Finding Drugs with our eyes

I worked in drug discovery for a few years. From testing out my own personalized caspase products and engineering my own biological and small molecules, to working at a biotech company that used machine learning to find non-coding mRNA drug candidates, there are a few common patterns I found in the industry. Drug discovery is notoriously slow, expensive, and extremely risky. Out of thousands and thousands of potential candidate molecules, only a handful ever make it to market—It’s like finding a needle in a haystack.

By increasing the hit rate, rate of successfully identifying a promising candidate, removing hay to find the needle, we can transform the economics of pharmaceutical R&D. Scientists created high throughput screening (HTS) in 1980, just less than 50 years ago. We are now using machine learning to compliment HTS, significantly changing the game in finding the needle.

Using ML in HTS, we can rapidly test millions of compounds and variants in parallel. It’s like having a massive “searchlight” sweeping over chemical space. ML guided screening allows for models to learn from previous experients to prioritize candidates, take in real world data and human genetics to take into account differences (chemical and biological landscapes) and translational concerns for homogenous searches, increasing a higher likelihood of success.

Google DeepMind has taken advantage of this using their TeleProt framework, which currently outperforms standard directed evolution approaches in finding high-activity enzyme variants. DeepMind’s TeleProt discovered a nuclease variant with 19× higher activity than the original at physiological pH, using fewer rounds and fewer wasted experiments than traditional methods.

Nuclease is an enzyme that cleaves, or “breaks” DNA and RNA. It’s what is used in CRISPR and other genetic engineering platforms that allow us to treat very rare and life threatening diseases.

How does this work?

Think of it like human vision. Vision is one of the most incredible senses we have, allowing us to perceive and interpret the world around us through light. What seems to be an effortless process involves a complex journey that begins with light (input) and ends in the brain (output).

When light enters the eye through the cornea, the clear, dome-shaped surface that covers the front of the eye, the cornea bends (refracts) the light to help focus it. The light passes through the pupil, the black circular opening in the center of the eye which adjusts automotically to control light intensity. Behind the pupil lies the lens, a flexible, transparent structure that further focuses the light. Similarly in HTS, the instruments capture “raw sensory input” of large data and has instructions to personalize and focus the data coming in.

After passing through the lens, light is focused onto the retina, a thin layer of specialized cells at the back of the eye, which acts like the film in a camera, capturing the image thorugh photoreceptors called rods and cones. These are converted into electrical signals that carry detailed information about the image, such as brightness, color, and movement. The signal travels to the brain where it processes and interprets the electrical signals, combining input from both eyes to create a coherent and three-dimensional image of the world. Similarly in machine learning, the data is transferred to a computational domain, and like the brain, you use ML models to process the raw data, breaking it down to meaningful “images” just like how your brain detects edges, shapes, and colors. You start to create pattern recognition and improve decision-making to predict successful objects, or in DeepMind’s case, nuclease candidates.

This approach shifts the bottleneck in drug discovery from blindly searching chemical space to intelligently navigating it, the same way your brain focuses attention on the most relevant features in a complex visual scene.

If you are interested, here is an article on how Google Deepmind is using ML to find highly active nucleases: https://www.sciencedirect.com/science/article/pii/S2405471225000699