Since the introduction of high throughput data generation methods, the medical field has become increasingly data-driven. Large biomedical studies often comprise sampling and data generation at different centers, requiring the data to be subsequently shared for joint analysis. Though data sharing has thus become crucial, the inherently private nature of medical data demands caution and elaborate sharing protocols are necessary to ensure patient privacy. This is particularly important when sharing genomics data and the associated medical metadata. The concise risks of privacy breaches and the methods used by adversaries vary for different data types. In this review, we will assess such risks and techniques in the context of genomic data, more precisely, three different data types: genomic sequences, transcriptomic and epigenomic data. We will further elaborate on currently used methods for privately sharing data along with different laws that are meant to protect individuals in case of a data leak. We will additionally introduce current research areas for private sharing of genomic data and give an outlook on the necessary as well as expected changes in the upcoming years.

Traditionally, genomic data (Goodwin et al., 2016[36]) refers to data holding information on the base sequence in an individual's genome. Here, we are going to extend this notion of genomics data to also incorporate transcriptomic as well as epigenomic data, which are closely associated with and influenced by the genomic sequence (Figure 1a(Fig. 1)). This serves as a more holistic assessment of potential privacy breaches when working with either category of data by evaluating how these different data types impact and relate to one another and how weak points may translate across categories.

In order to assess privacy issues that arise when sharing genomic data, there is a need to discuss the motivation behind sharing it in the first place (Figure 1b(Fig. 1)).

Data sharing is required during multi-party studies, where separate datasets have been collected in a decentralized fashion and subsequently have to be joined together. Further, sharing is required to improve the reproducibility of results by publicly (or on request) providing the dataset that was used in a particular analysis. A set of principles targeting the findability, accessibility, interoperability and reusability of data, particularly by machines, was introduced by Wilkinson et al. as the FAIR-principles (Wilkinson et al., 2016[78]). Other scenarios that require data sharing are the reusing of data, in which case datasets are shared for further use in order to answer questions other than those originally posed when collecting the data, or sharing for reduction of model bias. The latter is especially important when the data was collected from minority groups, in which case public accumulation can battle common bias problems in population-based modeling approaches (Mehrabi et al., 2019[54]).

A rather recent development that urges the distribution of genomic data is the increasing application of machine learning techniques (Libbrecht and Noble, 2015[49]). These approaches typically require a large amount of training data, especially in the context of deep learning (Eraslan et al., 2019[24]), as otherwise the model tends to bias towards local minima that generalize poorly on unseen samples. The demand for big data is particularly strong in genomics because of its immense feature space, and oftentimes exceeds the scope of a single study by far, prompting the collection of data from multiple studies.

Moreover, researchers (Brittain et al., 2017[10]; Dand et al., 2019[18]) and politicians (The White House, 2015[71]; BMBF, 2018[8]) alike are promoting the advent of personalized medicine to enable the development of treatment plans tailored to the individual. To provide the knowledge base for personalized medicine, a great wealth of biomedical data, including genomics data, will have to be generated and shared, given that the sampling will have to be conducted in a decentralized fashion to properly reflect human diversity and the data has to be accumulated for subsequent analysis. Given the highly private nature of the data, this naturally demands a thorough discussion on how to guarantee sufficient privacy compliance.

In light of the aforementioned reasons to share genomic data and before introducing concrete data sharing methods, the problems that arise from privacy breaches in the context of genomics must be elucidated. The root of these problems and the reason why genomic data privacy requires thorough discussion is the issue of subject re-identification. Identifying the individual which the data was taken from can potentially have severe consequences starting with social issues such as stigmatization, which is often observed when there is knowledge about the presence of certain risk alleles in an individual's genome, especially in the context of mental health issues (Ward et al., 2019[75]). Additionally, known preconditions and increased disease risks can negatively affect chances for employment or health insurance (Godard et al., 2003[34]). There have also been cases where adopted children or children received by use of sperm donations identified their biological parent(s) due to publicly available genomic information (Erlich and Narayanan, 2014[26]). Another issue is posed by the longevity of genomic data: given the heritability of genomic and epigenomic alterations, re-identification does not only pose a risk to the individual itself but also to close relatives (Ayday et al., 2015[3]; Oprisanu et al., 2019[58]). While all living relatives can - in theory - be asked to give their consent to the collection of the data in addition to the consent given by the individual, privacy breaches may also impact unborn relatives or underage relatives whose consent cannot be given.

The risk of re-identification and the corresponding instantiated attack methods are different for the previously introduced data types. We will first discuss these with respect to sequence data and we will see that most re-identification risks associated with transcriptomic and epigenomic data fall back onto techniques that infer the underlying sequence information.

