Jochen Weile

Variants in the familial hypercholesterolemia gene LDLR —the most important genetic driver of cardiovascular disease—can raise circulating low-density lipoprotein (LDL) cholesterol concentrations and increase the risk of premature atherosclerosis. Definitive classifications are lacking for nearly half of clinically encountered LDLR missense variants, limiting interventions that reduce disease burden. We tested the impact of ~17,000 (nearly all possible) LDLR coding variants on both LDLR cell-surface abundance and LDL uptake, yielding sequence–function maps that recapitulate known biochemistry, offer functional insights, and provide evidence for interpreting clinical variants. Functional scores correlated with hyperlipidemia phenotypes in prospective human cohorts and augmented polygenic scores to improve risk inference, highlighting the potential of this resource to accelerate familial hypercholesterolemia diagnosis and improve patient outcomes.

Background: Medical documents are a crucial resource for medical research around the world. While troves of valuable health data exist, they are largely computationally inaccessible as hard copies of unstructured text. Moreover, the persistent prevalence of fax machines in medical settings contributes to further degradation of document quality. Manual data extraction from these documents is time-consuming and resource intensive. However, large language models (LLMs) have recently shown great promise for automated digitization and information extraction (IE), greatly improving upon previous technologies in terms of accuracy. Objective: Here we aim to survey the landscape of LLM-based IE tools with respect to their applicability to biomedical documents and to systematically benchmark them on synthetic documents with paired reference extractions. We put special focus on locally executable tools and models that suit privacy- and resource-constrained settings. Methods: We reviewed LLM-based IE tools from the literature and assessed them with respect to their suitability for use in biomedical research. We found only one of these tools (NuExtract2) to satisfy our selection criteria and compared it to LLM foundation models prompted to perform extractions. We created 200 mock molecular test reports with paired reference data using a bespoke procedural generation approach and evaluated the tools’ performance across different prompting strategies, input modalities, and document qualities. Results: We found model performance to be very sensitive to input modalities and quality. The best overall performance was observed for the proprietary GPT 4.1-mini, with an F ₁ score of 79.2% averaged across quality levels and modalities. The best performing open-weights model was Mistral-small 3.1 with an average F ₁ score of 76.2%. The NuExtract2 model was the only one able to run on regular laptop computers, but did not perform well. Surprisingly, we found the choice of one-shot over zero-shot prompts to only have a small effect size on extraction performance in most cases. Conclusions: LLMs place the automatic processing of highly complex medical documents in resource-constrained and privacy-sensitive settings within reach, but still require human oversight. Availability: Source code and data are available on GitHub [4] and on Zenodo [5], respectively.

BACKGROUND Cancer research literature is often riddled with technical jargon that is not digestible to the average person. Individuals interested in research studies may want to contribute through patient partner engagement or sample donation but find the relevant literature overwhelming. Through the generation of lay summaries, previously inaccessible research papers become easier to comprehend, especially for patient partners or data donors. With large language models (LLMs) continuing to advance, so does their capability to summarize large texts. OBJECTIVE In this study, we examined whether LLMs can produce lay summaries of scientific literature at-scale, while maintaining readability and accuracy to their source texts. METHODS We developed a tool to generate lay summaries of open-access article abstracts and their full texts with GPT-4-Turbo. Prompt development aimed for a target 8th grade reading level assessed with Flesch-Kincaid Grade Level. Human-review metrics were used to evaluate readability and accuracy when generated using abstracts versus full text articles. RESULTS The average Flesch-Kincaid Grade Level Score was 7.13 for abstract-based summaries and 7.39 for full text-based summaries, indicating summaries at around 7th grade reading level. Human-review metrics showed these summaries were of similar readability and accuracy when generated using abstracts versus full text articles, with mean accuracy scores from human review of 7.09 vs 7.42 out of 10 respectively. Additionally, qualitative patient-based assessment indicated these summaries would encourage participation in research studies. CONCLUSIONS By generating lay summaries for complex and lengthy research papers, their scientific information becomes accessible to a larger audience, including patient partners interested in contributing to cancer research. Summaries that are easy to understand will allow participants to make informed decisions about their involvement and appreciate the impact of their contributions if and when their results are published.

