What AI can recover when a tumour sample has no healthy comparison

A tumour archive can preserve the material a researcher needs and still be missing half of a useful comparison. Clinical testing and biobanks often retain tumour tissue without a matching healthy sample. That makes it harder to tell which DNA differences appeared in the cancer and which belonged to the person all along.[3]

The distinction matters because genetic variation has more than one history. People can inherit DNA changes; cells also acquire changes during life, including as they divide. A difference found in cancer tissue is not automatically a change acquired by that cancer.[4]

VarNet-T, published on 9 April in Nature Communications and highlighted by A*STAR on 8 October, addresses that missing comparator.[1][3]

The patient is the most useful reference

For some genomic tests, a healthy-cell sample comes from blood, saliva or skin. Comparing it with the tumour helps identify somatic mutations: changes that arose during the person's lifetime. A generic reference genome cannot by itself supply that personal history.[5][2]

Consider an illustrative DNA difference present in both samples. Its presence in healthy tissue is evidence against treating it as tumour-specific. Remove the healthy sample and that direct comparison disappears. A tumour-only method must instead combine clues from the sequencing reads with what is already known about inherited variation and technical errors.[2][6]

The original VarNet, reported in 2022, treated aligned DNA reads as images. Different channels represented the DNA letters, confidence in those letters, confidence in their alignment and other properties. A convolutional neural network learned patterns around possible mutations. This is image analysis applied to a structured display of sequencing evidence, rather than a photograph of tissue.[2]

VarNet-T's training used more than 300 tumour–normal pairs to establish labels, while giving the model only the tumour sequences. That arrangement teaches it from cases where a healthy comparison exists, then asks it to work where the comparison is missing.[3]

Its documented workflow also uses resources for filtering inherited variants and recurrent sequencing artefacts. The system therefore combines learned patterns with external genetic knowledge. The model has no way to recreate a missing blood sample; it can make a better-supported call from the information that remains.[6]

A count that changes when false mutations slip in

One application is tumour mutational burden, or TMB: a measure of how many mutations a cancer carries. This is among the biomarkers that can help inform the use of immune checkpoint treatments. Counting inherited variation as new tumour mutations can distort that measure.[5][1]

The researchers tested TMB estimation on 1,000 whole-exome samples from The Cancer Genome Atlas, with 100 samples from each of ten solid-cancer types. Paired-normal-based calls supplied the comparison. At a high-TMB threshold of at least 10 nonsynonymous somatic mutations per megabase—changes that alter protein sequence per million DNA bases—VarNet-T's overall misclassification rate was 5%, about threefold lower than the next-best method.[1]

Classification performance was strongest for European-ancestry samples. Rare inherited variants and uneven representation in genetic reference resources remain relevant to tumour-only interpretation.[1]

A better mutation count still belongs inside a wider clinical judgement. The National Cancer Institute explains that a biomarker matching a treatment does not guarantee a response: other tumour features matter, and cells within one tumour may carry different markers. The VarNet-T study tests mutation detection and burden estimation; it does not establish improved survival.[5][1]

For research archives, the attraction is immediate. Material collected without a healthy counterpart may answer more questions if its mutations can be characterised more accurately. Better interpretation can increase an archive’s scientific value long after the samples were collected.[3]

Report an error or suggest a correction