Hyunghoon Cho

Large Language Models (LLMs) have demonstrated significant potential in medicine, with many studies adapting them through continued pretraining or fine-tuning on medical data. However, a key question remains: to what extent do LLMs memorize medical training data—that is, recall or regenerate content seen during continued pretraining or fine-tuning. In this work, we investigate memorization of LLMs in medicine, assessing its prevalence (frequency), characteristics (what is memorized), volume (how much), and potential downstream impacts. We systematically analyze common adaptation scenarios: (1) continued pretraining on medical corpora, (2) fine-tuning on standard medical benchmarks, and (3) fine-tuning on real-world clinical data, including over 13,000 unique inpatient records from Yale New Haven Health System. The results demonstrate that memorization is prevalent and significantly higher than that in the general domain. Memorization has distinct characteristics during continued pretraining and fine-tuning, and it is persistent: up to 87% of content memorized during continued pretraining remains after fine-tuning. Memorization can be categorized into three types: beneficial (e.g., accurate recall of clinical guidelines), uninformative (e.g., templated language), and harmful (e.g., sensitive clinical content). We offer practical recommendations to facilitate beneficial memorization, minimize uninformative memorization, and mitigate harmful memorization to protect patient privacy and improve medical utility.

Genotype imputation servers enable researchers with limited resources to extract valuable insights from their data with enhanced accuracy and ease. However, the utility of these services is limited for those with sensitive study cohorts or those in restrictive regulatory environments owing to data privacy concerns. Although privacy-preserving analysis tools have been developed to broaden access to these servers, none of the existing methods support haplotype phasing, a critical component of the imputation workflow. The complexity of phasing algorithms poses a significant challenge in maintaining practical performance under privacy constraints. Here, we introduce TX-Phase, a secure haplotype phasing method based on the framework of trusted execution environments (TEEs). TX-Phase allows users’ private genomic data to be phased while ensuring data confidentiality and integrity of the computation. We introduce novel data-oblivious algorithmic techniques based on compressed reference panels and dynamic fixed-point arithmetic that comprehensively mitigate side-channel leakages in TEEs to provide robust protection of users’ genomic data throughout the analysis. Our experiments on a range of data sets from the UK Biobank and Haplotype Reference Consortium demonstrate the state-of-the-art phasing accuracy and practical runtimes of TX-Phase. Our work enables secure phasing of private genomes, opening access to large reference genomic data sets for a broader scientific community.

Sharing data across institutions for genome-wide association studies (GWAS) would enhance the discovery of genetic variation linked to health and disease^1,2. However, existing data-sharing regulations limit the scope of such collaborations³. Although cryptographic tools for secure computation promise to enable collaborative analysis with formal privacy guarantees, existing approaches either are computationally impractical or do not implement current state-of-the-art methods^4,5,6. We introduce secure federated genome-wide association studies (SF-GWAS), a combination of secure computation frameworks and distributed algorithms that empowers efficient and accurate GWAS on private data held by multiple entities while ensuring data confidentiality. SF-GWAS supports widely used GWAS pipelines based on principal-component analysis or linear mixed models. We demonstrate the accuracy and practical runtimes of SF-GWAS on five datasets, including a UK Biobank cohort of 410,000 individuals, showcasing an order-of-magnitude improvement in runtime compared to previous methods. Our work enables secure collaborative genomic studies at unprecedented scale.

The introduction of trusted execution environments (TEEs), such as secure enclaves provided by the Intel SGX technology has enabled secure and privacy-preserving computation on the cloud. The stringent resource limitations, such as memory constraints, required by some TEEs necessitates the development of computational approaches with reduced memory usage, such as sketching. One example is the SkSES method for GWAS on a cohort of case and control samples from multiple institutions, which identifies the most significant SNPs in a privacy-preserving manner without disclosing sensitive genotype information to other institutions or the cloud service provider. Here we show how to improve the performance of SkSES on large datasets by augmenting it with a learning-augmented approach. Specifically, we show how individual institutions can perform smaller scale GWAS on their own datasets and identify two sets of variants according to certain criteria, which are then used to guide the sketching process to more accurately identify significant variants over the collective dataset. The new method achieves up to 40% accuracy gain compared to the original SkSES method under the same memory constraints on datasets we tested on. The code is available at https://github.com/alreadydone/sgx-genome-variants-search. Contact cenk.sahinalp{at}nih.gov