Background: Cancer research literature is often riddled with technical jargon that is not digestible to the average person. Individuals interested in research studies may want to contribute through patient partner engagement or sample donation but find the relevant literature overwhelming. Through the generation of lay summaries, previously inaccessible research papers become easier to comprehend, especially for patient partners or data donors. With large language models (LLMs) continuing to advance, so does their capability to summarize large texts. Objectives: In this study, we examined whether LLMs can produce lay summaries of scientific literature at-scale, while maintaining readability and accuracy to their source texts. Methods: We developed a tool to generate lay summaries of open-access article abstracts and their full texts with GPT-4-Turbo. Prompt development aimed for a target 8th grade reading level assessed with Flesch-Kincaid Grade Level. Human-review metrics were used to evaluate readability and accuracy when generated using abstracts versus full text articles. Results: The average Flesch-Kincaid Grade Level Score was 7.13 for abstract-based summaries and 7.39 for full text-based summaries, indicating summaries at around 7th grade reading level. Human-review metrics showed these summaries were of similar readability and accuracy when generated using abstracts versus full text articles, with mean accuracy scores from human review of 7.09 vs 7.42 out of 10 respectively. Additionally, qualitative patient-based assessment indicated these summaries would encourage participation in research studies. Conclusion: By generating lay summaries for complex and lengthy research papers, their scientific information becomes accessible to a larger audience, including patient partners interested in contributing to cancer research. Summaries that are easy to understand will allow participants to make informed decisions about their involvement and appreciate the impact of their contributions if and when their results are published.

Pathogenic variants in the Autoimmune Regulator (AIRE) gene cause Autoimmune Polyendocrine Syndrome Type 1 (APS-1), a rare primary immunodeficiency disease with symptoms including hypoparathyroidism, adrenal insufficiency, and chronic mucocutaneous candidiasis. AIRE increases the expression and presentation of tissue-specific genes expressing 'self' antigens in the developing T cell niche, thus triggering the elimination of self-reactive T cells and preventing autoimmunity. Earlier diagnoses can benefit patients, and APS-1 diagnosis by AIRE sequencing is increasingly common. However, two thirds of reported clinical variants are missense, and more than half of these are "variants of uncertain significance" (VUS). Cell-based variant functional assays can provide strong evidence towards more informative variant classification, but these are carried out reactively, often years after clinical presentation. By contrast, proactively assessing all possible missense variants could provide immediate evidence to guide genetic diagnosis, even for never-before-seen variants. Here we used an insulin promoter-driven reporter to proactively assess the function of 9790 AIRE missense variants. The resulting AIRE variant effect map both validates and extends current biochemical knowledge, concords with pathogenicity annotations, and provides proactive evidence for 70% of previously-reported VUS. Placing our map in the context of both an international APS-1 cohort and the UK BioBank revealed quantitative genotype-phenotype correlations. Using current guidelines, we provide classifications for 32% of current VUS. Together, our proactive resource of AIRE variant impacts offers the potential to improve patient outcomes via more rapid and definitive APS-1 diagnosis.

Importance: Manual data extraction from genomic lab reports for on-line registries and databases is time-consuming for human resources such as clinical research coordinators. Automated tools, especially LLMs, can address these issues. Efficient and accurate data processing is crucial for building a reliable database. Objective: To streamline the data extraction and curation process for genetic testing lab reports using an LLM-based approach. Design: Nine sample molecular lab reports were selected for manual data extraction by two expert curators. The process was timed, and the results served as gold-standard for validating automated extraction. Eighteen fields from the OHRCN's data model were selected as extraction targets. Setting: The study was conducted within OHCRN, which unifies research, genomic, and clinical patient data from clinics and laboratories across Ontario, Canada. Participants: Nine laboratories agreed to share sample molecular lab reports and two clinical research coordinators affiliated with OHCRN participated as data curators. Exposure: LLM-based Extraction of Information (LEI), an automated data extraction pipeline, was developed using regular expressions, Trie search, and LLMs to extract data from molecular lab reports and structure it for inclusion into OHCRN's database. Main Outcomes and Measures: LEI was evaluated by measuring the F1-score on the extraction task of 18 entity types. These measures were compared against 15 extraction tools in the biomedical domain. Extraction time was also measured and compared against manual extraction times. Results: LEI demonstrated quality on par with and surpassing other existing LLM-based extraction methods. Reference tools showed F1-scores around 70%, while LEI achieved an average score of 87.4%. LEI reduced extraction time by approximately 2-fold, with an average time of 7.59 minutes per report including results review by curators, compared to 14.88 minutes per report for manual extraction. Conclusions and Relevance: LEI facilitates standardized, accurate, and efficient healthcare data extraction from unstructured texts, significantly improving the current OHCRN workflow. By automating the extraction process, LEI allows expert curators to focus on validating results rather than performing manual data entry. LEI's simple interface enables researchers to easily guide extraction tasks and supports adaptability across diverse biomedical scenarios. Future improvements in accuracy may be achieved through fine-tuning techniques and ongoing advancements in LLM technologies.