Many countries have developed sophisticated laws and guidelines to protect individuals from misuse of their genomic information (Figure 1d(Fig. 1)). However, the wording is often ambiguous towards the different types of genomic data and has therefore prompted researchers to question their validity in particular situations.

For example, the German “Gendiagnostikgesetz” prohibits discrimination of citizens based on their genetic characteristics. Genetic characteristics are here defined as inherited or between conception and birth acquired hereditary information of human origin (Deutscher Bundestag, 2009[19]). This unambiguously covers genomic sequence information as it is present at birth. It is unclear, however, if this also covers quantitative gene expression information which can be affected by the environment and not only by underlying sequence information. Further, if it covers epigenomic data, which is often altered by environmental factors and therefore can be lost or acquired after birth and is not necessarily inherited. Lastly, if it also refers to genomic sequence information that has changed after birth, e.g., mutations due to external stimuli such as UV-radiation. The law also states that it does not apply to genetic examination and analyses and the handling of genetic samples and data for the purpose of research. The Council of Europe had previously issued a convention in 1997 that prohibits “any form of discrimination against a person on grounds of his or her genetic heritage”, where the term genetic heritage is not defined further. This convention was signed and ratified by 29 countries (Council of Europe, 1997[16]). In the United States, a similar effort has been made to protect citizens from discrimination in health insurance and employment based on genetic information by introducing the Genetic Information Nondiscrimination Act (GINA) (U.S. Equal Employment Opportunity Commission, 2008[73]). Genetic information is here defined as information on an individual's genetic test, genetic tests of their family members and diseases or disorders that have manifested in family members of the individual. A genetic test is defined as “an analysis of human DNA, RNA, chromosomes, proteins or metabolites, that detects genotypes, mutations, or chromosomal changes”. This definition leaves the same questions unanswered as above. Additionally, it explicitly excludes the U.S. military from the list of employers that are prohibited to use genetic information as well as any employer with less than 15 employees.

In addition to GINA, disclosure of health information is regularized by the Health Insurance Portability and Accountability Act (HIPAA). It provides three standards for the disclosure of patient health data that do not require authorization by the patient. Those three standards are the Safe Harbor standard, the Limited Dataset standard, and the statistical standard (El Emam, 2011[23]). Safe Harbor regulates the de-identification of health data by removing 18 different identifying elements, among those the name, certain geographic information, all elements of dates except for the year, phone and fax numbers, e-mail addresses, social security numbers, and many more (El Emam, 2011[23]). The Limited Dataset standard only removes 16 potential identifiers but additionally requests a data sharing agreement between data custodian and data recipient and the statistical standard requires expert evaluation and classification of the de-identification risk as very small (El Emam, 2011[23]). Given its clarity and simplicity, Safe Harbor is often referred to for data de-identification. However, as pointed out by El Emam, not only does it often result in the removal of information that could have been useful for the data evaluation, it also does not sufficiently ensure the protection of the individual with regard to re-identification. In this context, there is special emphasis on the lack of protection through genetic data, such as SNPs, longitudinal data, widely used diagnosis codes, sampling size, and free-form text. A more detailed elaboration can be found in El Emam (2011[23]).

In 2018, major advancements have been made in the EU with respect to data protection due to the release of the General Data Protection Regulation, GDPR (European Union, 2018[30]; Shabani and Borry, 2018[66]). It applies in all EU member states and aims to unify data protection across countries. GDPR overcomes the ambiguous definition of genetic data as outlined above by defining it as “personal data relating to the inherited or acquired genetic characteristics of a natural person which give unique information about the physiology or the health of that natural person and which result, in particular, from an analysis of a biological sample from the natural person in question” (“Recital 34 - Genetic data - GDPR.eu,” 2018[33]). This definition includes the wide range of modern genomics data types and therefore offers more thorough protection from open sharing and un-consented processing than the regulations discussed above. While passed in the EU, the law applies worldwide if the data that is processed was sampled from a citizen of an EU member state.

Additional EU guidelines have been introduced recently that specifically aim to protect EU citizens from threats posed by Artificial Intelligence (AI) systems. The “Assessment List for Trustworthy Artificial Intelligence” (ALTAI) was published by the European Commission in 2020 as a guideline for the development of trustworthy AI (European Commission, 2020[28]). Early in 2021, the European Commission additionally published a draft of the Artificial Intelligence Act (AIA) (European Commission, 2021[29]). Subject to this act is any provider of AI applications worldwide, if those applications are used by EU citizens. It categorizes AI systems into different categories of threat, connected with strict obligations that have to be met by the provider. These regulations also address AI used in the health context, therefore including genomic data. While the motivation behind these guidelines and regulations is most reasonable, there are some concerns regarding the implications they might have on data privacy. This is particularly important because they demand the models to be fair and unbiased by using representative, non-discriminatory and complete training, validation and test datasets. However, in order to test a model's fulfillment of these requirements, the inspecting authority is likely to need access to the datasets, which - e.g., in the context of genomic data - are often highly private.

