Isoswitching for brain age prediction

Isoswitching drives the aging process in human brains
Beril Erdogdu et al. preprint
https://www.biorxiv.org/content/10.1101/2025.05.05.652255v1

Transcript

1. Isoswitching drives the aging process in human brains Beril Erdogdu et al. Center for Computational Biology, Johns Hopkins University, Baltimore, MD https://www.biorxiv.org/content/10.1101/2025.05.05.652255v1 Journal club slides by Geo Pertea Lieber Institute for Brain Development
2. Courtesy of: Vitting-Seerup K, Sandelin A. The Landscape of Isoform Switches in Human Cancers. Mol Cancer Res. 2017;15(9):1206-1220. DOI: 10.1158/1541-7786.MCR-16-0459 Vitting-Seerup et al. IsoformSwitchAnalyzeR: Analysis of changes in genome-wide patterns of alternative splicing and its functional consequences. Bioinformatics (2019) What is isoform switching ?
3. Other transcriptomic age clock models in literature • Wang K et al. (2018). Comprehensive map of age-associated splicing changes across human tissues and their contributions to age-associated diseases. Sci Rep. DOI: 10.1038/s41598-018-29086-2 ◦ comprehensive splicing events study across 48 human tissues (GTEx), including brain, showing superior age prediction vs. gene expression • Wang F et al. (2020). Improved Human Age Prediction by Using Gene Expression Profiles From Multiple Tissues. Sci Rep. DOI: 10.3389/fgene.2020.01025 ◦ single and multi-tissue (GTEx) age prediction • Ren X et al. (2020). RNAAgeCalc: A multi-tissue transcriptional age calculator. PLoS One. 2020 Aug 4;15(8):e0237006. DOI: 10.1371/journal.pone.0237006 • Shokhirev & Johnson (2021). Modeling the human aging transcriptome across tissues, health status, and sex. Aging Cell. DOI: 10.1111/acel.13280 ◦ meta-analysis of 3,060 human samples achieving high accuracy R²=0.98 in brain tissue (age 1-107) • Mboning L et al.(2025) BayesAge 2.0: a maximum likelihood algorithm to predict transcriptomic age. Geroscience. DOI:10.1007/s11357-024-01499-0
4. case-control DTU Data and methods Longitudinal study of 341 DLPFC samples from LIBD with age ranging from prenatal to 85+ years of age. 341 DLPFC samples Salmon CHESS 3.1.1 transcripts SPIT (DTU) isoform fractions prenatal (56) adult CTL postnatal
5. Methods and findings • first findings: prenatal to postnatal isoform switching (e.g. SNAP25a to SNAP25b) • systematic isoform shifts around birth with a clear prenatal/postnatal separation SPIT (DTU) prenatal (56) adult CTL postnatal
6. Methods and findings initial random forest regressor • initial random forest regressor using IFs from about 30k transcripts (~9600 genes), predicted brain age with R² = 0.856 accuracy isoform fractions
7. • feature ablation used to show that isoswitching is pervasive in brain tissue • log-transformed chonological age was used to give greater weight to expression changes in earlier life stages, which improved accuracy • the model was further refined through feature selection, using only top 100 genes (341) transcripts, resulting in R² score of 0.977 feature importances feature selection 341 transcripts (100 genes) initial random forest regressor isoform fractions Methods and findings
8. • a stacked model was devised incorporating lasso regression to predict age from IF data, across 5 overlapping age ranges 341 transcripts refined random forest regressor final prediction feature importances feature selection initial random forest regressor isoform fractions lasso regressors' stack [-1, 1] [10, 60] [25, 100] [0, 10] [1, 25] initial prediction Methods and findings
9. • the stack of lasso regressors with overlapping age ranges allow for further refining of the initial prediction of the random forest regressor Methods and findings
10. Methods and findings • each lasso regressor was trained on a specific age range • this allows for better biological interpretability of the variable contributions of isoform switches in the 100 genes, across different age ranges
11. Validation on independent data • the random forest model trained on the LIBD data was validated on a dataset of 53 human forebrain samples (Cardoso-Moreira et al., 2019): R²= 0.924 • to assess tissue specificity, a similar model was trained on IFs from 261 GTEx left-ventricle heart samples (ages 20-70) - where it performed very poorly (R²= 0.08) • reproducing the method on macaque transcriptome ◦ generate a better transcritpome annotation (realigning RNAseq to the genome), focusing on strongest predictive isoswitches in PALM and SNAP25 ◦ patterns similar to human isoform switching were observed there, suggesting evolutionary conservation of isoswitches across age ranges
12. Functional impact of isoswitching • selective protein products analysis - impact of age-related isoform switches on protein structures, protein domain alterations, loss-of-function and other age-related biomolecular changes • protein modifications can impact biological pathways and cellular functions, and can provide a better understanding of the molecular underpinnings of brain development and aging