Imran Haque, SVP of AI and Digital Sciences at Recursion, walks us through the treasure trove of data found in images, explaining how we use algorithms to structure pixels, how we are leading discovery in the field, and the massive advances that have been made in training foundation models to make new predictions of promising therapeutic targets.
Some highlights:
• Imaging and algorithms can provide as much or more information than RNA sequencing and cost far less money, so you can run more experiments - and you don’t have to kill the cells to do it.
• With ML foundation models acting as a lever, we only need a small amount of data to perform as well as a much larger set. With areas as vast as chemistry and biology, you want to build as much as you can to break these scale barriers.
• The models - called masked autoencoders - can reconstruct some 75-90% of the images of cells that have been masked .
• Once trained, we use these models to turn an unmasked image into a list of numbers and then match these to the database of images of all the cells we’ve taken in the past to see how similar or different the biology is that is created by different conditions (or perturbations) for those cells..
• This technique - called similarity comparison - is central to our digital maps of biology. It’s how we discovered RBM39 provided a new means to inhibit CDK12, a promising therapeutic target for cancer.
#ai #ml #tech #techbio #chemistry #biology #models