We have developed an Artificial Intelligence model which is one of the best applications of AI (Artificial Intelligence) for pharmaceutical companies. This model take drug names as inputs and according to chemical and biological features and examining different types of drugs like enzymes, chemical etc. It also predicts side effects which are associated with new drugs which can be made from two or more drugs. So, let’s see some insight about this model. How this model is working and what are the features and concepts are used to make this model work in an efficient way.
Data set used for AI model
This model is built using deep neural networks which is a multi-perceptron model. Hereby every neuron, we mean some kind of function and activation function for that layer. These neurons are inspired by the human brain and learn new things in the same way.
For training our model of neural nets data of 832 Medicines is used where each medicine have 40260 features. We have used two hidden layers in our program which 2200 and 202 neurons in each layer consecutively.
Training of Model
Now, if we come to training part that which type of function and activation functions are used for this model. In this model for training we used method of Logistic Regression which is also known as ‘logit’ model. In the beginning, input layer have 40260 neurons for input layer and 1385 neurons in output layer which tells about probability of a side effect to occur out of 1385 classes of side effects. Every neuron has ReLu function as its activation function. At the output layer we used softmax function for multi label output.
All we have to is deciding weights for every feature and through back propagating we need to find that. If we have back propagating through neural network in order to improve weights of inputs we use optimization techniques to reduce loss (Measure of difference between actual and predicted value. We have used Adam Optimizer for reducing cost (loss).
Functioning of Model
This model takes drug IDs as inputs from API’s and then take feature matrix of all drugs given from ‘metrices.mat’ file available with data set and then tries to find the features matrix of a drug which can be formed from these drugs. After evaluating feature matrix of a new drug, it is given to neural network to predict its side effects. Neural networks give 1385 values at each neuron in output layer and we take top probable values from output layer and gives output as side effects which are more likely to happen. This model has achieved its accuracy up to 80% for this small data set and can be extended further with larger data sets.
This AI model is can be used by pharmaceutical with large amount and data and thousands of medicines and features to predict and evaluate every new medicine without actually making it which will surely save time and cost. And more accurate results can be obtained.