Masterprüfung mit Defensio, Nikola Vinko

- 25.01.2022 16:30

“Automated classification of T cell subsets in single-cell RNA-seq experiments”

Single-cell transcriptomicsenablestheclassificationofthousandsofcellsthroughtran- scriptome profiling.Theabilitytoaccuratelypartitionoridentifythedifferentsubsetsof heterogeneouscellpopulationslikeTcellsiscriticalforasuccessfulanalysisofscRNA-seq datasets. Withtheexponentialgrowthofdatasetcountsandsizes,automatedcelltypeanno- tation hasthepotentialtoenhanceandstreamlinethesingle-cellanalysispipelinesubstan- tially.However,manysingle-cellannotationmodelssufferfromthelowqualityofreference datasets causedbythelowerdynamicrangeofscRNA-seq.Multimodalsingle-celltechnolo- gies likeCITE-seqenablesimultaneousprofilingofRNAandsurfaceproteinabundancein single cells,facilitatingcelltypeannotation. In thisthesis,wegeneratedacollectionof12annotatedPBMCandTcelldatasets,of which 9wereCITE-seqdatasets.ThisdatawasusedforbenchmarkingSingleR,linearSVM, and XGBoostclassifiermodels.Theperformanceofthemodelswasevaluatedwithin(intra) datasets andacross(inter)datasets.Inaddition,weevaluatedmodels’sensitivitytonumber of cells,inputfeatures,andtheirperformanceacrossunbiasedandtargetedexperiments,as well asacrossdifferentcellannotationlevels.OurresultsshowthatXGBoostoutperforms other modelsinallaspectsbyasignificantmargin.Basedonthesefindings,weusedXG- Boost andtrainedproductionmodelsforcelltypeclassification. The developedXGBoostmodelswereusedfortheautomaticcelltypeclassificationofa CROP-seq panTcellexperiment.TheidentifiedsubsetsrevealthateachmainTcellsubset (CD4+, CD8+,andgamma-delta)formsatranscriptionalcontinuumthatprogressesfrom naive tocentralandeffectormemoryTcells.Finally,weshowthatspecificcellknock- outs determinetheTcell’sresponsetoactivationbyinfluencingitsgeneprogramsand differentiationprogress.Ourautomaticannotationpipelineenablesfast,reproduciblehigh- resolutioncellannotation,furtheringtheanalysiscapabilitiesofnewtypesofsingle-cell CRISPR screensinheterogeneouscelltypes,likeTcells.

Organiser:

SPL 5

Location:
online Videokonferenz