The tumor suppressor gene STK11 encoding Serine/Threonine Kinase 11 (STK11) is associated with Peutz-Jeghers Syndrome (PJS), a heritable gastrointestinal disease that increases lifetime cancer risk, and with somatic variation that contributes to ∼30% of lung and 20% of cervical cancers. Although identifying pathogenic variants is clinically actionable, over 94% of STK11 missense variants that have been observed clinically lack a definitive classification. We therefore measured the impact of STK11 variants at scale in a mammalian cell-based assay, scoring 6,026 (73% of all possible) amino acid substitutions across the full-length gene. Functional scores—which were consistent with biochemical properties, smaller-scale assays, and pathogenicity annotations—identified a subset of PJS patients with germline STK11 variants diagnosed later in life, as well as somatic STK11 variants found in cancer patients that had comparable overall survival estimates to wild-type STK11 . Our scores provided new evidence for 350 annotated VUS STK11 missense variants and ∼80% of missense variants that have not yet been reported clinically, but we might expect to observe in the future. Thus, our effect map provides a proactive resource for gaining sequence-structure-function insights and evidence for actionable interpretation of clinical missense variants.

Amyotrophic lateral sclerosis (ALS) is a progressive motor neuron disease for which important subtypes are caused by variation in the Superoxide Dismutase 1 gene SOD1 . Diagnosis based on SOD1 sequencing can not only be definitive but also indicate specific therapies available for SOD1 -associated ALS (SOD1-ALS). Unfortunately, SOD1-ALS diagnosis is limited by the fact that a substantial fraction (currently 26%) of ClinVar SOD1 missense variants are classified as “variants of uncertain significance” (VUS). Although functional assays can provide strong evidence for clinical variant interpretation, SOD1 assay validation is challenging, given the current incomplete and controversial understanding of SOD1-ALS disease mechanism. Using saturation mutagenesis and multiplexed cell-based assays, we measured the functional impact of over two thousand SOD1 amino acid substitutions on both enzymatic function and protein abundance. The resulting ‘missense variant effect maps’ not only reflect prior biochemical knowledge of SOD1 but also provide sequence-structure-function insights. Importantly, our variant abundance assay can discriminate pathogenic missense variation and provides new evidence for 41% of missense variants that had been previously reported as VUS, offering the potential to identify additional patients who would benefit from therapy approved for SOD1-ALS.

Motivation: Long-read sequencing technologies, an attractive solution for many applications, often suffer from higher error rates. Alignment of multiple reads can improve base-calling accuracy, but some applications, e.g. sequencing mutagenized libraries where multiple distinct clones differ by one or few variants, require the use of barcodes or unique molecular identifiers. Unfortunately, sequencing errors can interfere with correct barcode identification, and a given barcode sequence may be linked to multiple independent clones within a given library. Results: Here we focus on the target application of sequencing mutagenized libraries in the context of multiplexed assays of variant effects (MAVEs). MAVEs are increasingly used to create comprehensive genotype-phenotype maps that can aid clinical variant interpretation. Many MAVE methods use long-read sequencing of barcoded mutant libraries for accurate association of barcode with genotype. Existing long-read sequencing pipelines do not account for inaccurate sequencing or nonunique barcodes. Here, we describe Pacybara, which handles these issues by clustering long reads based on the similarities of (error-prone) barcodes while also detecting barcodes that have been associated with multiple genotypes. Pacybara also detects recombinant (chimeric) clones and reduces false positive indel calls. In three example applications, we show that Pacybara identifies and correctly resolves these issues. Availability and implementation: Pacybara, freely available at https://github.com/rothlab/pacybara, is implemented using R, Python, and bash for Linux. It runs on GNU/Linux HPC clusters via Slurm, PBS, or GridEngine schedulers. A single-machine simplex version is also available.

The tumour suppressor CHEK2 encodes the serine/threonine protein kinase CHK2 which, upon DNA damage, is important for pausing the cell cycle, initiating DNA repair and inducing apoptosis. CHK2 phosphorylation of the tumour suppressor BRCA1 is also important for mitotic spindle assembly and chromosomal stability. Consistent with its cell cycle checkpoint role, both germline and somatic variants in CHEK2 have been linked to breast and multiple other cancer types. Over 90% of clinical germline CHEK2 missense variants are classified as variants of uncertain significance, complicating diagnosis of CHK2-dependent cancer. We therefore sought to test the functional impact of all possible missense variants in CHK2. Using a scalable multiplexed assay based on the ability of human CHK2 to complement DNA sensitivity of a S. cerevisiae lacking its ortholog RAD53 , we generated a systematic ‘missense variant effect map’ for CHEK2 missense variation. Map scores reflect known biochemical features of CHK2 and exhibit good performance in separating pathogenic from benign clinical missense variants. Thus, the missense variant effect map for CHK2 offers value in understanding both known and yet-to-be-observed CHK2 variants.