Abstract 1706P
Background
Precise tumor purity assessment is a crucial preparation step to determine target sequencing depth to detect oncogenic variants from tissue samples. Moreover, knowing tumor purity improves somatic variant calling during post-sequencing analysis. We developed an artificial intelligence (AI)-powered model to quantify tumor purity from whole slide images (AI-P), named Lunit SCOPE TP, to address these challenges. In this study, we analyzed the correlation between AI-P, variant allele frequency (VAF), and average depth of sequencing across 23 cancer types.
Methods
All data used in this publication is publicly available from TCGA. AI-P is calculated as the total number of tumor cells divided by the count of total cells in the WSI. Tumor variants were considered only if the variants were missense mutations with VAF under 0.5 and reported in COSMIC. Average VAF was calculated as the alternative count over the total depth per sample.
Results
The median values of average tumor VAF, average sequencing depth, and AI-P from TCGA cohorts were 25.2% (IQR 19.2%-30.5%), 108 (IQR 77-146) and 83.5% (IQR 67.8%-92.5%). The purity estimates from AI-P and VAF had high concordance across the 23 cancer types (|R| = 0.35, p < 0.001) and there was no correlation found between AI-P and mean sequencing depth (|R| = -0.02, p = 0.18). In a subgroup analysis, uveal melanoma had the highest average tumor VAF (median 37.9%) and AI-P estimates (median 99.0%). The AI-P estimates of BLCA (|R| = 0.43, p < 0.001), BRCA (|R| = 0.41, p < 0.001), and HNSC (|R| > 0.41, p < 0.001) had high concordance with the average tumor VAF, but PRAD (|R| = 0.03, p = 0.63), THCA (|R| = 0.10, p = 0.048), and KICH (|R| =-0.004, p = 0.98) were loosely correlated.
Conclusions
Our results underline the strong association between tumor purity assessed by Lunit SCOPE TP and VAF from sequencing data across 23 cancer types in TCGA. We believe that quantitative assessment of tumor purity will aid in estimating the minimum required sequencing depth to detect common oncogenic variants.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
The authors.
Funding
Lunit Inc.
Disclosure
G. Park, S. Kim, S. Song, H. Song, J. Ryu, S. Park, S. Pereira: Financial Interests, Personal, Full or part-time Employment: Lunit. K. Paeng: Financial Interests, Personal, Leadership Role: Lunit. C. Ock: Financial Interests, Personal, Full or part-time Employment: Lunit Inc.; Financial Interests, Personal, Invited Speaker: Ybiologics; Financial Interests, Personal, Stocks/Shares: Lunit Inc., Ybiologics. All other authors have declared no conflicts of interest.