The New Age of Computational Drug Discovery

Close-up view of a DNA double helix structure with a dark, blurred background.

What is needed to design a new drug?

Pathological pathways and structures of involved proteins are essential. Solving protein structures is one of the oldest problems in biological sciences and typically involves expensive experiments and much trial and error.

For over 50 years, researchers have attempted to build computational models that could predict protein structures from the underlaying protein amino acid sequence. Today the problem appears to be solved.

The latest Nature and Science headlines read, “It will change everything” and the “Game has Changed.” So the game changed, but why and how exactly?

For 45 years, academic and industrial groups have tried to apply physics, statistics, math, computational modeling, biology, chemistry, and everything else they could think of to the problem of predicting structures of proteins in the computer.

But in 2016 Deepmind had some fun, and in an internal hackathon realized that the amount of data about proteins available is sufficient to train a neural network. So they put highly paid professionals and best-in-class computational resources to the task, and in 2018 they outperformed all academic groups in a highly competitive blind test called CASP. Did they stop here? No, in 2020, they revealed an improved version of their deep learning method that consistently predicted protein structures with experimental quality.

Yes, it is possible to computationally predict experiment-quality protein structures!

Or is it?

Deepmind is a subsidiary of Google and has patents filed for the new algorithm. It can cost 1-3 years of PhD time to solve a protein structure by experiment, but it is possible within days with their algorithm. What would you do, if you had such a solution at your fingertips? Give it away for free?

Now we start to see how the game might change. Either Google will use the competitive advantage to develop drugs for unknown targets, or, more likely, they are going to sell access to their models for big $$$. This is the first of many future success stories of experts applying deep learning with access to the right data and computational power.

Academia will likely not be able to compete going forward, but the industry could if they knew how. ZONTAL has some of the world’s leading experts on scalable data management for artificial intelligence applications. We make data FAIR so that your teams can train the next Alphafold.

Get in touch to learn more!