When assessing privacy risks, it is always essential to weigh the risk of an individual against the possible gain that is associated with the vulnerable position the individual finds themself in. The group of individuals that was - and in many scenarios still is - the main target groups for genomic data collection are those that have a personal reason to participate in the respective studies, e.g., a difficult-to-treat or poorly researched disease (Esplin et al., 2014[27]). For these individuals, the potential gain that comes from participating in these studies often substantially outweighs the risk of re-identification. But the focus of the target group appears to gradually shift, people have started to have their genome sequenced out of curiosity rather than acute medical reasons and the field of personalized medicine advocates genomic data collection to become part of a medical care routine (Brittain et al., 2017[10]; Suwinski et al., 2019[68]). Therefore, the group at risk of re-identification is bound to change from those that have a high benefit-risk ratio to a more heterogeneous population. Additionally, the risk of cross-referencing genomic data and metadata to narrow down the set of individuals a data instance potentially belongs to, is likely to increase due to oversharing of personal information on social media, be it voluntary or involuntary. This can start with information as subtle as height, sex, and weight which can be inferred from pictures, geotags, and dates that put an individual into close spatial and temporal proximity of the conduction of a given study, and it can go as far as people openly sharing their health status or study participation. This can be expected to substantially increase the risk of re-identification and it is an issue that has lacked thorough attention in prior risk-evaluation strategies.

Current data sharing techniques come at different levels of security, as is illustrated in Figure 1e(Fig. 1). As discussed above, some data types are often shared publicly in plain text without restrictions other than de-identified sample descriptors. This is the case when the risk of re-identification for the participants is considered minimal. Genomic data with increased re-identification risk such as sequence data or data that gives direct information on partial sequences (reads, SNPs) are published under controlled access. In this scenario, the legitimacy of an access request is evaluated based on the applicant's personal information and the research project the data is intended to be used for. The use of the data also often comes with a series of constraints that regard a safe storing location, no sharing and no re-identification attempts. While there is still a lack of true oversight with respect to whether or not the data is shared after downloading, another controlled access strategy is to not allow the data to be downloaded but instead run protocolled queries on the data and only retrieve the results. This may severely restrict the flexibility with which the data can be analyzed (Erlich and Narayanan, 2014[26]).

To decrease the risk of subject identification, be it in publicly shared or controlled-access data, efforts are made to de-identify the data, i.e., to remove information or reduce its granularity such that identification of the individual becomes very unlikely. A common set of guidelines for de-identification is the mentioned Safe Harbor standard included in the HIPAA Privacy Rule.

Gürsoy et al. developed a method that sanitizes the raw reads underlying gene expression data such that sharing with reduced re-identification risk is possible, while keeping the necessary data manipulation minimal (Gürsoy et al., 2020[37]). They achieve this by transforming the original BAM file into a sanitized file, where information that reveals the presence of a variant (SNPs, insertions, deletions) is masked. For instance, information on variants as it is contained in reads is removed by replacing the called base at the site of the variant with that present in the reference genome. The true values of the sanitized elements are stored in a separate file which is meant to be under controlled access. This allows for sharing of the sanitized data while being able to reconstruct the original if access to the additional file is granted, though no formal or statistical guarantees on privacy are provided.

Classical encryption approaches have been leveraged as well to enable secure sharing of genomics data. One example is the crypt4gh file format introduced in 2019 by the Global Alliance for Genomics and Health (GA4GH) (“GA4GH File Encryption Standard,” 2019[32]). The format allows the genomic data to be encrypted while in storage, in transit, during reading and writing. It uses symmetric encryption as well as public-key encryption, it is confidential in the sense that it is only readable by holders of a secret decryption key, but it does not obscure the length of the file. The data is stored in blocks, the integrity of which is ensured using message authentication codes, however, blocks can be rearranged, removed, or added. Authentication of files encrypted with this method is not provided. Though as with any system that relies on secret decryption keys, privacy is lost in the case of a system breach that results in an untrusted party acquiring the key.

