Mary Rooney
Assistant Research Professor
Make your next move.
Prepare your application
Review drafts before you send them.
Explore related searches
Articles (4)
Large-scale plasma proteomic studies often use different high-throughput affinity platforms, and measurements of the same protein across platforms are often discordant. Discordance hinders cross-study integration. Improving proteomics data integration would enable more powerful meta-analyses, improve statistical power for biomarker discovery, and provide a better understanding of proteome–phenotype relationships. Here we present a cross-platform proteomics imputation variational autoencoder (cpiVAE), a deep generative model for bidirectional imputation of protein abundances between two widely used platforms: Olink and SomaScan. Using a training cohort of paired measurements from the China Kadoorie Biobank (CKB), cpiVAE learns a joint latent representation that enables cross-platform imputation. The cpiVAE method improves benchmarks provided by established methods, k-nearest neighbors (KNN) Weighted Nearest Neighbors (WNN, from Seurat v4). The cpiVAE method achieves up to 30% higher correlation between imputed and true values than KNN and WNN. The cpiVAE method also generalizes well to an independent cohort from the Atherosclerosis Risk in Communities Study (ARIC). Without retraining, cpiVAE maintains high performance compared to benchmarks. Associations of imputed protein levels with clinical phenotypes closely mirror results using the actual measurements and increases power in a meta-analysis scenario. A post-hoc feature importance matrix enables interpretation of this AI model. Protein pair features extracted from cpiVAE have significant overlap with known associations in the Search Tool for the Retrieval of Interacting Genes (STRING) database. In summary, cpiVAE offers an accurate, generalizable, and interpretable solution for cross-platform proteomic imputation, enabling integrated analyses across studies with proteomics measured on different platforms. This user-friendly framework and pre-trained model weights are available under a BSD2 open source license at https://github.com/joelbaderlab/cpiVAE_v1 .
Year:
2025
Measures from affinity-proteomics platforms often correlate poorly, challenging interpretation of protein associations with genetic variants (pQTL) and phenotypes. Here, we examined 2,157 proteins measured on both SomaScan 7k and Olink Explore 3072 across 1,930 participants with genetic similarity to European, African, East Asian, and Admixed American ancestry references. Inter-platform correlation coefficients for these 2,157 proteins followed a bimodal distribution (median r=0.30). Protein measures were associated with variants (pQTLs), and one-third of the pQTL signals were driven by protein-altering variants (PAVs). We highlight 80 proteins that correlate differently across ancestry groups likely due to differing PAV frequencies by ancestry. Furthermore, adjustment for PAVs with opposite directions of effect by platform improved inter-platform protein measure correlation and resulted in more concordant genetic and phenotypic associations. Hence, PAVs need to be accounted for across ancestries to facilitate platform-concordant and accurate protein measurement.
Year:
2025
The number of assays on highly-multiplexed proteomic platforms has grown ten-fold over the past 15 years from less than 1,000 to >11,000. The leading aptamer-based and antibody-based platforms have different strengths. For example, Eldjarn et al 1 demonstrated that the aptamer-based SomaScan 5k (4,907 assays, assessed in the Icelandic 36K) and the antibody-based Olink Explore 3072 (2,931 assays, assessed in the UK BioBank) had a similar number of cis -pQTLs among all targets (2,120 vs. 2,101) but Olink had a greater number of cis -pQTLs among the overlapping targets (1,164 vs. 1,467). Analysis of split plasma measures showed the SomaScan assays to be more precise: median coefficient of variation (CV) of 9.9% vs. 16.5% for Olink. 1 Precision of the newest versions of the platforms—SomaScan 11k (>11,000 assays, released in December 2023) and Olink Explore HT (>5,400 assays, released in July 2023)—has not yet been established. We assessed the reproducibility of the SomaScan 11k and Olink Explore HT using split plasma samples from 102 Atherosclerosis Risk in Communities (ARIC) Study participants. We found that the SomaScan 11k assays had a median CV of 6.8% (vs 6.6% for the subset of assays available on the SomaScan 5k) and the Olink Explore HT assays had a median CV of 35.7% (vs 19.8% for the subset of assays available on the Olink Explore 3072). Across Olink assays, the CVs were strongly negatively correlated with protein detectability, i.e., percent of samples above the limit of detection (LOD). For the 4,443 overlapping assays, the distribution of between-platform correlations was bimodal with a peak at r ∼0 and with another smaller peak at r ∼0.8. These findings on precision are consistent with the updated results by Eldjarn et al 1 but indicate that precision of these two leading platforms in human plasma has diverged as the number of included proteins has increased.
