Galaxy VS: A new AI-powered application that screens 100 billion drugs in seconds

GalaxyVS: A New Era in Drug Discovery

In an exciting development for the field of drug discovery, a new AI-powered application called GalaxyVS is making waves. Developed at the National Supercomputing Center in Tianjin, this innovative system is designed to be able to screen a staggering 100 billion compounds in just seconds, a feat that could significantly speed up the drug development process.

GalaxyVS is not just your average molecular docking tool. It combines hardware and software in a way that enhances its performance. At its core, it utilizes a technique known as dense vector retrieval, drawn from a framework called DrugCLIP. This method allows it to overcome some of the limitations faced by traditional docking applications such as Vina and Gninas, which have often been slow and cumbersome.

One of the standout features of GalaxyVS is its ability to handle the complexities of data processing. It addresses common challenges such as encoding representations offline and managing memory and input/output bottlenecks. This is achieved through a robust supercomputing infrastructure and a two-stage refinement protocol. This protocol not only ensures diversity in the compounds being screened but also employs a method called AlphaRank for re-ranking based on affinity, which helps in identifying the most promising candidates.

The performance of GalaxyVS is impressive, achieving a daily throughput of 1.5 × 10¹⁶ target-ligand pairs. In practical terms, this means it can screen nearly 100,000 protein structures across six different species in just 16 hours. This rapid processing capability offers researchers a comprehensive view of potential interactions, paving the way for more effective drug candidates.

Moreover, the results generated from GalaxyVS are compiled into a resource known as GalaxyDB, which will be made openly accessible to facilitate further research and accelerate data-driven drug discovery. This move not only enhances collaboration among scientists but also democratizes access to valuable data in the quest to develop new therapies.

In summary, GalaxyVS represents a significant advancement in the drug discovery landscape, combining cutting-edge technology with a user-friendly approach. As researchers continue to tackle complex diseases, tools like GalaxyVS will be crucial in identifying new and effective treatments at an unprecedented pace.

Beyond the Supercomputer Hype: How “GalaxyVS” Screened 100 Billion Compounds in Under a Day

Virtual drug screening, the process of using computers to test whether chemical compounds might bind to a disease-causing protein, has historically been a massive logistical headache. If you want to sift through billions of molecules using traditional molecular docking (which simulates the physical “lock-and-key” fit of a molecule into a protein pocket), you are going to need millions of computer cores and weeks of processing time. For example a traditional docking screen of only 8 million compounds recently performed by the Bond Lab, using the University of Bristol supercomputer BlueCrystal, took about 3 months of processing time.

A team of researchers from Tsinghua University and the National Supercomputing Center in Tianjin are trying to change that. In a recent preprint paper, they introduced GalaxyVS, a virtual screening framework designed to explore a 100-billion-compound chemical space in seconds.

Deployed on the Tianhe supercomputer, the system recently wrapped up a massive cross-species screening campaign. It managed to evaluate nearly 100,000 protein structures against a 100-billion-compound library in just 16 hours. A feat that resulted in an open-access dataset called GalaxyDB.

Shifting the Bottleneck Offline

To understand why GalaxyVS is fast, you have to understand why traditional methods are slow. Classic molecular docking requires a computer to sample multiple shapes (conformations) of a molecule and physically calculate its binding energy against a high-resolution protein structure. GalaxyVS bypasses this by treating drug screening like a search engine query. It builds on an AI paradigm called DrugCLIP, which uses a dual-encoder setup. One encoder converts the 3D features of a chemical compound into a long string of numbers (a dense vector embedding), while the other encoder does the same for a protein’s target pocket.

Because these two encoders operate independently, the heavy lifting can be done ahead of time. The researchers paid a “one-time sunk cost” to map out their massive library. This involved:

• 1,000 CPU nodes ran for 12 days to generate 3D shapes of the molecules.
• 5,000 accelerator nodes spent 18 days translating those shapes into vector math.

Once that 100-billion-molecule library was built and saved to disk, running an actual screen became incredibly lightweight. Instead of simulating physics on the fly, the system simply runs a mathematical comparison (cosine similarity) between a new protein pocket vector and the pre-computed drug vectors.

Two Ways to Run a Screen

Supercomputers are great, but most everyday biology labs don’t have uninhibited access to tens of thousands of server nodes. To address this, the developers built two distinct operational modes into the framework:

The Accessible Mode works by using a tool called PipeANN, which searches the data directly from hard drives rather than demanding hundreds of terabytes of RAM to keep the whole library live.

The Extreme Mode achieves its speed by proactively slicing the 100-billion library into balanced shards and pre-loading them into local memory caches across the supercomputer, avoiding data traffic jams on the shared storage drives.

Tackling “Diversity Collapse” and Finding Novel Fits

In virtual screening, a massive library can occasionally be a disadvantage. If you simply ask a computer for the top 30,000 matches out of 100 billion, you often get “diversity collapse”: thousands of virtually identical molecular cousins that share the exact same structural backbone.

To prevent this redundancy, GalaxyVS groups its library into roughly 100,000 structural families ahead of time. During a search, a dynamic diversity filter caps how many candidates can be pulled from any single structural cluster.

When tested against standard screening benchmarks, this filtering method pulled a significantly higher variety of unique molecular scaffolds and fragments compared to conventional million-scale libraries.

[100B Compound Space]
       │
       ▼  (Stage 1: Fast Vector Retrieval via PipeANN)
[~10M Raw Candidates] ──► (Diversity Filter caps redundant molecular families)
       │
       ▼  (Stage 2: Precision Re-ranking via AlphaRank)
[Chemically Diverse, High-Affinity Binders]

Furthermore, because vector searching trades fine-grained physical modeling for pure speed, the researchers added a secondary refinement step. The diverse candidate pool is passed to AlphaRank (an affinity prediction model built on the AlphaFold3 backbone), which explicitly analyses 3D atomic interactions to make sure the final suggestions hold up under physical scrutiny.

What’s Next?

To prove the platform’s stability, the researchers ran the entire structural proteomes of six species—including E. coli, yeast, fruit flies, mice, and humans—through the system. Looking at roughly 100,000 proteins and evaluating over 4 million potential binding pockets, the system completed 10¹⁷ pocket-ligand evaluations in 16 hours.

The resulting interaction data, dubbed GalaxyDB, has been made open-access to help scientists look for everything from new human therapeutics to better antimicrobials and pesticides. The software framework successfully moves the heaviest logistical hurdles of ultra-large screening to a one-time offline process. However, as the authors note in their conclusion, the ultimate test for these virtual candidates will be moving them out of the supercomputer and into physical wet labs for experimental validation.