Others have explored fully homomorphic encryption (FHE) to securely operate on genomic data. FHE allows to conduct computations on the genomic data while it remains in its encrypted state, receiving the encrypted results and decrypting them with a personal key (Erlich and Narayanan, 2014[26]). This allows for secure computations on cloud systems even if the system itself is not. While this approach was long assumed too computationally expensive to be reasonable in the context of genomic data, Blatt et al. recently presented an improvement in run-time of FHE for Genome Wide Association Studies (GWAS) by introducing parallelization and crypto-engineering optimizations, which allegedly outperforms secure multiparty computations (SMPC) (Blatt et al., 2020[7]). SMPC is another approach to secure computation, in which two or more parties that hold private data can compare and perform computations on the data without ever revealing the data itself to the other party or another third party (Erlich and Narayanan, 2014[26]). While a prominent point of criticism with SMPC models is the oftentimes extensive communication required between computing parties, Cho et al. demonstrated an approach for SMPC in GWAS in which run-time scaled linearly to the number of samples and was reduced to 80 days for 1 million individuals and 500,000 SNPs (Cho et al., 2018[14]). Though this is still a substantial amount of time, efforts such as this have worked on optimizing the procedure in the past years. However, in the case of Cho et al., the privacy guarantee only holds for the semi-honest security model, in which participants are assumed to not deviate from the conduction protocol.

Besides homomorphic encryption and multi-party computing, there are also hardware-based approaches to handling sensitive information such as genomic data. An example is the Intel Software Guard Extension (SGX), which has also been used in combination with homomorphic encryption on GWAS data (Sadat et al., 2019[64]). While such techniques based on SGX can also be used to protect model and/or data (Hanzlik et al., 2021[39]), implementations of SGX however have been troubled with serious security issues themselves (Van Bulck et al., 2018[74]; Lipp et al., 2021[51]).

Another widely applied concept to allow privacy preserving data analysis is that of Differential Privacy (DP) (Dwork and Roth, 2014[20]). The idea behind DP is to allow private data analysis by assuring that the addition or removal of a subject to or from the dataset does not significantly alter potential query results and therefore does not disclose whether or not the individual is part of the dataset. To achieve this, different levels of noise have to be added to the data before its release, where the amount of noise necessary increases when the sample size of a dataset decreases. The privacy loss of the noised dataset is quantified using the epsilon parameter, where a value of 0 indicates total privacy, though decreasing values come with increasing added noise and therefore less utility. In contrast to security measures, cryptography or privacy heuristics, DP comes with strong guarantees that are not susceptible to misuse of the system or typical breaches. Although privacy preserving and statistical analyses do not follow an adverse goal, today's techniques are often affected by a reduced utility of the analysis. For an exhaustive review of privacy-enhancing technologies in the genomic context, please refer to the work done by Mittos et al. (2019[55]).

Putting the risks of re-identification using genomic sequence data into perspective, data privacy is a concern that needs to be taken seriously. At this point in time, while all the above-mentioned methods are eligible threats, the costs involved and the expertise needed are still rather high and one could wonder, how realistic the threat really is. But as always in the field of data security and privacy, it is not only important to assess what is but also what potentially will be. While today, the motivation to acquire the genomic sequence of most people is considerably low given the entailed costs and effort, further decrease in sequencing price, increased insight into the genome and better computing performance are likely to make it more attractive in the future. The future in mind, present day scientists are required to address the issues of genomic data privacy to assure responsible research. This entails active communication with lawmakers to provide non-discrimination laws that protect study participants in the case of data leakage and which are up to date with the science. Further, increased collaboration of data security and privacy researchers with life scientists is essential to develop privacy-preserving data sharing techniques that are specifically tailored to genomics data in the near future. In this context, we can expect an increased necessity for bioinformaticians, computational biologists, biomathematicians, and others to optimally communicate the needs of the genomics community to computer scientists, in order to enable easier, yet secure sharing of genomic data.

This work was supported by the HGF Helmholtz AI grant Pro-Gene-Gen (ZT-I-PF-5-23), the HGF Incubator grant sparse2big (ZT-I-0007), HGF Incubator grant “Trustworthy Federated Data Analytics (TFDA)" (ZT-I-OO1 4), by NaFoUniMedCovid19 (FKZ: 01KX2021, project acronym COVIM), by the German Research Foundation (DFG) (INST 37/1049-1, INST 216/981-1, INST 257/605-1, INST 269/768-1, INST 217/988-1, INST 217/577-1, INST 217/1011-1, INST 217/1017-1 and INST 217/1029-1), ImmunoSep (grant 84722) and the BMBF-funded excellence project Diet-Body-Brain (DietBB) (grant 01EA1809A).

The figure was created with BioRender.com.

The authors declare that they have no conflict of interest.