Towards standardization guidelines for in silico approaches in personalized medicine

1 Introduction

A key strategic objective of EU-STANDS4PM is to engage with relevant stakeholders. Through specific awareness actions, including workshops, the project provides a forum to assess strategies for health data integration (such as genetic, expression, proteomics, demographic, clinical, and lifestyle exposures) as well as data-driven in silico modelling^[1] [1] approaches. This proactive networking process is central to the major deliverable of EU-STANDS4PM, which is to jointly develop universal standards as well as guidelines and recommendations for in silico methodologies relevant for personalized medicine in a concerted action.

To initiate this process, EU-STANDS4PM consulted the COMBINE (Computational Modelling in Biology Network) community to address crucial requirements with respect to the development of data and model standards as well as data integration tasks in research and clinic, including ethical and legal aspects. These areas were discussed at the annual meeting of COMBINE in moderated round table workshops covering six central topics presented below and that conclude in a first set of recommendations to key actors.

2 Standards as drivers for reproducibility and data quality

As a major requirement for any scientific result reproducibility ensures high quality of investigations and vice versa high-quality investigations ensure reproducibility. Particularly, for personalized medicine, reproducibility of clinical research and development ensures patient safety. Even more than in classical modelling applications, revision-safe and reproducible documentation of studies is necessary when using modelling as a tool in clinical research settings. Standards may not be necessary to replicate a result, but they are indeed the drivers for reproducibility and they ensure higher quality of data, thus increasing trust in the scientific findings. Most standards in systems medicine are so-called de facto standards, meaning they are not legally binding. However, regulations for medical treatments are backed up by law and documentation, for example, in a hospital. They must be revision safe. This gap needs to be bridged and hence existing standards must be moved to the level of de jure standards, especially for cases that require revision safety (i.e. ensuring that all data and software code is being archived in a verifiable and transparent manner that allows to obtain the original, unchanged data or software code at later points in time).

To move systems medical standards to the level of de jure standards, it will be necessary to have them approved by formal authorities like the International Organization for Standardization (ISO). These organizations offer a critical and thorough assessment of all elements of a standard before release. Particularly, for the implementation of virtual studies in the clinic – that is the investigation in a biomedical system by means of modelling and simulation [2] – it is indispensable to ensure traceability of all analysis steps, boundary conditions, and assumptions through proper standardization. A first step into formalizing systems medicine standards has already been taken with the work of EU-STANDS4PM in different ISO and German Institute for Standardization (DIN) committees (see section “Community and Formal Standards” below).

Today virtual studies already predict disease progression [3], [4], support decision making [5], enable cross-validation of possible treatment outcomes [6] and are used for educational purposes [7]. They typically consist of (i) one or more models, (ii) one or more simulations, and (iii) a collection of results. These three ingredients need to be well-documented and each component must be tested for correctness. Reproducibility then requires standard formats to represent the data, detailed descriptions following the Good Scientific Practices described in Minimum Information Guidelines, and semantic annotations [8], [9]. The computational biology community has already developed standards for all parts of a typical virtual study and the authors are convinced that these well-established COMBINE standards [9] shall be thoroughly evaluated for use in predictive in silico models in personalized medicine. Equally important is the correctness of the software code to run a computational model. In addition, certification procedures, usually complex and time consuming, will be necessary for any model to be run in a clinical setting. One step into this direction, are efforts by regulatory authorities, such as the US Food and Drug Administration (FDA) and industry, to advance the use of computational modelling as medical advices, including the development of the respective standards [10], [11].

Finally, standards and standard operating procedures are necessary for the definition and execution of software pipelines; this applies in particular to complex simulations. Software pipelines, for example built-in workflow systems like Galaxy [12], can in themselves be considered software products, requiring similar procedures as for the model code and the simulation software. Furthermore, the COMBINE Archive format is the one development to allow for the exchange of multi-file experiments [13] and first applications of eNotebooks and docker have proven successful [14], [15], [16].

In summary, standards are needed to encode data, model(s), simulation, and results – but furthermore, standards need to be evaluated for method development and protocols, for documents in general and for meta-data encoding. Although many community projects and their efforts in developing standards are of considerable value for personalized medicine, satisfying the strict requirements of clinical evaluation of security and reporting guidelines are remaining and high hurdles.

3 Community and formal standards

Standards provide a specific set of rules, guidelines and characteristics to ensure that products, processes and services (including data and models) fit the individual purpose of the users and consumers in the best possible way. As such standards record in a comprehensible manner what information users can expect in a data set and, at the same time, specify which rules, requirements or conditions should or shall be followed. In the context of personal/patient-derived data in a clinical setting, standards are needed to ensure data security, quality and availability for any data-driven application. However, especially in the life sciences and in clinical research, there are still many challenges associated with defining what a standard really is and there are basically two independent worlds developing formal and community standards. In the following section, we will analyse the major differences, and briefly, discuss the challenges associated with bridging these two systems.

