Biomedical engineers have created an AI-based system that crafts short proteins known as peptides. These peptides have the ability to latch onto and eliminate previously untreatable proteins that cause diseases. Drawing inspiration from OpenAI’s image generation technology, this innovative algorithm quickly selects which peptides to test experimentally.
Biomedical engineers at Duke University have created an AI-based system that designs short proteins, known as peptides, which can attach to and eliminate disease-causing proteins that were once deemed untreatable. Influenced by OpenAI’s image generation model, their groundbreaking algorithm efficiently selects which peptides should undergo experimental testing.
The research was published on January 22 in the journal Science Advances.
One strategy in treating diseases is to develop drugs that can specifically target and destroy the proteins responsible for the illness. Some of these critical proteins have clear, defined shapes, like a neatly folded origami figure, allowing traditional small molecule drugs to easily connect with them. However, over 80% of disease-causing proteins are more chaotic and disorganized, resembling a tangled ball of yarn. This disorder makes it very challenging for standard therapies to find a connection point to bind and perform their function.
To tackle this challenge, researchers have examined the potential of using peptides to bind to and break down these problematic proteins. Due to their smaller size, peptides do not require specific binding pockets; instead, they can connect to various amino acid sequences within the protein. However, existing binding solutions have not been effective for unstable or overly tangled protein structures. The search for new binding proteins continues, but these efforts still depend on detailed 3D structural information of the targeted proteins, which is often unavailable for disordered proteins.
Instead of attempting to construct these 3D diagrams, Pranam Chatterjee, an assistant professor of biomedical engineering at Duke, and his team drew inspiration from generative large language models (LLMs) to formulate a new approach. This led to the development of PepPrCLIP, or Peptide Prioritization via CLIP. The first part of their tool, PepPr, employs a generative algorithm trained on a substantial library of natural protein sequences to create new ‘guide’ proteins with targeted features. The second part, CLIP, uses an algorithm framework initially designed by OpenAI to align images with their corresponding captions, adapting it to assess which peptides match their intended proteins. In this case, the CLIP model simply requires the target sequence.
“OpenAI’s CLIP algorithm connects text with images. For instance, if you input ‘dog,’ you should see an image of a dog,” explained Chatterjee. “In our adaptation, we trained it to connect peptides with proteins instead. PepPr generates the peptides, and our modified CLIP algorithm evaluates these peptides to determine the best matches.”
When compared to RFDiffusion, a current platform for peptide generation based on the 3D structure of a target protein, PepPrCLIP demonstrated superior speed and was more successful in creating peptides that closely matched their intended proteins. To test how effectively PepPrCLIP operates with both ordered and disordered proteins, Chatterjee and his lab collaborated with other research teams from Duke University Medical School, Cornell University, and Sanford Burnham Prebys Medical Discovery Institute.
In their initial experiment, they confirmed that peptides developed by PepPrCLIP could successfully bind to and inhibit the activity of UltraID, a straightforward and stable enzyme protein. Subsequently, they employed PepPrCLIP to design peptides that could attach to beta-catenin, a complex and disordered protein involved in signaling for various cancer types. They generated six peptides that CLIP indicated should bind to this protein and discovered that four were effective in bonding with and breaking down their target. By dismantling the protein, they could slow down cancer cell signaling.
In their most challenging experiment, the team designed peptides aimed at a highly disordered protein linked with synovial sarcoma, a rare and aggressive cancer that primarily affects soft tissues in children and young adults. Chatterjee described it as “like a bowl of spaghetti. It’s possibly the most disordered protein known.”
The researchers examined ten designs by inserting their peptides into synovial sarcoma cells and found that the PepPrCLIP-generated peptides likewise managed to bind to and degrade the protein, similar to their findings with simpler targets. If they can eliminate this protein, they could pave the way for a therapy that targets a previously undruggable form of cancer.
Looking ahead, Chatterjee and his team aim to further refine their platform and work with medical and industry professionals to develop peptides that might eventually be utilized in new treatments for diseases caused by unstable proteins, such as Alexander’s Disease, a fatal neurological disorder that mainly affects children, along with various cancers.
“These complex, disordered proteins have rendered numerous cancers and diseases nearly impossible to treat because we lacked the ability to create molecules that could target them,” elaborated Chatterjee. The success of PepPrCLIP in addressing even the most intricate proteins opens up a world of exciting clinical opportunities.
