A recent study used machine learning to identify unoptimized short linear peptides, which represent promising candidates for blood glucose regulation for the treatment of type II diabetes.
Type II Diabetes
Type II diabetes mellitus (T2DM) is a chronic condition and represents over 90% of all diabetes mellitus (DM) incidences. In T2DM, cells fail to respond to insulin, or an insufficient amount of insulin is produced by the pancreas. Insulin normally allows glucose to enter cells from the blood, thereby reducing blood glucose levels. Currently, the global incidence of DM continues to grow, with the condition currently effecting over 450 million people. Researchers predict this figure to rise to almost 700 million by 2045.
DM drug discovery has seen some advances over the last couple of decades. However, the widespread prevalence of the disease stresses the urgent need for novel and effective anti-diabetic treatments that have improved safety profiles and are well tolerated for chronic use. Existing pharmacological interventions for T2DM include metformin, a small molecule drug known to work via several mechanisms, including AMP-activated protein kinase and mitochondrial activity. An important and efficacious group of T2DM therapeutics includes groups of peptides such as insulin analogues and incretin mimetics. This group of drugs act as either agonists or antagonists of endogenous human hormones. However, these existing peptide therapeutics require modifications for them to be able to exert their therapeutic effects. These peptide therapeutics comprise a significant proportion of the modern anti-diabetic drugs, as well as those in the development pipeline.
Short Linear Peptides
A largely unexplored alternative class of peptides, which may offer good safety profiles and tolerance for chronic use, are short linear peptides with few modifications. Research has shown that short linear peptides are capable of modulating intracellular signalling without modifications. These peptides can be highly selective, having multiple points of contact with their target, which may also result in decreased toxicity. Moreover, as they are comprised of amino acids, the peptides can be easily metabolised over time. This thereby avoids the tolerance issues associated with chronic administration of drugs.
However, a major drawback of the use of these peptides remains, as they are readily broken down during gastrointestinal digestion, so issues with low bioavailability when administered orally remain problematic. This has resulted in a turn towards developing optimised peptides to enhance therapeutic properties (such as cyclisation), although there are instances where anti-cancer linear peptides outperform their cyclic counterparts, indicating that this class of peptides should be further investigated.
To this end, the researchers behind this study used a machine learning method to identify short linear peptides which could modulate an effect on blood glucose levels, GLUT4 expression and/or glycated haemoglobin levels, while being non-toxic and showing no off-target effects. To ensure this, the researchers validated the peptide candidates in both in vitro and in vivo assays.
Machine learning algorithm for identification of short linear peptides
The researchers used a machine learning model previously utilised by Kennedy et al., comprising an ensemble of neural networks. To build the training set for the model, the researchers used structural data from public databases (bioactivity annotations, biological pathways, and structural annotations) and unstructured data extracted from peer-reviewed scientific papers and patents. The researchers used diabetes, blood glucose regulation and GLUT4 as the initial descriptors used to query the data sources. Processing of the structured data was achieved through a combination of graph-based techniques and manual curation. Whereas Natural Language Processing (NLP) techniques were applied to the unstructured data.
The researchers used the resulting dataset of peptides with a known effect on blood glucose regulation to train the predictive model for bioactivity in fold cross-validation. The model predicts glucose uptake efficiency of novel peptides’ from a large input of peptide sequences. The researchers achieved additional testing and refinement by incorporating a predict-test-refine loop, which is an example of active learning. Integration of the in vitro testing results biased the model towards the prediction of peptides with specific GLUT4 translocation activity. Thereby, a set of peptides, which the model found difficult to classify (those which an efficacy prediction close to 50%), were selected for experimental testing in vitro. The researchers then fed resulting data back into the predictive model.
The machine learning model predicted over 100 peptide candidates to potentially possess anti-diabetic functionality via blood glucose regulatory activity. A collection of tools to filter out peptides with undesirable physiochemical properties further refined the set of small linear peptides. The researchers discounted any peptides of more than 20 amino acids in length to reduce manufacturing costs. This resulted in a set of five peptides, which had no homology to one another, or any published bioactive compounds.
The five identified peptides were:
To validate the bioactivity of these predicted small linear peptides, the researchers used an in vitro glucose method. The researchers observed no significant glucose uptake activity for pep_RTE62G and pep_QT5XGQ, so did not progress the peptides further. Whereas, pep_1E99R5, pep_37MB3O and pep_ANUT7B, demonstrated the ability to significantly increase glucose uptake in human skeletal muscle cells. Moreover, pep_37MB3O and pep_ANUT7B displayed a stronger effect on glucose uptake than that of insulin.
These results suggest an in vitro success rate of 60% for the researcher’s predictive model. Moreover, this study indicates that these sequences should be further examined in relevant models as they represent a promising opportunity for the pharmaceutical industry, with reduced manufacturing costs and good tolerance in chronic use.
This study utilised an existing machine learning algorithm to predict small linear peptides, which could be used for the treatment of T2DM. Although further work is required to confirm the bioactivity, mechanism of action and clinical efficacy, this study presents initial evidence that unoptimized predicted peptides can display more enhanced bioactivity in vitro than insulin.
Image credit: xb100 – FreePik