ECP 2022 Abstract Book

Virchows Archiv (2022) 481 (Suppl 1):S1–S364 13 relevant for germline variation. Interestingly, we often find that statistical measures of variance in morphology are more relevant than measures of central tendency, and that the relationships are frequently non-linear and best modelled using tree ensembles. We further uncover pervasive batch effects and describe an approach to address these. Conclusion: We defined quantitative relationships between histo- logical phenotype and both germline and somatic genetic variation in tumour tissue using explainable machine learning. This approach has the potential to influence how clinical grade molecular inference models are optimised for generalisability in the future and allow histopathologists to gain intuition into the predictions made by deep models. Funding: This work was supported by: the MRC Human Genetics Unit core funding programme grants (MC_UU_00007/11 and MC_ UU_00007/16), MRC Toxicology Unit core funding (RG94521), Cancer Research UK Cambridge Institute core funding (20412), European Molecular Biology Laboratory, ERDF/Spanish Ministry of Science, Innovation and Universities-Spanish State Research Agency/DamReMap Project (RTI2018-094095-B-I00), and the Wellcome Trust (WT202878/B/16/Z). Edinburgh Genomics is partly supported through core grants from NERC (R8/H10/56), MRC (MR/K001744/1) and BBSRC (BB/J004243/1). J.C. is supported by a Wellcome Trust PhD Training Fellowship for Clinicians (WT223088/Z/21/Z) as part of the Edinburgh Clinical Academic Track (ECAT) programme. S.J.A. received a Wellcome Trust PhD Training Fellowship for Clinicians (WT106563/Z/14/Z) and National Institute for Health Research (NIHR) Clinical Lectureship. CP-02-007 Successful deployment of an AI solution for primary diagnosis of prostate biopsies in clinical practice M. Aslam*, A. Bansal, M. Atkinson, K. Sawalem, M. Mehdi, H. Abdelsalam, A. Heath, P. Huey, R. Nicholson, J. Theunissen, M. Grinwald, M. Vecsler, C. Linhart *Betsi Cadwaladr University Health Board, United Kingdom Background & objectives: This project aimed to validate, clinically deploy and integrate an AI decision support solution for prostate biopsies into the digital pathology workflow as a first read for primary diagnosis. Methods: The project included a technical validation and integra- tion phase of the AI solution into the lab workflow prior to the deployment. Seven pathologists underwent training and used the solution for prospective primary diagnosis of consecutive prostate core needle biopsies, reporting on 334 cases (1197 H&E slides). AI-assisted diagnoses were compared to the ground truth (GT = concordance of two pathologists). Results: The AI solution demonstrated high performance when pre-classifying slides with highest likelihood to be benign or malig- nant, with NPV = 98.8% (331 / 335) and PPV = 99.8% (399 / 400), respectively. 32% of slides have been classified as undetermined by AI. In 4 out of 7 discrepancies that were compared subsequently to the GT, the AI classification was correct. User experience survey, as reported by pathologists, showed high satisfaction marks for the AI solution. Pathologists felt more confident to review and report both benign and cancerous slides using the AI system and prefer to continue working with the system compared to a microscope. Conclusion: We report here successful implementation of a multi feature AI solution that automatically imparts clinically relevant diagnostic parameters regarding prostate cancer and other patho- logic features. The solution demonstrated its ability to accurately triage cancerous prostate cases and improve diagnostic quality. Thus, Galen Prostate AI solution could be used as significant aiding tool for pathologists in clinical decision-making in routine pathology practice. Funding: SBRI Centre of Excellence MD-01 | Molecular Diagnostics Pathology Symposium: Selected Abstracts MD-01-001 Comparison of whole genome with broad gene panel sequencing to identify actionable targets for cancer treatment D. Leunissen*, L. Kester, E. Driehuis, A. zur Hausen, P. Van Diest, W. de Leng, E. Speel *Maastricht UMC+, The Netherlands Background & objectives: DNA mutation analysis by broad panel NGS and WGS is currently used to guide cancer treatment. WGS can detect all genetic alterations, however, its implementation in daily clinic holds practical considerations. We evaluated the poten- tial of WGS alternatives in diagnostics. Methods: Publicly available WGS data of lung (n=86), colon (n=118), melanoma (n=63) and ovarian (n=42) cancers was used to identify clinically relevant variants using variant interpretation software (VarSome Clinical). We compared reported variants between the whole genome and targeted panels and their clinical relevance in Dutch routine care or clinical trials. Results: For each tumour type unique single nucleotide variants were identified (on average 1834 (likely) pathogenic variants (LPV) per tumour type). Structural variants were not included in this study. After applying in silico filters for commercially available cancer hotspot panel (CHP), Foundation Medicine (FMI) and TSO500 panels (50, 324 and 523 genes, respectively), of the LPV detected by WGS, an average of 12.4%, 6.6% and 3.5% was predicted to be detected by TSO500, FMI and CHP, respectively. Of the detected variants, all that were deemed clinically relevant were detected by the broad TSO500 and FMI gene panels while 15% would be missed using a smaller CHP gene panel. Conclusion: Of the clinically actionable LPV detected by WGS in four tumour types, 100% is assumed to be identified by broad gene panels NGS (TSO500, FMI) and 85% by CHP. We conclude that in current clinical practice, the added value of WGS compared to broad gene panels is limited for clinically actionable single nucleo- tide variant detection in the tumour types analysed. MD-01-002 Validation of TruSight TM Oncology Comprehensive (EU) assay V. Sementchenko*, M. Harris, N. Haseley, A. Yazdanparast, P. Wenz, J. Dockter, T. Pawlowski *Illumina, Inc., USA Background & objectives: TruSight TM Oncology Comprehensive (TSO Comp) is a CE-marked comprehensive genomic profiling (CGP) assay designed to interrogate solid tumours for relevant single nucleotide variants, multi-nucleotide variants, insertions, deletions and gene amplifications from DNA, and gene fusions and splice variants from RNA. Methods: It is an enrichment-based next-generation sequencing assay that targets 517 genes for detection of small DNA variants, 2 genes for detection of gene amplifications, 23 genes for detection of gene fusions and 2 genes for detection of splice variants. TSO Comp performance was evaluated in various analytical studies, including limit of detection/blank, accuracy, precision, utilizing FFPE-derived DNA and/or RNA samples. Results: Performance of the assay was assessed for multiple vari- ant classes: small variants and gene amplifications (amps) from S58