AI platform enables million-fold increase in drug screening speed

Chinese researchers have developed an artificial intelligence-powered virtual drug screening platform that achieves a millionfold increase in screening speeds and a significant improvement in prediction accuracy over conventional methods, according to Tsinghua University.

The platform, DrugCLIP, has allowed the research team to conduct the first-ever virtual screening project on a human-genome scale, opening new avenues for innovative drug discovery. The related study was published online in Science on Friday.

Currently, the vast majority of the human genome's potential targets and compounds remain untapped. The human genome encodes over 20,000 proteins, and only a small fraction have been explored in drug targets. Screening for promising lead compounds across this vast chemical and target space has become a major bottleneck in the field.

While traditional docking tools could take months to screen just 10,000 protein-ligand pairs on a single workstation, AI-driven models can complete these tasks in a fraction of the time, overcoming a longstanding hurdle in matching novel targets with potential drugs.

DrugCLIP reduces this timeframe dramatically, compressing what would take years into a single day on one computing node — a fundamental unit in a high-performance or distributed computing system.

Previously, finding a new drug was like trying to find a specific key for a specific lock by physically trying every key in the world one by one. DrugCLIP is able to "look" at the key and the lock simultaneously and predict the fit almost instantly.

While AlphaFold, whose developers won the 2024 Nobel Prize in Chemistry, has addressed the challenge of protein structure prediction, DrugCLIP makes a complementary leap by establishing a critical bridge from protein structure to drug discovery, enabling large-scale virtual screening across the human genome.

At its core, DrugCLIP's breakthrough is in reformulating traditional molecular docking as a high-efficiency semantic search for protein pockets and small molecules in vector space. Running on a computing node with a 128-core CPU and 8 GPUs, the platform can score trillions of protein pocket-small molecule pairs daily, delivering a million-fold speed increase over conventional docking approaches.

In its inaugural genome-scale screening project, the team applied DrugCLIP to approximately 10,000 protein targets and 20,000 protein pockets, analyzing over 500 million small, drug-like molecules. The effort yielded more than 2 million potential active molecules, resulting in the largest known protein-ligand screening database to date.

This database is now freely available to the global scientific community, offering powerful data support for fundamental research and early-stage drug discovery.

A companion screening service platform has also been launched, allowing users to submit custom targets and protein pockets for analysis. As of the study's publication, the platform has served over 1,400 users and completed more than 13,500 screening tasks within a span of six months.

The DrugCLIP team said it plans to collaborate closely with academic and industry partners to accelerate the discovery of novel targets and first-in-class therapies in areas such as oncology, infectious diseases and rare disorders.

The team added it will continue to enhance the platform's performance, expand its capabilities, and contribute to a more intelligent, efficient and accessible global ecosystem for drug innovation.

Xinhua