A large library of chemical compounds may provide potential treatment methods for various diseases such as cancer or heart disease. Ideally, scientists would like to experimentally test each compound in order to target all possible receptors, but this screening process is time-consuming.
In recent years, researchers have started using computational methods to screen these libraries in order to expedite drug discovery. However, many of these methods also require a long time because most of them calculate the three-dimensional structure of each target protein based on the amino acid sequence and then use these structures to predict which drug molecules it will interact with.
Researchers from MIT and Tufts University have now designed an alternative computational method based on a large-scale language model powered by artificial intelligence algorithms. These models, such as the well-known ChatGPT, can analyze large amounts of text and identify which amino acids are most likely to occur together. This new model, called ConPLex, can match target proteins with potential drug molecules without the need for computationally intensive calculations of molecular structures.
Using this approach, researchers can screen over 100 million compounds in a single day, far surpassing any existing models.
Bonnie Berger, one of the senior authors of the study and a professor of mathematics at MIT, said, "This work addresses the need for efficient and accurate in vitro screening of potential drug candidates, and the scalability of the model enables large-scale screening to evaluate off-target effects, drug repurposing, and determine the impact of mutations on drug binding."
The research was published this week in the Proceedings of the National Academy of Sciences, with Lenore Cowen, a professor of computer science at Tufts University, also being a senior author. Robert Sledzieski is a research scientist at CSAIL, Samuel Sledzieski is a graduate student at MIT, and Bryan Bryson, an associate professor of biological engineering at MIT and a member of the Ragon Institute of MGH, MIT, and Harvard, is also one of the authors. In addition to this paper, the researchers have made their model available for other scientists to use online.
Predicting protein structures based on amino acid sequences has made significant progress in recent years. However, predicting the interaction between a large potential drug library and cancer proteins using these models is challenging, mainly because calculating the three-dimensional structure of proteins requires a significant amount of time and computational power.
Furthermore, these types of models have not effectively eliminated compounds known as decoys, which are very similar to successful drugs but do not actually interact well with the target.
"One long-standing challenge in the field is that these methods are brittle, in the sense that if I give the model something that looks almost like the real thing but is slightly different in some subtle way, the model might still predict that they interact," said Sledzieski.
Researchers have designed models to overcome this brittleness, but they are often tailored to specific classes of drug molecules and are not well-suited for large-scale screening due to the long computation time.
The research team at MIT decided to take a different approach based on their protein model developed in 2019. The language model works with a database of over 20,000 proteins, encoding this information into meaningful numerical representations for each amino acid sequence, capturing the relationship between sequence and structure.
"With these language models, even proteins that have very different sequences but potentially have similar structures or similar functions can be represented in a similar way in this language space, and we can leverage that to make predictions," said Sledzieski.
In their new study, researchers applied the protein model to determine which protein sequences would interact with specific drug molecules, both having numerical expressions that are transformed into a shared space by a neural network. They trained the network based on known protein-drug interactions, teaching it to associate specific features of proteins with their binding abilities to drugs without the need to calculate the three-dimensional structure of any molecule.
"By using this high-quality numerical representation, the model can completely bypass the need for atomic representations and predict whether this drug will bind," said Sledzieski. "The benefit of doing this is that it avoids the need for atomic representations, but the numbers still contain all the information you need."
Another advantage of this approach is that it takes into account the flexibility of protein structures, which can "twist" and present slightly different shapes when interacting with drug molecules.
To prevent their model from being easily deceived by decoy molecules, researchers also employed the concept of contrastive learning during the training phase. In this approach, researchers provided the model with examples of "real" drugs and decoys and taught it to distinguish between them.
Then, researchers screened a library of approximately 4,700 candidate drug molecules to detect their binding abilities to 51 enzymes called protein kinases.
From the top-ranked drugs, researchers selected 19 pairs of drug-protein interactions for experimental testing. The experiments showed that out of these 19 targets, 12 had strong binding affinities (in the nanomolar range), while almost all other possible drug-protein pairs had no affinity. Four of these pairs exhibited extremely high subnanomolar affinities, so strong that even tiny drug concentrations, approximately one billionth, would inhibit the protein.
Although the focus of this study was primarily on screening small molecule drugs, researchers are working to apply this method to other types of drugs, such as therapeutic antibodies. This modeling approach can also be used for toxicity screening of potential drug compounds to ensure they do not produce any adverse reactions before animal model testing.
"One reason drug development is so expensive is the high failure rate. If we can reduce the failure rate by informing early on that a drug is unlikely to be successful, it would greatly reduce the cost of drug development," said Sledzieski.
Eytan Ruppin, director of the Cancer Data Science Laboratory at the National Cancer Institute, who was not involved in the study, said that this new approach "represents a major breakthrough in drug target interaction prediction and opens up more opportunities for future research to enhance its capabilities," such as incorporating structural information into the potential space or exploring molecular generation methods for generating fake-like molecules, which can further improve prediction capabilities.
Source: https://news.mit.edu/2023/new-model-offers-speedy-drug-discovery-0608
Alphabet Q1 paid subscribers reached 350M, fueled by YouTube & Google One. YouTube ad revenue slightly missed at $9.88B amid Premium growth. Total Q1 revenue hit $109.9B, beating forecasts.
Anthropic releases Claude connectors for Adobe, Blender, and Ableton, enabling AI workflow automation and app control. It also sponsors the Blender Development Foundation.
Apple introduces an App Store model splitting 12-month plans into monthly installments. On iOS 26.4+, it lowers upfront costs without reducing total price, adding billing tracking.
Google launches Aadhaar digital IDs in Wallet for India, enabling secure QR/NFC verification. A phased partner rollout targets key sectors, while passport-based credentials expand globally.
YouTube tests "Ask YouTube," an AI tool blending text and video for complex queries. Initially for US Premium users, it supports interactive follow-ups, expanding Google's AI search.
Google grants US DoD classified AI access, joining OpenAI and xAI in defense deals amid Anthropic's dispute. Contractual safeguards against surveillance and lethal weapons remain unenforceable.
Mistral AI launches Workflows, a Temporal-powered orchestration platform for secure, Python-driven enterprise AI automation featuring human-in-the-loop approvals.