HTML web-based documents accessible for the clinical investigators via a secure online platform2. These reports are discussed in weekly virtual molecular tumor board meetings, in which members of each CCE center connect from different locations and agree on clinical recommendations. As discussed before, genomic alterations are agged in the MTBP report according to prede ned expert actionability criteria, and all the results appear systematically organized in a userfriendly, readily interpretable view ( Figure 2B). In addition, further information and variant annotation details are accessible via interactive elements of the HTML report, which empowers an in-depth revision of the content and supporting evidence. Although the MTBP can also distribute summarized reports in PDF format, we believe that working with interactive data-rich documents is more appropriate in the context of academic medical centers, in which molecular tumor boards discuss more complex cases and serve as a venue for continued education in genomics-driven oncology. Of note, we observed a learning curve to use the MTBP system lasting for approximately 25 patients ( Figure 2C). After that, the amount of time devoted to discussing each case averaged less than three minutes, which is key for scaling the process to a large number of patients.
At the moment of writing this manuscript, the MTBP supports the interpretation of genomics data and has recently incorporated gene expression analysis. As one of the founding principles of CCE is to share new tools with the cancer research community, we have created an open version of the MTBP analytical framework available at http://mtbp.org (Supp Figure 4). This public resource supports a general interpretation of the gene variants uploaded by the user, which (although more narrowed than that deployed for CCE projects) may be of interest for researchers outside of our network. In order to support innovative clinical trials for next generation precision medicine, we are currently working on the incorporation of emerging technologies such as proteomics, ex-vivo drug screening and digital pathology into the MTBP. Ultimately, we envision the platform as a catalyzer for systems-based precision oncology strategies capable of integrating multiple levels of molecular and imaging data and inform treatment decisions throughout the patient's disease course.

Discussion
The MTBP implements an e cient process to support the biomarker-driven clinical decision-making at scale and creates a uni ed platform for developing new trials in a truly collaborative manner. As the complexity of cancer biomarkers continues to grow, automating the interpretation and reporting of sequencing results decreases the need for manual procedures and facilitates rapid, comprehensive and consistent clinical decision-making. In addition, the MTBP creates the infrastructure to systematically gather the molecular and clinical information in a "bio-repository" of data, which supports the discovery of new cancer biomarkers and insights for future trial designs. However, deploying the MTBP across the CCE network raised multiple challenges beyond the development of the platform itself, such as: (i) ensuring the interoperability with the information technology systems of each connected center; (ii) automating the retrieval of clinical and sequencing data provided by different facilities; (iii) developing user-friendly interfaces for distinct user types, such as medical practitioners, project managers and data analysts; (iv) coordinating the efforts to agree on variant interpretation criteria and actionability prioritization; and (v) creating the associated resources, such as a database with up-to-date information of the clinical trials open for recruitment across the network. These tasks require expertise from domains such as medical software regulation, cyber-security and front-end development, which is not easily available in the academic setting and thus creates needs for collaboration with industry partners. In conclusion, we believe that streamlining digital systems like the MTBP at the point of care is key to address current challenges of biomarker-driven precision oncology, but the success of these initiatives relies on the long-term investment needed to develop and maintain the technology.

Declarations
Pharma; research funding for Bayer & Novartis. Janne Lehtiö reports research funding from AstraZeneca, Novartis and GE healthcare, and is co-founder and shareholder of FenoMark Diagnostics Ab.

Supplementary Files
This is a list of supplementary les associated with this preprint. Click to download. SupplementaryInformation.pdf