DeepMind’s latest AlphaFold model is designed to predict more than protein structures. The system can generate predictions for nearly all molecules in the Protein Data Bank, and its expanded capabilities may give drug researchers another way to study how biological molecules interact.
From protein structures to molecular interactions
AlphaFold first drew attention for predicting structures of many proteins in the human body. DeepMind later released AlphaFold 2 in 2020. The newest model extends that work to other kinds of molecules involved in biology.
DeepMind says the model can predict structures for ligands, which bind to receptor proteins and can change how cells communicate. It can also predict nucleic acids, which contain key genetic information, and post-translational modifications: chemical changes that occur after a protein is created.
This wider scope matters because biological molecules do not act in isolation. Understanding how a ligand fits with a protein, for example, can help researchers identify or design molecules that might eventually become drugs. A structural prediction gives scientists a way to examine those possible interactions computationally.
A different route from docking methods
Pharmaceutical researchers currently use computer simulations called docking methods to assess how proteins and ligands might interact. These methods need a reference structure for the protein and a proposed position where the ligand might bind.
The new AlphaFold model is intended to work without those inputs. DeepMind says it can make predictions for proteins that have not been structurally characterized and model how proteins and nucleic acids interact with other molecules. That could let researchers investigate some molecular relationships even when they lack the starting structure or binding position that docking requires.
In drug discovery, that difference could expand the questions scientists can explore with computer models. Instead of beginning with a known protein structure and a suggested binding site, they may be able to use the model to predict structures and interactions together. Such predictions can support investigation and design; they do not, by themselves, establish that a molecule will work as a treatment.
Early results and remaining limits
DeepMind reports that early analysis found the newest model substantially outperformed the previous generation on some protein structure prediction problems relevant to drug discovery, including antibody binding. The company presents this improvement as evidence of AI’s potential to strengthen scientific understanding of the molecular systems in the human body.
There is also a clear limitation. In a whitepaper describing the system’s strengths and weaknesses, researchers at DeepMind and Isomorphic Labs say it does not match the best-in-class method for predicting RNA structures. RNA molecules carry instructions for making proteins, so this gap leaves an important area of molecular prediction less well served.
The report describes a system with broader reach, not a finished answer to every structural problem. Its value for drug discovery will depend on the quality of its predictions for the molecular tasks researchers need to study, as well as on how scientists use those results in later work.
Isomorphic Labs is applying the model to drug design
Isomorphic Labs, a DeepMind spin-off focused on drug discovery, co-designed the new model and is already applying it to therapeutic drug design, according to a post on the DeepMind blog. The company is using it to characterize different molecular structures that matter in treating disease.
That work connects the model’s technical capabilities with a practical research aim: improving how scientists examine possible therapeutic molecules. AlphaFold’s expanded predictions may help researchers explore interactions that older docking approaches could not address in the same way, while its RNA shortfall shows that important challenges remain.