Researchers at MIT created DiffDock, a model that has the potential to uncover new medications more quickly than conventional methods and with fewer adverse side effects. 

DiffDock develops a library of protein-ligand binding poses, which has the potential to significantly alter the current drug discovery process as it is now understood. Protein molecules in various colours are depicted here as they bind to their counterparts in grey.

The model's distinct approach to computational drug design represents a paradigm shift from the current state-of-the-art technologies used by most pharmaceutical companies, presenting a significant opportunity for a comprehensive overhaul of the traditional drug development pipeline.

Overview

Drugs often work by interacting with proteins found in our bodies, as well as proteins found in bacteria and viruses. Molecular docking predicts the atomic 3D coordinates of a ligand (i.e., drug molecule) and protein bonding to understand these interactions. 

While molecular docking has resulted in the successful identification of drugs that now treat HIV and cancer, with each drug taking a decade to develop and 90 per cent of drug candidates failing costly clinical trials (most studies estimate average drug development costs to be between $1 billion and $2 billion per drug), it's no surprise that researchers are looking for faster, more efficient ways to sift through potential drug molecules. Currently, most in-silico drug design molecular docking tools employ a "sampling and scoring" approach to find the optimal ligand "pose" for the protein pocket. This time-consuming technique examines and ranks many alternative postures depending on how efficiently the ligand binds to the protein.

Structure

It is helpful to describe how generative diffusion models function using image-generating diffusion models. In this case, diffusion models gradually add random noise to a 2D image in a succession of steps, eroding the image's data until there is nothing but grainy static. Next, a neural network is trained to retrieve the original image by reversing the noising process. Finally, the model can produce new data by beginning with a random configuration and progressively reducing noise.

In the case of DiffDock, the model can effectively detect multiple binding sites on proteins that it has never met before after being trained on various ligand and protein configurations. Then, instead of creating new picture data, it makes new 3D coordinates to assist the ligand in determining viable orientations for fitting into the protein pocket.

Conclusion

These advancements provide new opportunities for biological research and drug development. In phenotypic screening, for instance, researchers observe the effects of a drug on a disease without knowing which proteins the drug is acting upon. The discovery of the drug's mechanism of action is crucial for comprehending how the drug can be enhanced and its potential side effects. Unfortunately, this process, known as "reverse screening," can be complicated and expensive. Still, a combination of protein folding techniques and DiffDock may enable a large portion of the process to be performed in silico, allowing potential "off-target" side effects to be identified early on before clinical trials.

Image source: Unsplash