Genetic Prediction of Cancer Recurrence: Scientists Verify Reliability of Computer Models
In biomedical research, machine learning algorithms are often used to analyse data—for instance, to predict cancer recurrence. However, it is not always clear whether these algorithms are detecting meaningful patterns or merely fitting random noise in the data. Scientists from HSE University, IBCh RAS, and Moscow State University have developed a test that makes it possible to determine this distinction. It could become an important tool for verifying the reliability of algorithms in medicine and biology. The study has been published on arXiv.
Machine learning methods help analyse complex biological data, ie for predicting the likelihood of cancer recurrence based on gene expression, which reflects the activity levels of specific DNA regions within cells. However, it is not always clear whether these algorithms are detecting meaningful patterns or merely fitting random noise in the data.
A team of scientists from HSE University, IBCh RAS, and Moscow State University has developed a test to assess how reliably the classifier distinguishes between different patient groups. In this case, the two groups were patients who experienced a recurrence of the disease and those who did not. A model performs correctly if it effectively captures biologically meaningful differences. If the algorithm simply separates the data at random, its accuracy may appear deceptively high. The researchers focused on linear classifiers, one of the most widely used ML tools in biomedicine.
Anton Zhiyanov
'We aimed to test whether randomly generated (synthetic) data could be separated by a linear classifier as effectively as real biological samples. To do this, we calculated an upper bound on the p-value, which indicates the likelihood that the model is merely "guessing." The lower this p-value, the more reliable the classifier,' explains Anton Zhiyanov, Research Fellow at the HSE Laboratory of Molecular Physiology.
The researchers conducted a series of experiments using synthetic data, allowing them to precisely control the degree of differences between classes. They then applied the new test to real-world medical models that predict the risk of breast cancer recurrence.
The results showed that most classifiers failed to capture any meaningful differences between patients with and without recurrence. Further analysis revealed that 559 out of 570 models produced results consistent with random chance. This suggests that many algorithms may appear accurate, while in reality their predictions are driven by coincidences rather than genuine patterns.
However, the researchers also identified reliable models that reveal biologically meaningful patterns. One such model was a classifier that focused on the activity levels of the ELOVL5 and IGFBP6 genes. This algorithm was further tested on an independent data sample, confirming that differences in the expression of these genes are indeed linked to the risk of cancer recurrence.
Each point on the graph represents a patient, with the expression levels of two genes measured: IGFBP6 on the X-axis and ELOVL5 on the Y-axis. The orange dots represent patients with a recurrence, while the blue dots represent those without. In the first graph, these points (patients) are clearly separated by a straight line, representing a linear classifier. In the second graph, the points are randomly distributed, and the classifier fails to identify any patterns between gene expression and actual recurrence.
Alexander Tonevitsky
'Our test could become an important tool for verifying the reliability of algorithms in biology and medicine. It helps prevent false conclusions and emphasises models that truly identify important patterns, which is crucial for making decisions about patient treatment,' comments Alexander Tonevitsky, Professor at the HSE Faculty of Biology and Biotechnology.
The study was conducted with support from HSE University's Basic Research Programme within the framework of the Centres of Excellence project.
