UniLingo3DMol: a unified 3D molecular language model for de novo and fragment-based drug design

The search for new medicines is a monumental task, often compared to finding a specific needle in a haystack that is constantly changing shape. For decades, scientists have relied on two main strategies to design these needles. One approach looks at the chemical shape of molecules that are already known to work, tweaking them slightly to see if they become better. The other approach looks at the three-dimensional shape of the disease target itself, like a lock, and tries to design a key that fits perfectly into its hollows. While both methods are useful, they have traditionally been kept separate. Artificial intelligence has begun to help with both, but most computer programs are specialized: they are either good at tweaking existing chemicals or good at designing new shapes for a specific lock, but rarely both at the same time. This separation forces researchers to switch tools and mindsets, slowing down the discovery process and preventing the computer from learning the full picture of how a drug interacts with a disease target.

A team of researchers has now introduced a new artificial intelligence system called UniLingo3DMol that bridges this gap. Instead of using separate tools for different stages of drug design, this system uses a single, unified language to describe molecules in three dimensions. It can generate entirely new drug candidates from scratch, or it can take a specific piece of a known drug and build a new molecule around it, all while keeping the complex three-dimensional shape of the target disease in mind. The researchers tested this system on over one hundred different biological targets and found that it could produce drug-like molecules that fit their targets better and more accurately than existing computer models. To prove it worked in the real world, the team used the system to design a new inhibitor for a protein called CBL-B, which is involved in regulating the immune system. The resulting compound showed strong activity in laboratory tests and, when combined with an existing immunotherapy antibody, successfully shrank tumors in mice by 76 percent.

The core innovation of this work lies in how the computer "sees" a molecule. Traditional methods often describe a molecule as a flat list of atoms and bonds, which makes it difficult for the computer to understand how that molecule would look in three-dimensional space when it tries to fit into a protein. The new system uses a dual-sequence approach. It breaks a molecule down into meaningful chunks, like rings of atoms or side chains, and then describes how those chunks connect to one another using a set of pointers. This allows the computer to rearrange the order of the chunks without breaking the molecule, much like shuffling a deck of cards while keeping the suits intact. This flexibility is crucial because it allows the system to learn from two very different types of data at once: the vast chemical space of all possible molecules, and the specific, high-resolution structures of molecules that are already known to bind to proteins.

To teach the system, the researchers built a massive database called RComplex, containing millions of protein-ligand complexes. They organized this data into three levels of confidence, ranging from the most certain structures, which were determined directly by X-ray crystallography, to less certain structures that were predicted by computer docking. The training process happened in three stages. First, the system learned the basic rules of chemistry and how atoms arrange themselves in space using a library of eight million virtual molecules. Next, it learned how these molecules interact with protein pockets, using the less certain but abundant data from the database. Finally, it refined its understanding using only the highest-quality experimental structures. This step-by-step approach ensured the system learned the fundamental language of chemistry before trying to master the complex dialect of protein binding.

The researchers evaluated the system against six other leading artificial intelligence models using a standard set of one hundred and two protein targets. The new system outperformed all of them. It generated a higher percentage of molecules that looked like real drugs, meaning they had the right balance of size, complexity, and chemical properties. More importantly, the three-dimensional shapes it produced fit into the protein targets with greater precision. When the researchers checked if the generated molecules could reproduce the binding patterns of known active drugs, the new system succeeded in more than seventy percent of cases, a significant jump from the less than thirty-five percent success rate of the other models. It also worked much faster, generating five hundred valid molecules in about thirty seconds, while the other models took anywhere from thirteen minutes to over three hours to do the same task.

To demonstrate that this was not just a computer exercise, the team applied the system to a specific challenge: designing an inhibitor for CBL-B. This protein is an attractive target for cancer immunotherapy because it acts as a brake on the immune system; turning it off could help the body fight cancer more effectively. The team started with a known inhibitor bound to the CBL-B protein and asked the system to generate new molecules that kept two specific parts of the original drug but changed the rest to find a better fit. The system generated hundreds of thousands of candidates, which were then filtered down to a few promising leads. One of these, a compound they named Cmpd. 7, was synthesized and tested. It showed strong activity in the lab, and its crystal structure confirmed that it bound to the protein exactly as the computer had predicted.

Building on this success, the researchers used the system again to optimize the compound further. They kept the parts of Cmpd. 7 that were working well and asked the system to redesign the remaining section to improve its interaction with the protein. This round of generation produced a new lead, Cmpd. 20. When tested in mice with tumors, Cmpd. 20 alone showed some activity, but when combined with a standard PD-1 antibody, it resulted in a 76 percent reduction in tumor growth. The compound also showed a favorable safety profile in early tests, with low inhibition of certain liver enzymes that often cause drug interactions. The entire process, from the initial computer generation to the final animal testing, highlighted how a unified approach can streamline the path from a digital idea to a potential medicine.

The study also explored the limits of the current technology. While the system is highly effective, it relies on static images of proteins, which are actually flexible and move around in the body. The researchers noted that future versions could incorporate the movement of proteins to make the designs even more robust. They also suggested that the system could be improved by retrieving specific chemical fragments from external databases in real-time, allowing it to adapt to new scientific discoveries without needing to be retrained from scratch. These potential upgrades point toward a future where artificial intelligence acts as a continuous partner in drug discovery, constantly learning and refining its designs as new data becomes available.

The success of UniLingo3DMol suggests that the future of drug design may not lie in specialized tools for specific tasks, but in unified systems that can handle the entire process. By teaching artificial intelligence to speak a language that encompasses both the chemistry of the molecule and the geometry of the target, the researchers have created a tool that can navigate the complex landscape of drug discovery more efficiently than ever before. The ability to generate high-quality, three-dimensional drug candidates in seconds, and to validate them in living organisms, marks a significant step forward in the application of artificial intelligence to human health. The work demonstrates that when computer models are designed to mimic the holistic way human experts think about drug design, they can produce results that are not only theoretically sound but practically effective.