Karakatsoulis, G. | Search bioRxiv

bioRxiv2024

Synth4bench: a framework for generating synthetic genomics data for the evaluation of tumor-only somatic variant calling algorithms

MotivationSomatic variant calling is a key activity towards identifying genomic alterations; yet, the evaluation of the respective tools remains challenging due to the scarcity of high quality ground truth datasets. To overcome this limitation, we developed synth4bench, a synthetic data generation pipeline for robust benchmarking. Using a systematic process to create distinct synthetic datasets, we thoroughly evaluated five variant callers (Mutect2, FreeBayes, VarDict, VarScan2 and LoFreq). We compared tool outputs against our synthetic ground truth across key sequencing aspects (such as depth and read length) to assess their capacities and shed light on their underlying algorithmic principles. ResultsSynth4bench is an approach for evaluating tumor-only somatic variant callers that relies on a systematic definition of fully controlled ground-truth datasets. Our analysis revealed significant inconsistencies among the tool outputs and a strong dependence of caller performance on sequencing parameters. Indels remain the hardest-to-call variant type, driven by errors at low allele frequencies. Algorithmic choice is also critical; the most robust callers displayed the highest precision in allele frequency estimation, while the most sensitive caller was best for maximizing true positive recovery. Conversely, the least suitable caller exhibited systematic errors along with the poorest overall performance. These findings indicate that there isnt a one-solution-fit-all; sequencing optimization together with caller selection are necessary to maximize sensitivity and reliability. Furthermore, the pronounced inconsistencies suggest that current algorithms are not yet able to capture all mutational mechanisms adequately, with the modeling of the underlying processes remaining an open challenge. Availabilitycode: https://github.com/sfragkoul/synth4bench/ and data: https://zenodo.org/records/16524193 Graphical Abstract O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=69 SRC="FIGDIR/small/582313v2_ufig1.gif" ALT="Figure 1"> View larger version (16K): org.highwire.dtl.DTLVardef@1ac0fe5org.highwire.dtl.DTLVardef@1479cddorg.highwire.dtl.DTLVardef@8b7d1borg.highwire.dtl.DTLVardef@1c2a81b_HPS_FORMAT_FIGEXP M_FIG C_FIG

bioinformatics↗

bioRxiv2024

APNet, an explainable sparse deep learning model to discover differentially active drivers of severe COVID-19

MotivationComputational analyses of plasma proteomics provide translational insights into complex diseases such as COVID-19 by revealing molecules, cellular phenotypes, and signaling patterns that contribute to unfavorable clinical outcomes. Current in silico approaches dovetail differential expression, biostatistics, and machine learning, but often overlook nonlinear proteomic dynamics, like post-translational modifications, and provide limited biological interpretability beyond feature ranking. ResultsWe introduce APNet, a novel computational pipeline that combines differential activity analysis based on SJARACNe co-expression networks with PASNet, a biologically-informed sparse deep learning model to perform explainable predictions for COVID-19 severity. The APNet driver-pathway network ingests co-expression and classification weights to aid result interpretation and hypothesis generation. APNet outperforms alternative models in patient classification across three COVID-19 proteomic datasets, identifying predictive drivers and pathways, including some confirmed in single-cell omics and highlighting under-explored biomarker circuitries in COVID-19. Availability and ImplementationAPNets R, Python scripts and Cytoscape methodologies are available at https://github.com/BiodataAnalysisGroup/APNet Contactggeorav@certh.gr Supplementary informationSupplementary information can be accessed in Zenodo (10.5281/zenodo.10438830).

systems biology↗