Researchers from the University of Gothenburg (Sweden) have developed an artificial intelligence (AI) model to improve the diagnosis of potential cancer via the analysis of sugar molecules, glycans.
Glycan analysis, potential to reveal cancer
Glycans are structures of sugar molecules in our cells. These molecules can constitute cancer biomarkers, if analysis by mass spectrometry (read below) reveals that their structure has changed.
There mass spectrometry is an analysis method which aims to identify the different molecules and their quantity present in the samples studied. This analysis occurs by transforming molecules into charged ions, then sorting and detecting them based on their mass and charge.
Currently, data measured by spectrometry must be carefully analyzed by humans to see whether forms of cancer are occurring or not. This painstaking work requires hours, even days, and the expertise of a few specialists who have been trained for years, hence the usefulness of artificial intelligence which would speed up the analysis.
This is why a team of researchers led by Dr Daniel Bojar, associate lecturer in bioinformatics at the University of Gothenburg (Sweden), has developed an AI model to facilitate cancer detection. “ In principle, the method can be applied to any types of cancer since all forms of cancer that we know possess characteristically modified glycans that can be measured by mass spectrometry, explains Danier Bojar, main author of this work published on 1Er July 2024 in the magazine Nature Methods.
Read alsoCancer: how sugar facilitates the growth of tumors
CandyCrunch, an AI model that detects indications of cancer in seconds
This new AI model named “CandyCrunch”, in reference to the famous video game “Candy Crush Saga”, needs only a few seconds to detect possible indications of cancer via the identification of sugar molecules or glycans in the samples observed. .
CandyCrunch was trained for this purpose using a database of more than 500,000 examples of different glycan fragmentations and structures. The researchers used this data to train a “deep learning” artificial intelligence model (capable of learning on its own) to recognize other glycans in the future.
“ During training, we show the model many such spectra, in which the model can see the pieces of glycans, and then ask the model to predict which glycan belongs to this spectrum, explains Daniel Bojar. Then the model gives us a glycan as text and we “reward” the model or not based on how close the predicted glycan is to the true answer, until the model has learned to correctly associate glycans to the way they break down in the mass spectrometer.“
Read alsoE407a, E407, E471: food additives singled out by an epidemiological study
According to Daniel Bojar, “ training allowed CandyCrunch to calculate the exact sugar structure contained in a sample in 90% cases ».
The CandyCrunch model has proven to be more effective than the Glycoforest method (read the box below) and represents a considerable advance since it is capable of providing correct results in a record time.
The Glycoforest method uses high-resolution mass spectrometry to identify and annotate glycan structures from fragmentation spectra.
“ Of course, the exact time difference varies a lot depending on the sample., specifies Daniel Bojar, but we routinely get 10x faster analysis of glycan data through CandyCrunch, although it is important that this includes quantification for CandyCrunch but not for Glycoforest.“
The principal investigator of this study also notes a better performance of CandyCrunch than of Glycoforest. “As we train on a much larger dataset, our model can also identify a much wider range of glycans than Glycoforest”, he concludes.
English 
French