The abilities of artificial intelligence has seen massive improvements in the past few years. Nowadays, this technology is being implemented in several fields including the medical industry. By finding patterns from large sets of biological data, AI methods can help doctors uncover the secrets behind many previously unsolvable medical cases. For example, AI techniques such as deep neural networks can help in the prediction of disease associated proteins, discovery of novel biomarkers and even in the design of small molecule drug leads. So, by using this powerful technology, researchers from the University of Hong Kong designed a robust deep learning approach to predict disease associated mutations of the metal binding sites in proteins.
Such a feat was never achieved before. Infact, this was the first time that a deep learning approach was used to predict disease associated metal relevant site mutations in metalloproteins. This new finding could prove to be a pivotal point in medical sciences as metal ions are instrumental to several structural and functional aspects in the human body. So naturally, when the balance of such metal ions within the body gets disrupted, it can cause severe side effects. Even when mutations occur to human genomes, it is seen that when such changes happen within the coding region of DNA, the metal binding sites of proteins are the main regions that get affected which eventually lead to the development of harmful diseases. From this, the scientists understood that learning more about the disease associated mutations at the metal binding sites of proteins is key discovery and development of new drugs.
Therefore in an attempt to further their understanding, the researchers began by integrating omics data from different databases to build a detailed training data set. Then, by analysing the statistics from the extracted data, the team observed that different metals have different disease associations. For example, mutations in calcium and magnesium sites are linked with muscular and immune system diseases while mutations in zinc sites can cause diseases associated with the breast, liver, immune system, kidney, etc.
The next step involved the usage of an energy based affinity grid map to obtain spatial features from the metal binding sites. Following this, the scientists merged the spatial features with physicochemical sequential features to train the model. By doing so, the team was able to get highly positive results; with the help of the spatial features, the performance of the prediction curve was enhanced with an area under the curve of 0.90 and accuracy of 0.82.
Since existing techniques in the metallomics and metalloproteins domain still lack refinement, this new method could prove to be a big leap in this field and could potentially help predict mutations that are associated with diseases like cancer, cardiovascular diseases etc. Moreover, the vast amount of data that scientists have collected over the years, can easily train these models and could help convert such data into valuable knowledge.