March 2023 - April 2023
Source: https://www.chemistryworld.com/news/ai-tool-screens-107-million-molecules-discovers-potent-new-antibiotics/4011233.article
Active learning is a technique that iteratively selects a subset of data from a large sample of unlabeled data to be labeled by a domain expert or oracle. The newly labeled points are then used to train machine learning models. These samples are chosen in a way that minimizes the number of samples needed to label in order to model the data. Screening for potency typically involves working with several small molecules, which can be an expensive process at the lab-scale. Conducting active learning can decrease the number of potential target molecules and reduce expenditure by a significant margin.
We worked with a set of state-of-the-art techniques that not only allow us to detect small molecules with antibacterial activity, but also to actively screen for them. One such approach that was solely implemented by me is the Deep Reinforcement Approach for Active Learning. This method has not been studied a lot but we believe it has a lot of potential for application in a biological context.
Model Architecture for Deep Reinforcement Active Learning
We compared the RL approach to a simpler Query-by-Committee method and Random Sample, and observed much higher accuracies for RL.
Performance of RL is much better than any other model
We also tried study the exploration vs exploitation relationship, and found it useful to have exploration in the initial stages and then switching exploitation. In the figure below, we do 80% exploration till 100 samples in the observed dataset and then switch to 10% exploration.
Python, sklearn, PyTorch, AWS, Tensorflow, Keras, RDKit
Mirudhula Mukundan, Sriram Kidambi, Daniel Lee