npj Precision OncologySource publication:
Loading content…
At the Dresden University Cancer Center's MASTERsg precision oncology program, a gradient-boosting model re-reads the raw bin-level copy-number data from TSO500 panel sequencing to find focal PTEN deletions that the standard pipeline misses. TSO500 averages read counts across bins within a gene, so focal events fall short of its fold-change threshold; the model flagged deletions in 33 of 204 samples classified as PTEN wild-type, and 15 of the 20 stainable cases showed absent or reduced PTEN protein.
On the same TSO500 panel sequencing data, a machine-learning pipeline reports focal deletions in samples that the standard pipeline classifies as PTEN wild-type. The TSO500 pipeline averages read counts over 50–250 bp bins within a gene to estimate gene-level copy number, so focal deletions stay below its fold-change threshold; no published independent deletion-calling algorithm had been based on TSO500 output. Among 204 MASTERsg samples classified as PTEN wild-type by TSO500, the model flagged 33 deletions, 18 of which CNVkit also did not report; 15 of the 20 cases with tissue available for staining showed absent or markedly reduced PTEN protein.
The model reads bin-level counts directly instead of relying only on the gene-level average. Two Extreme Gradient Boosting classifiers handle binary deletion detection and three-class zygosity prediction, using normalized bin counts, a log-transformed polynomial fit and its derivative, global fold change, and tumor cell content as inputs. The training/test cohort held 59 samples across 15 solid tumor entities; in an independent external cohort of 61 randomly selected samples from 24 entities, Model 1 reached an F1 of 0.75 and specificity of 0.92, against TSO500's F1 of 0.71 and sensitivity of 0.56.
The demonstrated scope is PTEN deletion detection and zygosity evaluation on TSO500 panel sequencing data, with training and validation samples from the MASTERsg program and the Technical University of Munich pathology institute across multiple advanced solid tumor entities. The authors position the models as screening tools whose positive calls still need functional confirmation such as immunohistochemistry, and they plan to extend the method to other genes.
Validation cohort B was selected by Model 1 predictions, so the authors state it cannot estimate positive predictive value or false-positive burden in an unselected population, and comparisons within it favor Model 1. Re-running with TSO500 v2.6.0.2 lowered Model 1's F1 from 0.76 to 0.67 and Model 2's from 0.80 to 0.72. The authors also note that copy-number change is only one route to PTEN inactivation, that bulk sequencing cannot fully exclude subclonal deletions, that the training cohort is small, and that the wild-type class in the three-class model reached an average precision of 0.64.