An Example of AI's Positive Role

Over the past century, AI has risen to stardom after its numerous appearances in big-budget films and best-selling books. However, movies focus on evil uses of AI when, in reality, AI has been a very beneficial tool. Take the recent Covid-19 outbreak as an example. Ai has been employed to model, track, diagnose, and help prevent the spread of the virus. One potential use of AI being explored is using AI to quickly generate a list of medical compounds that can bind to the virus, significantly speeding up the process of finding effective medicine. 

Unlike antibiotics, antiviral medication binds to and disables a specific protein, not the entire virus. For example, scientists are targeting covid-19’s spiky surface proteins, peplomers. Peplomers bind to human ace2 proteins to enter our lung cells. Disabling the peplomers makes the virus ineffective. 

Unfortunately, traditional methods of predicting how the tangled mess of amino-acids will look are time-consuming and expensive. I mean look at all the steps that it take in real life. A protein’s primary structure is a chain of amino acids joined by peptide bonds through dehydration synthesis. Hydrogen bonds turn it into an alpha helix, a parallel or antiparallel beta-pleated sheet. (Basically, a coil, a zig-zag, or a fancy zig-zag.) The tertiary structure of a protein is more complex. It’s partially determined by the polarity of the R-group. Polar molecules mix with water so they stay on the outside. Non-polar molecules don’t. Tertiary structures often bind with each other to form the final, quaternary structure. Yikes.

Enter AI. AI is a vague name for programs that analyze and utilize information in an intelligent way. Machine learning and deep learning uses structures called neural nets to analyze more information and learn more complicated things. Neural nets mimic human brains and are made up of neurons and channels. The neurons are grouped into layers. They take many different forms or architectures. Protein folding neural nets consist of convolutional residual networks and Dilated Residual Network. They might look scary but just remember, DRN’s are bigger and more accurate. Diagraming ResNets neuron by neuron is challenging. Instead, it does them in layers. ResNets take one layer’s information and connect it with another layer, allowing for more information to be analyzed. Each neuron “looks” at a specific trait and feeds that information forward. The next layer takes the information and analyzes a broader pattern. Of course, that lends the question of what information are we studying and how does the computer analyze it? 

Simply put: Protein ResNets study the distances and angles between amino acids in proteins and look for correlations between that and the initial amino acid sequences. 

All information is stored as a number. To teach the computer, it’s given 1000’s of amino acid sequences we know the protein structure and repeats the following process for each. Each channel has a weight and neurons have a bias. Both are initially randomly generated numbers. We multiply the weight with the previous layer and add it to the bias. Plug it into the activation function, ReLu to determine if it's useful or not. Repeat this once again, but instead of plugging this into ReLu, we use a skip connection to add this layer back and THEN plugging this into Relu. Do this until the output layer.  If the final results are wrong, the computer adjusts its weights and biases. If perfected, plugging in any amino acid sequence, such as covid-19’s, would yield the correct protein model. 

Unfortunately, this is not perfected yet but, there is great promise for it. One day, AI may work jointly alongside scientists to stop pandemics altogether. Until then, stay safe and stay healthy!