Year:
2024
BACKGROUND Low physical activity (PA) measured from accelerometers and low heart rate variability (HRV) measured from short-term ECG recordings are associated with worse cognitive function. Wearable long-term ECG monitors are now widely used. These monitors can provide long-term HRV data and, if embedded with an accelerometer, they can also provide PA data. Whether PA or HRV measured from long-term ECG monitors is associated with cognitive function among older adults is unknown. METHODS Free-living PA and HRV were measured simultaneously over 14-days using the Zio ® XT Patch among 1590 participants in the Atherosclerosis Risk in Communities Study [aged 72-94 years, 58% female, 32% Black]. Total amount of PA was estimated by total mean amplitude deviation (TMAD) from the 14-day accelerometry raw data. HRV indices (SDNN and rMSSD) were measured from the 14-day ECG raw data. Cognitive factor scores for global cognition, executive function, language, and memory were derived using latent variable methods. Dementia or mild cognitive impairment (MCI) status was adjudicated. Linear or multinomial regression models examined whether higher PA or higher HRV was cross-sectionally associated with higher factor scores or lower odds of MCI/dementia. Models were adjusted for demographic and medical comorbidities. RESULTS Each 1-unit higher in total amount of PA was significantly associated with 0.30 higher global cognition factor scores (95% CI: 0.16-0.44), 0.38 higher executive function factor scores (95% CI: 0.22-0.53), and 62% lower odds of MCI (OR: 0.38, 95% CI: 0.22-0.67) or 75% lower odds of dementia (OR: 0.25, 95% CI: 0.08-0.74) versus unimpaired cognition. Neither HRV measure was significantly associated with cognitive function or dementia. CONCLUSIONS PA derived from a 2-week ECG monitor with an embedded accelerometer was significantly associated with higher cognitive test performance and lower odds of MCI/dementia among older adults. By contrast, HRV indices measured over 2 weeks were not significantly associated with cognitive outcomes. More research is needed to define the role of wearable ECG monitors as a tool for digital phenotyping of dementia. CLINICAL PERSPECTIVE What Is New? This cross-sectional study evaluated associations between physical activity (PA) and heart rate variability (HRV) measured over 14 days from a wearable ECG monitor with cognitive function. Higher total amount of PA was associated with higher global cognition and executive function, as well as lower odds of mild cognitive impairment or dementia. HRV indices measured over 2 weeks were not significantly associated with cognitive outcomes. What Are the Clinical Implications? These findings replicate positive associations between PA and cognitive function using accelerometer data from a wearable ECG monitor with an embedded accelerometer. These findings raise the possibility of using wearable ECG monitors (with embedded accelerometers) as a promising tool for digital phenotyping of dementia.
Year:
2024
Collaborators (19)
Ciprian Crainiceanu
Professor
Johns Hopkins University
Josyf Mychaleckyj
Professor
University of Virginia
Bing Yu
Professor
The University of Texas Health Science Center at Houston
Peter Ganz
University of California, San Francisco
Elsayed Soliman
Director/Professor
Wake Forest School of Medicine
Laura Raffield
Assistant Professor
University of North Carolina at Chapel Hill
Peifeng Ruan
Assistant Professor
University of Texas Southwestern Medical Center
Lin Chen
Associate Professor
University of Minnesota
Josef Coresh
Professor
Johns Hopkins University
Amil Shah
Professor of Medicine, Director of Population Sciences
University of Texas Southwestern Medical Center
Hau-Tieng Wu
Professor
Courant Institute of Mathematical Sciences
Robert Gerszten
Chief of Cardiology
Beth Israel Deaconess Medical Center
Joel Bader
Professor
Johns Hopkins University
Jennifer Deal
Associate Professor
Johns Hopkins University
Elizabeth Selvin
Johns Hopkins University
Quan Sun
Research Assistant Professor
University of Pennsylvania
Lekki Wood
Assistant Professor
Baylor College of Medicine
Yun Li
Professor
University of North Carolina
Karen L. Mohlke
Professor
University of North Carolina at Chapel Hill
Make your next move.
Prepare your application
Review drafts before you send them.