A multi-layer computational framework for structural prioritization of transcriptome-derived proteins

In the vast library of life, every cell carries a set of instructions written in a code called RNA. These instructions tell the cell how to build proteins, the molecular machines that perform nearly every task required for an organism to survive and function. For decades, scientists have been able to read these instructions with incredible speed, generating massive lists of protein sequences from plants, animals, and microbes. However, knowing the sequence of letters in a protein is only the beginning. To understand what a protein actually does, researchers need to see its three-dimensional shape. A protein's form determines its function, much like how the shape of a key determines which lock it can open. The challenge is that while computers can now read millions of genetic instructions, predicting the complex, folded shapes of the resulting proteins remains a difficult bottleneck. Without these shapes, it is nearly impossible to know which proteins might interact with specific targets in the body, such as the signaling molecules that drive inflammation.

This is where a new study steps in, offering a systematic way to navigate this mountain of data. Researchers from the Federal University of ABC in Brazil developed a multi-layered computational framework designed to sift through thousands of candidate proteins and identify the most promising ones for further study. They focused on the Cereus jamacaru, a cactus native to the Brazilian Caatinga biome, known for its potential anti-inflammatory properties. While previous research on this plant had mostly looked at small chemical compounds, this team wanted to explore the proteins hidden within its genetic code. Their goal was not to prove that these plant proteins directly cure diseases, but to create a reliable, step-by-step method for finding structurally sound protein candidates that could interact with human inflammatory signals. By combining several different computer-based techniques, they successfully narrowed down a list of over 129,000 translated proteins to just nine high-confidence candidates, demonstrating a scalable strategy that could be applied to any organism's genetic data.

The journey began with a massive dataset containing 129,052 protein sequences derived from the cactus's root and epidermis. To make sense of this overwhelming number, the researchers first applied a broad filter based on known structures. They compared the cactus proteins against a database of experimentally determined protein shapes, keeping only those that shared more than half of their sequence with a known structure. This initial step reduced the pool to about 14,800 candidates. They then used a scoring system to evaluate the quality of the predicted shapes, discarding those that looked unstable or unreliable. This process whittled the list down to approximately 3,500 high-quality structural models, creating a manageable set of candidates for deeper analysis.

Next, the team turned their attention to three specific human proteins known as cytokines: tumor necrosis factor-alpha, interleukin-1 beta, and interferon-alpha. These molecules are central regulators of the body's inflammatory response and are often involved in conditions like rheumatoid arthritis. The researchers wanted to see if any of the cactus proteins could physically interact with these human cytokines. To do this, they employed a two-step docking strategy. First, they used a fast, rigid-body screening method to test how well the 3,500 cactus proteins fit against the three human cytokines. This step identified a smaller group of proteins that showed promising initial matches. To refine these results, they used a second, more flexible docking method that allowed for slight movements in the protein shapes, mimicking the way real molecules might adjust when they come together. By looking for proteins that performed well in both methods, the researchers identified nine candidates that consistently showed strong structural compatibility with the human cytokines.

Before declaring these nine proteins as the final winners, the team subjected them to a series of rigorous validation tests to ensure their findings were not just computer artifacts. They first refined the shapes of these nine proteins using a highly advanced prediction tool called AlphaFold 3, which generated new, high-confidence models to confirm the initial structures were accurate. They then analyzed the specific points where the cactus proteins touched the human cytokines. They counted the number of chemical bonds, such as hydrogen bonds and salt bridges, that formed at these interfaces. A high number of these connections suggested a stable and tight fit. Furthermore, they checked whether the parts of the cactus proteins involved in these interactions were evolutionarily conserved, meaning they had remained unchanged across millions of years of evolution. This conservation often indicates that a specific region of a protein is critical for its function or stability, adding another layer of confidence to the results.

To test if these interactions would hold up under real-world conditions, the researchers ran dynamic simulations. They placed the four most promising protein pairs into a virtual environment filled with water molecules and simulated their behavior over time. These simulations ran for 100 nanoseconds each, repeated three times for every pair to ensure the results were reproducible. The goal was to see if the proteins would stay locked together or if they would drift apart as they moved and jiggled. The results showed that the interactions remained stable throughout the simulations. The proteins maintained their shape and their connection points, even as they moved in the virtual water. This dynamic stability confirmed that the predicted interactions were not just static snapshots but represented robust, physically plausible complexes.

The study also took a closer look at what these nine prioritized proteins actually are. Through functional analysis, the team discovered that they belong to diverse families of enzymes and regulatory proteins, including oxidoreductases and RNA-binding proteins. These are typical components of plant metabolism, involved in processes like energy production and stress response. The researchers emphasized that the selection of these proteins was driven entirely by their structural compatibility with human cytokines, not by any assumption that they naturally function as human drugs. In fact, the study explicitly argues against the idea that these plant proteins are direct functional equivalents of human cytokine regulators. Instead, the findings suggest that the three-dimensional shapes of these plant proteins happen to create surfaces that can physically fit with human inflammatory signals, likely due to shared principles of geometry and chemistry rather than shared biological history.

The ultimate value of this work lies in the framework itself. The researchers did not claim to have discovered a new medicine, but rather a reliable method for finding candidates. By integrating multiple layers of evidence—structural modeling, dual-docking approaches, evolutionary analysis, and dynamic simulations—they created a workflow that reduces uncertainty at every step. This approach allows scientists to move from massive, unmanageable datasets to a small, curated list of high-confidence targets that are ready for experimental testing. The study concludes that this strategy is transferable, meaning it can be applied to any transcriptome or large protein dataset to identify structurally robust candidates. It offers a practical path forward for bioprospecting, turning the overwhelming flood of genetic data into a focused set of hypotheses that can guide future biochemical research and protein engineering efforts.