AI and molecular dynamics combine to map drug-protein interactions in cryo-EM

Seeing a drug bind to its target protein at atomic resolution has long been one of structural biology's central puzzles. Cryo-EM—the technique of freezing proteins in place and imaging them with electrons—has revolutionized how we visualize these molecular machines. But there's a persistent problem: while the protein itself often comes into sharp focus, the small molecule drug nestled in its pocket frequently remains blurry, its exact position ambiguous. A team of researchers has now shown how to sharpen that blurry picture by combining three tools: artificial intelligence trained to predict protein structures, the actual experimental cryo-EM images, and molecular simulations that let atoms move and settle into their true positions.

The researchers started with a straightforward question: what if you fed an AI system just the protein's amino acid sequence and a chemical description of the drug, then used the experimental cryo-EM map to refine what the AI predicted? They tested this on ten real pharmaceutical targets—two kinases involved in cancer and Parkinson's disease, two receptors that regulate inflammation and allergies, and six transporters that move molecules across cell membranes. None of these complexes had been seen by the AI during its training, so this was a genuine test of whether the approach could work on new systems.

The pipeline worked in three stages. First, they used Chai-1, an open-source AI model with architecture similar to AlphaFold3, to generate five possible predictions of how the drug and protein fit together. The AI needed only the protein sequence and the drug's chemical formula—no prior knowledge of where the drug actually binds. Second, they aligned the best-looking prediction to the cryo-EM map using standard rigid-body fitting. Third, for cases where the AI's guess wasn't quite right, they ran molecular dynamics simulations guided by the experimental density map, allowing both the protein and drug to shift and adjust until they matched the observed structure as closely as possible.

The results were striking. In three cases—two kinases and one receptor—the AI prediction alone was already accurate enough, achieving 90 percent or better agreement with the experimentally determined structure. For seven other cases, the initial prediction was rougher, with accuracy ranging from 40 to 71 percent. But after running the density-guided simulations for just two nanoseconds, accuracy jumped to 82 to 95 percent. The protein structures improved too: pocket residues went from 63-86 percent accuracy to 84-96 percent, and the full protein from 71-89 percent to 84-99 percent. In one particularly stubborn case—a large transporter bound to an estrone sulfate molecule—the initial prediction had the drug flipped relative to its true orientation. The researchers generated twenty additional AI predictions and found one with better chemical plausibility, which then refined cleanly to 92 percent accuracy.

What makes this approach powerful is its accessibility. A researcher doesn't need to know where the drug binds, what the protein's functional state should be, or how to manually build a molecular model. The pipeline requires only the sequence, the drug's chemical structure, and the cryo-EM map. The method also sidesteps a common bottleneck: traditional flexible-fitting approaches often require a pre-built protein structure and a masked density map showing only the drug. Here, the protein and drug refine together, accommodating conformational changes that occur when a molecule binds. This matters especially for transporters, which cycle between different shapes as they move cargo across membranes—a challenge that stumped earlier methods but that the density-guided simulations handled with reasonable accuracy.

The researchers acknowledge limitations. The approach depends on the AI generating an initial model that's at least in the right ballpark—within about four angstroms of the true structure. If the AI produces something wildly wrong, flexible fitting alone may not rescue it. Classical force fields used in molecular dynamics simulations can also be imperfect for small molecules, though the strong forces from the cryo-EM density map help compensate. As AI structure prediction continues to improve, and as researchers develop enhanced sampling strategies and iterative refinement approaches, these constraints should loosen. For now, the pipeline offers a straightforward, semi-automated way to decode how drugs recognize and bind their protein targets—a capability that could accelerate both basic research into molecular recognition and the practical work of designing better medicines.