Camille and Henry Dreyfus Distinguished Chair in Chemistry Vijay Pande and his students have published findings that indicate potential for deep learning algorithms to boost the field of drug development. The team found success in a subset of machine learning known as “one-shot learning algorithms” to help in the decision making processes involved in developing new drugs.
Most machine learning algorithms require millions of data points to train computers, a challenge that one-shot learning overcomes.
“We’re trying to use machine learning, especially deep learning, for the early stage of drug design,” Pande told Stanford News. “The issue is, once you have thousands of examples in drug design, you probably already have a successful drug.”
Pande directs Stanford’s Pande Lab, which uses computer simulation, mechanics and statistics to tackle problems in chemical biology and biomedicine.
One-shot learning is a categorization of machine learning algorithms that aims to discover information and useful patterns using a limited number of data points. This allows the development and creation of new processes without the need for an exhaustive list of tests and supporting data.
To see whether one-shot learning would be successful with drug design, the researchers first reorganized the molecular information of the testing drugs into graphs using the connections between atoms. This made the molecular data more digestible to supply their algorithms with inputs.
The researchers then used one-shot learning to train an algorithm on two datasets — one to learn about the drugs’ chemical toxicity and another to learn about side effects of existing drugs. The predictions made by the algorithms following their training showed improved accuracy over predictions made with random chance.
“We worked on some prototype algorithms and found that, given a few data points, they were able to make predictions that were pretty accurate,” said fifth-year computer science Ph.D. student Bharath Ramsundar. Ramsundar, a researcher at the Pande Lab, co-authored the study published in ACS Central Science.
The group envisions algorithms like these as “an experimentalist’s helper,” according to Ramsundar. This automation will not be replacing the human scientists designing experimental drugs but will rather give a computational basis to support the decisions they make early on in research. Deciding which molecules to pursue from a set of candidates, for example — a choice now made from intuition — may be guided by predictions from the algorithms.
“This paper is the first time that one-shot has been applied to [drug design],” Pande said. “This is not the end of this journey — it’s the beginning.”
Contact Susannah Meyer at smeyer7 ‘at’ stanford.edu.