4 Pitfalls in developing and harmonizing standards

The development of standards in the life sciences, especially for personalized medicine, is a challenging task. First of all, a substantial amount of time and effort is needed to define a standard for a certain purpose and field, as it should cover not only one, but many potential use cases. Such an investment of time and resources is demanding, if possible at all, for researchers who deal with standardization as a side job. On the other hand, and if the necessary resources are available, chances are that the development of a standard is driven distinctively by the researchers’ genuine scientific work. This can lead to a biased standard tailored towards a specific topic, process or product. Thus, not necessarily the best – or most appropriate for the scientific community at large – standardization concept wins, but the one with the most supporting resources. This situation provides a competitive advantage for those stakeholders involved in the development of the winning concept. Sometimes there are even competing standards, as for example, in the field of life science microscopy imaging with a whole range of different, mostly proprietary standards [21]. A more recent example is the development of standards for genome compression [22], [23]. In the most extreme cases, such a competition can lead to an overrun of existing and well-established standards by newly developed ones.

A lack of maintainability adds to the difficult situation. Standards development (and their implementation in software tools) requires long-term available resources that are often not present given the current system of research funding (see above). Standards need to be maintained in a sustainable manner, also beyond the first release of a standard; otherwise, they become outdated. The additional problem of competing standards is especially crucial for the field of personalized medicine. Here, the obligation of long-term storage of patient data (and data derived from the primary patient data) makes sustainability a prerequisite.

The long development time for most standards (up to 5 years through the ISO route) may even lead to the preposterous situation that a standard can already be outdated once released. On the other hand, especially in evolving and modern scientific fields such as personalized medicine, a standard developed too early, based on technological methods that are not mature enough, might not cover all major aspects. Thus, the timing of standardization is absolutely crucial.

A central problem when developing standards is acceptance by the whole research community. Developed standards become useless if not known to the community or not properly adopted by software tools, researchers and/or clinicians. Thus, it is crucial for the acceptance that a standard is drafted by a representative part of the community of the corresponding domain, and not by single individuals or institutions that would like to promote their own workflows or tools. That ensures that a standard reflects the best practice in the corresponding domain. However, even when a standard is developed by a representative part of the community that is dedicated to put efforts into standardization, it still might be unknown to the majority of the scientific community. Consequently, standards have to be promoted in the communities, and they have to be developed close to existing workflows and data structures models to enable simple implementation procedures, ensuring a wide-spread adoption of the standard. For the same reasons, it is crucial that regulatory and governance bodies releasing and monitoring regulatory standards work closely with the scientific communities when establishing such official and mandatory standards. Moreover, over-standardization should be avoided to not hinder the dynamic development in a scientific field and hence the loss of flexibility therein. This is of crucial importance in personalized medicine, as here the field is quickly developing while, at the same time, highly regulated by authorities.

5 Standards relevant for personalized medicine

The future development of personalized medicine is dependent on an exchange of data from different sources. Patients will benefit from computational models that predict treatment outcomes. For example, models of a specific molecular characterization can be tested based on broad exchange of data from different clinical studies. This will allow comprehensive comparison of treatment outcomes under different settings and predictions for an optimal individual and personalized therapy. Figure 1 illustrates a typical workflow followed in personalized medicine starting with the clinical question, followed by identification, access, and harmonization of relevant data, the development of data model(s), the model validation and finally application in a clinical setting.

Figure 1:

A typical workflow for personalized medicine. A personalized medicine approach to improve patient health typically starts with identifying which aspect of health that is to be addressed and modelled, for example prediction of disease, prediction of severity of a particular disease or prediction of response to treatment. The next step is typically identification of relevant data sources, which can be of many different types, as illustrated. These data often have to be harmonized, a task that is made easier whether common data standards have been used. Once this has been completed, modelling to predict the clinically relevant state takes place. Usually the models then have to be validated in an independent setting. Once that has been completed, the models can be used in a clinical environment to help improve patient health. Picture source (licence free): www.pixabay.com.

A successful interchange of data, whether sensitive/personal or not, is highly dependent on the standardization of data sources and encoding of models. Different types of data (e.g. genetic, expression, proteomics, demographic, clinical, and lifestyle exposures) can be of relevance for personalized medicine. Data standards already exist for many types of data [24] and are widely used (Table 2); while for other types of data, there are very few established standards. In the following section, we will briefly discuss the latter ones and highlight specific challenges.

6 Integration of clinical and research data

The integration of clinical and research data, while essential, is not trivial. Challenges include establishing semantically consistent disease annotations and medical vocabulary, handling different types of patient populations and overcoming highly diverse registration procedures of measurements and interventions. Finally, different legal and ethical rules apply to clinical and research data in European countries (and beyond). While some of the problems arise due to differences inherent in the data capture setting (patient populations are inherently different from research populations, research projects can allot more resources to information gathering that does not directly affect patient treatment), most of the challenges are similar to any data integration effort between different medical systems, be they clinical or research based.

