GFlow Neural Networks to help accelerate molecules, candidates generation: experts

HANOI, Vietnam, Nov 28: The demand to find new ways to combat disease is rising by the day. Humanity is always at risk of a new pandemic, and the mutation of viruses creates resistance to antibiotics. According to experts, it has caused both high fatalities and economic value losses. “There’s already 1.2 million deaths per year, and it’s going to grow to 10 million deaths per year,” said Professor Yoshua Bengio (Mila Quebec AI Institute). “Economic costs are also rising, and it’s projected to be 100 trillion US dollars by 2050.

To combat this, prof. Bengio has been looking into utilizing Generative Flow Networks, or GFlow Nets - his team’s ML technique for generating compositional objects at a frequency proportional to the associated reward - to discover new drug molecules and generate candidates. These findings were published in 3 recent papers at renowned AI conferences.

According to prof. Bengio, one of the greatest areas of growth is at the intersection of AI and biotechnology for the next decade, thanks to its ability to reprogram the DNA of organisms and synthesize new drug molecules.

However, the number of potential new drugs is vast and it takes decades to test which drug would work in treating which subject. Here, ML can be used to represent sample candidate experiments and shorten the time to give an educated guess as to “what chance a candidate drug is going to do the job”. But questions arise regarding ML ability to acknowledge its limitations and uncertainty, and to create diversity in candidates.

“We would like our ML system to generate candidates to be different from each other,” said prof. Bengio. “It’s important as if one drug candidate does not work, “we still have other candidates that are quite different. At the end of the day, we have a greater chance of having a drug that’s going to work.”

To decide which candidates to zoom in on since the number would be too large, prof. Bengio and his team propose using generative models, benefiting from neural nets’ ability to imagine - usually for synthesizing new images. In this case, instead of images, neural nets can synthesize molecules, and through training can be used in scientific experiments.