Finding A Cure To All Diseases
How Deep Learning can revolutionize Medicine
Ever wondered why can’t we magically take a pill and cure all the diseases? Let’s dive into how an oral drug works. When we take up the medicine, it enters our stomach and the pill dissolves in the acidic medium. Next, it enters the small intestine, where the small intestine lining absorbs it to reach the bloodstream. Through our circulatory system, a pill can go to any place or organ in the body. After entering the bloodstream via the small intestine, the pill has to enter the liver, a remarkable organ that regulates the chemical levels in the body. Here, the medicine breaks down into small molecules or metabolites. As a result, only a part of the medicine can reach the target place and it binds to the receptors of the cells of that organ.
So now you know why a single pill cannot magically work for everything. It is because different organs have different types of cell receptors. After the drug binds to its receptor, it fulfils its action by either promoting production(agonists) or inhibition(antagonists) of certain chemicals/proteins which depend on the drug design and function. Also, it is worth noting that there can be a chance for the drug to go off target and cause unfavourable side effects. Hence, the specificity of drugs is very crucial.
After the job is done, eliminating drugs is a vital part of the body to prevent them from becoming toxic. So our liver is put to work again, and it excretes unnecessary substances through the kidneys. This is an overview of the working of the medicine. Thousands of cellular proteins that are complex molecules for growth and maintenance of the body, genes that are codes for life, and metabolites are involved in this complex process even though they are not in the spotlight.
All this medicine processing in the body is understood by researchers through pharmacokinetics, which involves many mathematical equations and analyses to determine where the medicine enters the body, the required dosage of the drug, and how long it takes for the body to eliminate the body medicine. But it is studied from a few human subjects and is assumed to work similarly in general populations. In contrast, how a drug works are affected by several factors such as age, sex, genetics, physiological state, etc. Hence, there is an individual variation in medicine response. Therefore, we cannot ignore the fact that the side effects of drugs can be too fatal. A book “The Brain that Changes Itself ~Norman Doidge” describes a heart-rending disability of a woman, Cheryl Schiltz who loses her vestibular function (vestibular apparatus is responsible for balance and is located in the inner ear) as a side effect of the drug Gentamicin. There are various other cases where drugs show their detrimental effects in the long run and can also lead to cancer. Thus, a paradigm shift towards personalized medicine became crucial.
Deep Learning Is The Answer...
We can build a much brighter future where humans are relieved of menial work using AI capabilities. ~Andrew Ng
To achieve this far-fetching goal of personalized medicine, we need a large amount of data to draw significant insights. It will take ages to manually make sense of all of it. Thanks to the development of machines, we have big data in biology ranging from genomics to proteomics and proteogenomics. Next, to analyze data with high precision, we need a magical tool and Deep Learning has the potential to achieve that. So what is Deep Learning? Deep Learning is a computational model that mimics the working of the human brain(So, it is also known as neural networks) and can process large amounts of data. It has the potential to unlock trends and patterns in datasets beyond human limits.
Currently, many Deep Learning models are being applied in biology which already displayed promising results like the AlphaFold of Google Deepmind which was able to solve the greatest challenge of protein folding with amazing speed and precision.
Getting Started with Graph Convolution Networks
To get the essence of how Deep learning works in biology, let’s take an exciting example of how we can predict properties such as solubility, binding affinity to medicine, etc of protein molecules using Graph Convolutional Networks(GCNs). GCNs are powerful tools for analyzing molecules. Note that the choice of the model is a rough approximation and the type we need to use is a huge subject in itself.
As you may have guessed, We input molecules to the computer in a graphical representation in GCNs. The diagram above represents atoms of molecules as nodes and the bond between molecules as edges. This is known as featurization.
In the figure above, X represents the structure of a molecule and Y represents an output predicted vector where each element of the vector describes some property of the molecule. The black box represents the Graph Convolution Network model. You may think that the function f(x) might be a simple mathematical function but it is a very complicated one and extracting information about molecules and giving output is not an obvious task.
Moving on, we arrive at the coolest part in Deep Learning, you allow the computer to learn the function f(X) which fits X and Y all by itself! This means that we define a model(Here, GCN) which has classes of functions. The model learns the right set of parameters, resulting in the right function that fits based on training data.
Diving into the inside working of a GCN model, the nodes and edges of the graph (graph represents a molecule) are represented with an input vector at the initial stage. This vector might contain values of chemical properties of molecules such as charge, hybridization state, etc. So how does the model propagate or in other words how do the vector values get updated? The model consists of hidden layers which are mathematical functions and the next graph convolution layer computes new vectors based on the value of the input vector. A node vector takes in the values of neighbouring nodes to perform certain calculations(like finding average) and then it gets new values(now each node is an aggregate of its neighbourhood).
The model then propagates forward to the next hidden layer of graph convolution in a similar way. Each hidden layer generates a unique set of vector outputs. The information passes from one hidden layer to another and the output of one layer becomes the input of the next layer. The final output is computed by applying a learned convolutional kernel(convolution kernel is sort of a matrix) to the nodes and edges of the output graph shown in the above diagram.
Don’t worry about the fancy-looking ReLU function in the diagram, it is an activation function that is a way of applying non-linearity to the function so that the model can learn a much wider range of functions. Finally, we get a trained model which can predict the properties of various molecules! Similarly, we can use deep learning to predict structure-activity relationships, generate new drug molecules and a lot more.
Various other graph-based powerful techniques yield promising results such as graph autoencoder networks which can learn to make a compact representation of graphs and are useful for unlabeled graph data. The main limitations of Graph-based neural networks are addressed through algorithms such as DeepWalk, GraphSage, etc. and all these advancements improve their efficiency and raise our hope in the development of Medicine through Deep Learning.
You need not know the detailed working of Deep learning models to get started. So, these trained models are made available to us through amazing packages like Molflash created by the Boltzmann team, so that you can focus a lot more on understanding cool aspects of biology and medicine. Molflash is also providing an exciting platform to create new drug molecules, a unique aspect of this software.
Stay curious and explore more about Deep learning models to create cool stuff!😁