For the purposes of data integration, the distinction between research and clinical data is narrowing, as systematized data collection and knowledge generation are becoming a characteristic not only of research but also of clinical practice [36], [37].

As clinical data become more systematized, digitized, and linkable, researchers are confronted with well-known data integration problems when working with clinical and research data. One difference in integrating these different data may be that clinical care data are very complex and less harmonized than most project/research-generated data, since the data sets are larger and patients are not stratified, thus heterogeneous, and often multi-morbid. Data can be integrated in different ways, and with different levels of identifiability.

7 Using patient-derived data for personalized medicine: legal and ethical aspects

Legal and ethical governance of personalized medicine at European level is composed of a patchwork of international, regional, and national laws, as well as non-binding recommendations, that seek to protect patients from breaches of their privacy and confidentiality, and from discrimination on the basis of health data [39]. This spans international bioethics treaties from the Council of Europe (CoE), namely the Biomedicine Convention [40], and non-binding recommendations on bioethics drafted by CoE and United Nations Educational, Scientific, and Cultural Organization (UNESCO) [41]. At European (EU) level, the European Union General Data Protection Regulation (GDPR) entered into force in 2018 and has important implications for the processing of personal data in treatment and research contexts [42]. However, this legal and ethical framework is subject to several challenges. Bioethical and legal norms were established post World War II with research on human subjects in mind, not computational models, meaning that regulations often do not fit the big data landscape. Furthermore, international treaties, including the GDPR [43], leave significant discretion to national legislatures, resulting in varied implementation among member states and potential legal uncertainty. In the case of CoE norms, there are limited means of enforcement, in comparison with GDPR where breaches can result in significant fines.

Regarding the development of in silico models, there are several challenges. Firstly, while the GDPR has direct effect in all EU member states, it also provides for national variations. Crucially for researchers, Article 89 GDPR allows member states to enact derogations from the rights referred to in Articles 15, 16, 18, and 21 when processing personal data for scientific research purposes. This can lead to national variances that have implications for cross-border research collaboration. In other words, despite the aim of harmonization, national data protection legislation continues to differ among member states and may place varying requirements on researchers who inevitably conduct research across borders.

In silico modelling also raises concerns regarding the type of information that must be supplied to research subjects to meet GDPR consent requirements. This can be a challenge for in silico models where the research hypothesis is unclear. Another concern surrounds the principle of data minimization (data should be limited to what is necessary) as enshrined by Article 5(1)(c) GDPR and the level of anonymization required for data to avoid falling under GDPR.

From an ethical perspective, willingness to donate data, specifically whether the public has adequate knowledge of the possible implications, is of concern. The familial implications of donating one’s genetic data for research/other purposes also require study.

8 Conclusions

A key output of the discussed workshop topics was a summary of challenges associated with the implementation of data-driven in silico methodologies in personalized medicine and clinical practice. In this section, we highlight these challenges and provide a set of recommendations directed towards different key actors.

EU-STANDS4PM consortium

Rolf Apweiler¹, Stephan Beck², Catherine Bjerre Collin³, Niklas Blomberg¹, Søren Brunak³, Tom Gebhardt⁴ Eugenijus Gefenas⁴, Martin Golebiewski⁶, Kalle Günther⁷, Mette Hartlev³, Vincent Jaddoe⁸, Marc Kirschner⁹, Ingrid Kockum¹⁰, Sylvia Krobitsch⁹, Lars Küpfer¹¹, Stamatina Liosi², Vilma Lukaseviciene⁵, Ali Manouchehrinia¹⁰, Arshiya Merchant¹, Neha Mishra¹², Heike Moser¹³, Miranda Mourby¹⁴, Wolfgang Müller⁵, Flora Musuamba Tshinanu¹⁵, Katharina Eva Ó Cathaoir³, Uwe Oelmüller⁷, Tito Poli¹⁶, Philip Rosenstiel¹², Dagmar Waltemath¹⁷, Olaf Wolkenhauer⁴, Amonida Zadissa¹

¹ European Bioinformatics Institute (EBI-ELIXIR), United Kingdom

² University College London, United Kingdom

³ University of Copenhagen, Denmark

⁴University of Rostock, Germany

⁵ Vilnius University, Lithuania

⁶ HITS gGmbH, Germany

⁷ Qiagen GmbH, Germany

⁸ Erasmus University Rotterdam, The Netherlands

⁹ Forschungszentrum Jülich GmbH, Project Management Jülich, Germany

¹⁰ Karolinska Institutet, Sweden

¹¹ Bayer AG, Germany

¹² University of Kiel, Germany

¹³ German Institute for Standardization, Germany

¹⁴ Univeristy of Oxford, United Kingdom

¹⁵ Federal Agency for Medicines and Health Products, Belgium

¹⁶ University of Parma, Italy

¹⁷ Medical Informatics Laboratory, Institute for Community Medicine, University Medicine Greifswald, Germany