A targeted cancer panel does not read every chromosome continuously. It samples selected genomic regions and infers copy-number changes from patterns in sequencing depth. A large deletion can be obvious. A focal deletion that clips only part of a gene may leave a subtler shape—especially when tumour purity, coverage and neighbouring regions vary.
Researchers in the MASTERsg precision-oncology programme trained an extreme-gradient-boosting model to recognize deletion patterns in PTEN, a tumour-suppressor gene. In a retrospective reanalysis of 204 samples that the TSO500 pipeline had not classified as PTEN-deleted, the model flagged 33 candidates. Eighteen had also not been reported by CNVkit, another established read-depth caller.[1]
The candidate-validation chain
| Measure | Value | Context |
|---|---|---|
| Previously unflagged samples reanalysed | 204 | Advanced-cancer samples in the retrospective MASTERsg cohort.Source 1 |
| Model-positive candidates | 33 | The model identified patterns consistent with possible PTEN deletion.Source 1 |
| Also missed by CNVkit | 18 of 33 | The candidate set was not simply reproducing one alternative caller.Source 1 |
| Candidates with tissue for staining | 20 | Immunohistochemistry was limited by tissue availability.Source 1 |
| Absent or reduced PTEN protein | 15 of 20 | Protein expression supported the biological relevance of most stainable candidates.Source 1 |
The model learned from the shape of coverage, not from a new sequencing assay
Each genomic bin contributes a signal: how many reads arrived relative to expectations and how that pattern changes across the gene. Conventional rules may require a broad drop before calling a deletion. The model can combine smaller, uneven changes across bins and learn patterns associated with known deleted and wild-type samples.
That makes the approach attractive for old data. The laboratory does not have to resequence every patient to ask a new question; it can re-evaluate existing panel output. It also creates a dependence on the original assay and cohort. A pattern learned from one panel, tumour mix and processing pipeline may shift when coverage chemistry or sample quality changes.
Three layers of evidence
| Layer | Role | What a positive result means |
|---|---|---|
| Panel pipeline | Routine clinical copy-number call | The standard system found enough evidence to classify a deletion.Source 1 |
| Machine-learning screen | Detect subtler multibin patterns | The sample deserves closer review; it is not automatically reclassified.Source 1 |
| Protein staining and expert review | Check whether PTEN expression and genomic shape support the call | Converging evidence strengthens the case but still requires clinical interpretation.Source 1 |
The strongest result came from focal deletions
In the held-out test data, the first model categorized 12 of 13 test patients correctly and identified every focal deletion that the TSO500 pipeline had missed. Synthetic tests suggested that the models could recognize smaller affected regions than the conventional threshold. Those results show sensitivity to a real failure mode; they do not establish performance across every cancer panel or every gene.[1]
The researchers also tested generalization in an external cohort of 61 advanced-cancer patients across 24 solid-tumour entities. External validation is important because a classifier can otherwise memorize the quirks of its development programme. The study reports useful performance, while still describing larger prospective studies as necessary before routine use.[1]
A candidate deletion is not a therapy recommendation
PTEN loss can be relevant to signalling pathways and treatment hypotheses, but an inferred deletion does not by itself establish that a patient will benefit from a particular inhibitor. Zygosity, tumour heterogeneity, coexisting mutations, protein expression and the evidence for the proposed therapy all matter. The model improves what the laboratory notices; it does not collapse diagnosis and treatment into one score.
The value of the result is not that machine learning overruled pathology. It found a class of cases that the ordinary pipeline might never have sent to pathology for another look. In precision oncology, a better question can be clinically useful before it becomes a definitive answer.
Sources & further reading
- Machine learning uncovers previously undetectable genetic deletions in cancer panel sequencing npj Precision Oncology · 16 Oct 2026 · Accessed 17 Oct 2026