Genotype imputation servers enable researchers with limited resources to extract valuable insights from their data with enhanced accuracy and ease. However, the utility of these services is limited for those with sensitive study cohorts or those in restrictive regulatory environments due to data privacy concerns. Although privacy-preserving analysis tools have been developed to broaden access to these servers, none of the existing methods support haplotype phasing, a critical component of the imputation workflow. The complexity of phasing algorithms poses a significant challenge in maintaining practical performance under privacy constraints. Here, we introduce TX-Phase, a secure haplotype phasing method based on the framework of Trusted Execution Environments (TEEs). TX-Phase allows users’ private genomic data to be phased while ensuring data confidentiality and integrity of the computation. We introduce novel data-oblivious algorithmic techniques based on compressed reference panels and dynamic fixed-point arithmetic that comprehensively mitigate side-channel leakages in TEEs to provide robust protection of users’ genomic data throughout the analysis. Our experiments on a range of datasets from the UK Biobank and Haplotype Reference Consortium demonstrate the state-of-the-art phasing accuracy and practical runtimes of TX-Phase. Our work enables secure phasing of private genomes, opening access to large reference genomic datasets for a broader scientific community.

The rapidly growing scale and variety of biomedical data repositories raise important privacy concerns. Conventional frameworks for collecting and sharing human subject data offer limited privacy protection, often necessitating the creation of data silos. Privacy-enhancing technologies (PETs) promise to safeguard these data and broaden their usage by providing means to share and analyze sensitive data while protecting privacy. Here, we review prominent PETs and illustrate their role in advancing biomedicine. We describe key use cases of PETs and their latest technical advances and highlight recent applications of PETs in a range of biomedical domains. We conclude by discussing outstanding challenges and social considerations that need to be addressed to facilitate a broader adoption of PETs in biomedical data science.

Finding relatives within a study cohort is a necessary step in many genomic studies. However, when the cohort is distributed across multiple entities subject to data-sharing restrictions, performing this step often becomes infeasible. Developing a privacy-preserving solution for this task is challenging owing to the burden of estimating kinship between all the pairs of individuals across data sets. We introduce SF-Relate, a practical and secure federated algorithm for identifying genetic relatives across data silos. SF-Relate vastly reduces the number of individual pairs to compare while maintaining accurate detection through a novel locality-sensitive hashing (LSH) approach. We assign individuals who are likely to be related together into buckets and then test relationships only between individuals in matching buckets across parties. To this end, we construct an effective hash function that captures identity-by-descent (IBD) segments in genetic sequences, which, along with a new bucketing strategy, enable accurate and practical private relative detection. To guarantee privacy, we introduce an efficient algorithm based on multiparty homomorphic encryption (MHE) to allow data holders to cooperatively compute the relatedness coefficients between individuals and to further classify their degrees of relatedness, all without sharing any private data. We demonstrate the accuracy and practical runtimes of SF-Relate on the UK Biobank andAll of Usdata sets. On a data set of 200,000 individuals split between two parties, SF-Relate detects 97% of third-degree or closer relatives within 15 h of runtime. Our work enables secure identification of relatives across large-scale genomic data sets.

Finding related individuals in genomic datasets is a necessary step in many genetic analysis workflows and has broader societal value as a tool for retrieving lost relatives. However, detecting such relationships is often infeasible when the dataset is distributed across multiple entities due to privacy concerns. Although cryptographic techniques for secure computation offer ways to jointly analyze distributed datasets with privacy guarantees, the sheer computational burden of operations required for identifying kinship, such as pairwise sequence comparison of all individuals across large datasets, presents key challenges to developing a practical privacy-preserving solution. We introduce SF-Relate, a secure federated algorithm for identifying genetic relatives across data silos (Fig. 1) that scales efficiently to large datasets that include hundreds of thousands of individuals. We leverage the key insight that the number of individual pairs to compare can be vastly reduced, while maintaining accurate detection, through innovative locality-sensitive hashing of individuals who are likely to be related together into buckets and then testing relationships only between individuals in corresponding buckets across parties. To this end, we construct an effective hash function that captures identity-by-descent (IBD) segments in genetic sequences, which, along with our novel micro-bucketing strategy is key to achieving accurate and practical private relative detection. To guarantee privacy, we devise an efficient algorithm based on multiparty homomorphic encryption (MHE) to allow the parties to cooperatively compute the relatedness coefficients between pairs of individuals, and to further classify their degrees of relatedness, all without sharing any private data. We demonstrate the accuracy and practical runtimes of SF-Relate on real genomic datasets of varying sizes, from the UK Biobank and All of Us datasets. On the largest dataset of 200K individuals split between two parties, SF-Relate securely detects 94.9% of third degree relatives, and 99.9% of relatives second-degree or closer within 15 h of runtime. Our work enables secure identification of relatives across large-scale genomic datasets. Our code is available at https://github.com/froelich/sf-relate, and the full manuscript preprint is available at https://doi.org/10.1101/2024.02.16.580613.