from Postmortem Human Prefrontal Cortex Louise A. Huuki-Myers1,*, Kelsey D. Montgomery1,*, Sang Ho Kwon1,2, Sophia Cinquemani1, Nicholas J. Eagles1, Daianna Gonzalez-Padilla1, Sean K. Maden3, Joel E. Kleinman1,4, Thomas M. Hyde1,4,5, Stephanie C. Hicks3,4,6,7,8, Kristen R. Maynard1,2,4,+, Leonardo Collado-Torres1,3,6,+ 1.Lieber Institute for Brain Development, Johns Hopkins (JH) Medical Campus, 2. The Solomon H. Snyder Department of Neuroscience, JH School of Medicine (JHSOM), 3. Department of Biostatistics, JH Bloomberg School of Public Health, 4. Department of Psychiatry and Behavioral Sciences, JHSOM, 5. Department of Neurology, JHSOM, 6. Center for Computational Biology, JH University (JHU), 7. Department of Biomedical Engineering, JHU, 8. Malone Center for Engineering in Healthcare, JHU * equal contributors; + co-corresponding authors; JHU & JHSOM: Baltimore, MD, USA Overview Dataset RNAScope Cell Type Proportion Deconvolution Methods Method Evaluation Marker Gene Selection Reference Dataset Features Pre-print & Resources Method Citation Approach Marker Gene Selection Availability Top Benchmark Performance DWLS (Dampened weighted least-squares) Tsoucas et al, Nature Comm, 2019 weighted least squares - R package on CRAN Cobos et al., Nat Commun., 2020 Bisque Jew et al, Nature Comm, 2020 Bias correction: Assay - R package on GitHub Dai et al., BioRxiv, 2023 MuSiC (Multi-subject Single- cell) Wang et al, Nature Communications, 2019 Bias correction: Source Weights Genes R package GitHub Jin et al., Genome Biol., 2021 BayesPrism Chu et al., Nature Cancer, 2022 Bayesian Pairwise t-test Webtool R package on GitHub Hippen et al., Genome Biol., 2023 hspe (dtangle) (hybrid-scale proportion estimation) Hunt and Gagnon- Bartsch, Ann. Appl. Stat. 2021 High collinearity adjustment Multiple options- default “ratio” 1vALL mean expression ratio R package on GitHub Dai et al., BioRxiv, 2023 CIBERSORTx Newman et al., Nat Biotech, 2019 Machine Learning Differential Gene expression Webtool, Docker Image Jin et al., Genome Biol., 2021 Conclusion Deconvolutio n Bechmark Pre-print Download this Poster DeconvoBuddies R package & dataset Funding: NIMH Grant R01 MH123183 & R01 MH111721 Related #deconvoChallenge Publications: Deconvolution Challenges: https://doi.org/10.1186/s13059-023-03123-4 Lute: https://doi.org/10.1101/2024.04.04.588105 Seminar Recording • 22 DLPFC tissue blocks from 10 neurotypical donors • Bulk RNA-seq samples with 6 different library type * RNA extractions (n=110) • snRNA-seq data with 56k nuclei across 7 broad cell types (n=19) • RNAScope/IF data labeling 6 broad cell types (n=12/13) • Produced multi-assay dataset with paired bulk RNA-seq, reference snRNA-seq, and RNAScope/IF cell type proportion measurements • Evaluated 6 popular deconvolution methods • Developed Mean Ratio method for identifying cell type specific marker genes useful for deconvolution • Different methods predict a wide range of cell type proportions • Some methods are impacted by library type & RNA extraction • Quality controlled for high-quality images with no technical artifacts • Similar ranges of cell type proportion with snRNA-seq, but not always correlated • Evaluated method performance by person correlation (cor) and root mean squared error (rmse) • Bisque and hspe were top performers • Top 25 Mean Ratio marker genes best balanced rmse and cor in hspe and Bisque Tran, Maynard et al., Neuron, 2021 & Mathys et al., Nature, 2019 • Bisque preforms poorly with <4 donors in the reference dataset • Both methods performed well with the inclusion of AD samples in reference Run Deconvolution • Bisque and hspe were most accurate methods in brain tissue • Bisque is sensitive to features of the reference dataset • Mean Ratio top 25 marker genes improved the performance of top methods Presenter Louise Huuki-Myers Staff Scientist at LIBD @lahuuki
lahuuki
kevinliebholz
marcduiker
lauravandoore
malarkey
jonrohan
notwaldorf
jeffersonlam
aarron
pedronauck
thekraken
thoeni
addyosmani
