Machine learning-based proteomics analysis identified 23 protein biomarkers for enhanced risk prediction of myocardial infarction in Chinese and Europeans

BACKGROUND: Advances in proteomics assays enable measurement of thousands of plasma proteins, which could facilitate machine learning-based approaches to enhance risk prediction of disease. METHODS: We measured plasma levels of >9500 proteins among 4000 Chinese adults (1937 myocardial infarction [MI] cases and 2026 subcohort), using both Olink Explore (2923 proteins) and SomaScan (6597 proteins) platforms. Protein-based risk prediction models were constructed, using conventional and machine learning approaches, with model performance assessed via discrimination (C-statistic), classification (net reclassification index; NRI) and calibration. Protein scores were compared and combined with conventional risk factors and polygenic scores (PGS), and externally validated in 37,187 Europeans (630 MI cases) in UK Biobank. FINDINGS: Overall, 451 Olink and 530 SomaScan proteins were significantly associated (5% false discovery rate) with MI risk after adjustment for conventional risk factors. The protein prediction models based on Olink, SomaScan, or both platforms combined, outperformed conventional risk factors and PGS using discrimination, classification and calibration. The protein score using 23 Boruta-selected (Olink/SomaScan: 11/12) proteins yielded a 3.3-fold difference in 10-year cumulative incidence of MI between the top and bottom 20% risk groups (3.6 [95%CI: 3.1-4.1] vs 1.1 [0.9-1.2] per 100), compared with 2.5-fold (2.5 [2.2-2.9] vs 1.0 [0.9-1.2]) and 1.5-fold (1.9 [1.6-2.1] vs 1.3 [1.1-1.4]) differences for conventional and PGS, respectively. Adding this protein score to conventional risk factors increased C-statistics from 0.846 (0.829-0.862) to 0.864 (0.849-0.879), yielding an NRI of 72%. External validation in 37,187 participants in UK Biobank by adding 23 Boruta-selected proteins to conventional risk factors yielded comparable improvement in C-statistics for MI (from 0.710 [0.688-0.731] to 0.733 [0.710-0.756]) and 45% improvement in NRI (NRI 45%). CONCLUSIONS: Machine learning identified subsets of proteins for improved MI risk prediction beyond conventional risk factors and PGS, showing the potential for developing small panel proteomics assays to improve disease risk prediction.

Authors

Institutions

Publication Details

Journal
GigaScience
Published
2026-11-16
DOI
https://doi.org/10.1093/gigascience/giag102
Primary Topic
Acute Myocardial Infarction Research
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Machine learning-based proteomics analysis identified 23 protein biomarkers for enhanced risk prediction of myocardial infarction in Chinese and Europeans

Michael Hill, Daniel Avery, Mohsen Mazidi, Ling Yang et al.
GigaScience
Acute Myocardial Infarction Research
article

Machine learning-based proteomics analysis identified 23 protein biomarkers for enhanced risk prediction of myocardial infarction in Chinese and Europeans

Michael Hill, Daniel Avery, Mohsen Mazidi, Ling Yang, Andri Iona, Dianjianyi Sun, Canqing Yu, Dan Schmidt, Robert Clarke, Alfred Pozarickij, Pang Yao, Hannah Fry, Christiana Kartsonaki, Iona Y. Millwood, Robin G. Walters, Derrick Bennett, Baihan Wang, Neil Wright, Huaidong Du, Rory Collins, Yiping Chen, Zhengming Chen, Robert Clarke, Yali Yan, Pei Pei, Junshi Chen, Jun Lv, Liming Li
article en

Abstract

BACKGROUND: Advances in proteomics assays enable measurement of thousands of plasma proteins, which could facilitate machine learning-based approaches to enhance risk prediction of disease. METHODS: We measured plasma levels of >9500 proteins among 4000 Chinese adults (1937 myocardial infarction [MI] cases and 2026 subcohort), using both Olink Explore (2923 proteins) and SomaScan (6597 proteins) platforms. Protein-based risk prediction models were constructed, using conventional and machine learning approaches, with model performance assessed via discrimination (C-statistic), classification (net reclassification index; NRI) and calibration. Protein scores were compared and combined with conventional risk factors and polygenic scores (PGS), and externally validated in 37,187 Europeans (630 MI cases) in UK Biobank. FINDINGS: Overall, 451 Olink and 530 SomaScan proteins were significantly associated (5% false discovery rate) with MI risk after adjustment for conventional risk factors. The protein prediction models based on Olink, SomaScan, or both platforms combined, outperformed conventional risk factors and PGS using discrimination, classification and calibration. The protein score using 23 Boruta-selected (Olink/SomaScan: 11/12) proteins yielded a 3.3-fold difference in 10-year cumulative incidence of MI between the top and bottom 20% risk groups (3.6 [95%CI: 3.1-4.1] vs 1.1 [0.9-1.2] per 100), compared with 2.5-fold (2.5 [2.2-2.9] vs 1.0 [0.9-1.2]) and 1.5-fold (1.9 [1.6-2.1] vs 1.3 [1.1-1.4]) differences for conventional and PGS, respectively. Adding this protein score to conventional risk factors increased C-statistics from 0.846 (0.829-0.862) to 0.864 (0.849-0.879), yielding an NRI of 72%. External validation in 37,187 participants in UK Biobank by adding 23 Boruta-selected proteins to conventional risk factors yielded comparable improvement in C-statistics for MI (from 0.710 [0.688-0.731] to 0.733 [0.710-0.756]) and 45% improvement in NRI (NRI 45%). CONCLUSIONS: Machine learning identified subsets of proteins for improved MI risk prediction beyond conventional risk factors and PGS, showing the potential for developing small panel proteomics assays to improve disease risk prediction.

GigaScience
Peking University (CN), University of Oxford (GB), China National Center for Food Safety Risk Assessment (CN), Henan Provincial Center for Disease Control and Prevention (CN)
Acute Myocardial Infarction Research
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon