AlphaFold Predictions Shown to Be as Reliable as Experimental Structures

AlphaFold is transforming how researchers study proteins. By rapidly predicting a protein's unique 3D structure from its amino acid sequence, AlphaFold could dramatically accelerate drug discovery pipelines that have traditionally relied on laborious experimental methods. However, before AlphaFold can transform biomedicine, scientists need evidence that its predicted protein structures are as accurate as those obtained through techniques like X-ray crystallography. A new study in Science suggests AlphaFold may meet this bar.

Rockefeller University researchers used AlphaFold2, an earlier iteration of the AI, to predict the structures of two drug target proteins. They then employed sophisticated software to screen billions of compounds, looking for potential drug candidates that could bind to these predicted structures.

They found the AlphaFold2 structures performed just as well as experimentally-determined structures for identifying promising drug leads. This contradicts previous work suggesting AlphaFold would underperform experimental data for drug discovery applications.

"We estimate in about one-third of cases, an AlphaFold-predicted structure could significantly expedite a project by up to a few years compared to obtaining a new experimental structure," explains study author Jiankun Lyu.

The latest version, AlphaFold3, represents a major upgrade—able to predict not just single protein chains but complexes with other molecules like DNA, RNA, and potential drugs. However, its developers have not publicly released the model code, limiting academic research.

Lyu advises cautious optimism about AI's role in drug discovery moving forward. While deep learning shows immense promise, its accuracy depends on the availability of high-quality experimental data to train the algorithms.

"Where AI is already succeeding happens to be in those areas in which basic science has generated a lot of data experimentally," Lyu notes. "So now that we have many AI architectures, we need to go back to the bench and generate more high-quality data, to feed these data-hungry algorithms until they produce better predictions. That’s when the breakthroughs will come."