People’s organs have a multiplicity of natural substances that act like therapeutic drugs to keep the whole body healthy. Now, researchers have developed a machine learning method to search for these natural drugs through the thousands of peptides, which can then be developed into actual drugs to combat disease.
The cells in our body swim in a soup of peptides that result from many biological processes.
Some peptides are bioactive and are absolutely essential to keep the body healthy, such as insulin and the gut hormone GLP-1, both of which have been commercialised to be among the world’s most commonly used drugs.
However, the vast majority of peptides result from the degradation of proteins and are therefore unimportant.
When researchers search for new drugs similar to insulin or GLP-1, they face the needle-in-a-haystack problem of identifying the few peptides that are bioactive. The hundreds of thousands of peptides in the body comprise a very large haystack in which to search for very few bioactive peptides.
Researchers have now solved this problem by training a machine learning model to help identify the few potentially useful peptides in the peptide soup, so that they can be examined individually in the search for new wonder drugs.
“Our problem is that when we examine a biological sample, this is a snapshot of protein degradation. However, because the many thousands of peptide fragments creates high background noise, identifying what is bioactive and what is not is very difficult. Our method finds the interesting signals in this background noise – like a magnet that can pull the needles out of the haystack,” explains a researcher involved in the study, Christian Toft Madsen, Senior Scientist, Novo Nordisk A/S.
The research has been published in Nature Communications.
Investigated 150,000 peptides
The researchers developed a machine learning model to identify the peptides in a biological sample that are probably bioactive and thus have potential as therapeutic drugs.
To show that the model works as intended, the researchers first purified peptides from 48 mice. The researchers individually examined peptides in the liver, muscles, brain, pancreas, intestine and two types of fat tissue.
To identify all the peptides, the researchers used mass spectrometry , which can map the peptide sequences based on the mass and charge of the peptides. This resulted in data on 157,857 unique peptide sequences.
The researchers then used machine learning to identify which of these 157,857 unique peptide sequences appeared to be bioactive.
The researchers had previously trained the machine learning model on known bioactive peptides, and the computer learned various features of bioactive peptides, including how the peptide is positioned in a pattern of other peptides and the contribution of a given peptide within a complex of other peptides.
Then the machine learning model could search for bioactive peptides in the soup of peptides in the sample and tell the researchers how close the individual peptides were to resembling bioactive peptides.
“We get a number from 0 to 1 for the individual peptide. The closer the score is to 1, the more likely the machine model has predicted that the peptide is bioactive,” says Christian Toft Madsen.
Discovered new peptide in insulin
Christian Toft Madsen says that 0.14% of the peptides in the investigated samples from mice appeared to have the most potential and were worth pursuing further in the studies of potential bioactivity.
In particular, the researchers wanted to discover peptides that can lower blood glucose, and they found a very interesting new insulin peptide in the pancreas.
Normally, insulin consists of three peptide fragments, but the researchers identified a fourth that arises from an alternative splicing of RNA into peptide.
This peptide scored very highly in the machine learning analysis of potential bioactivity but did not have the effect that the researchers thought.
“Our model predicted that this peptide definitely has a bioactive role, but we tested it for the ability to lower blood glucose, and it does not. Therefore, it must play another bioactive role, and we are currently considering whether we should proceed with studying this peptide’s role. We will also investigate whether the peptide is only relevant for mice or whether it is also relevant for people,” explains Christian Toft Madsen.
Can identify health-promoting peptides through exercise
Using the machine learning model on peptides from mice is a proof of concept that the method works, and the researchers are now carrying out various studies on samples from people.
The researchers want to determine how the peptide landscape in the body changes during exercise and whether it leads to the secretion of health-promoting bioactive peptides that could be potential therapeutic drug candidates.
Since exercise leads to increased insulin sensitivity in the body, machine learning may be able to pinpoint whether peptides could lead to this beneficial effect.
The researchers are also investigating this method in specific patient groups, including among people with heart failure.
The idea is that the peptide landscape can be a biomarker for heart failure since it differs from that of healthy people. This means that when new drugs are developed for people with heart failure, researchers can easily see whether the drugs actually change the person’s unique peptide landscape.
“We also aim to investigate whether the situation is the same for other organisms, including that perhaps mice or rats have the same heart failure peptide profile as people. This may indicate whether these organisms are good models of heart failure for people,” concludes Christian Toft Madsen.