← Latest papers
💻 bioinformatics

Proteome-scale inference of transcription factor binding motifs from predicted protein-DNA structures

This paper introduces ZebraMotif, a proteome-scale computational framework that leverages AlphaFold3 and NA-MPNN to predict DNA-binding motifs for all 2,543 zebrafish transcription factors by modeling protein-DNA structures, thereby significantly expanding the available motif database beyond what is possible with experimental data alone.

Original authors: Iovino, B., Bulaong, N., Boahene, L., Law, W., Hammond, N., Potts, M., Alvarez, R., DeRisi, J. L., Pak, J., Kim, Y. J., Royer, L. A.

Published 2026-11-19
📖 6 min read🧠 Deep dive

Original authors: Iovino, B., Bulaong, N., Boahene, L., Law, W., Hammond, N., Potts, M., Alvarez, R., DeRisi, J. L., Pak, J., Kim, Y. J., Royer, L. A.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Inside every cell, a vast network of molecular switches determines what the cell is and what it does. These switches are controlled by proteins called transcription factors. Each transcription factor acts like a specific key, designed to fit into a unique lock made of DNA. When the key finds its lock, it turns on or off the genes nearby, telling the cell to grow, divide, or change its identity. To understand how life works, scientists need to know exactly which key fits which lock. For humans and a few other well-studied animals, researchers have spent decades running experiments to map these connections. But for most species, including the zebrafish, a tiny fish that is a favorite for studying how bodies develop, this map is almost entirely blank. Without knowing which proteins bind to which DNA sequences, scientists cannot fully understand the instructions that build an organism.

For a long time, the only way to fill in these blanks was to run slow, expensive lab tests for every single protein, a task that is impossible to complete for thousands of proteins at once. Scientists have tried to guess the missing pieces by looking at similar proteins in other animals, but this method often fails when the animals are too different from one another. Now, a team of researchers has found a new way to solve this puzzle without stepping into a wet lab. They used a powerful computer program to predict the 3D shape of how these proteins grab onto DNA, and then used that shape to figure out the exact DNA sequence each protein prefers. By applying this method to the entire set of proteins in a zebrafish, they have created the first comprehensive map of DNA binding for this species, revealing thousands of new connections that were previously unknown.

The researchers began by asking a simple question: if a computer can predict the shape of a protein holding onto a piece of DNA, can it tell the difference between the correct piece of DNA and a random one? They used a sophisticated artificial intelligence system, known as AlphaFold3, which is famous for predicting the shapes of proteins. They fed the computer the sequence of a zebrafish protein and a specific DNA sequence, asking it to model how they would look when stuck together. They did this for thousands of combinations, pairing each protein with both the DNA it is supposed to bind and random DNA sequences that it should ignore. The computer assigned a confidence score to each prediction, essentially telling the researchers how sure it was that the two molecules would stick together in that specific way.

The results showed that the computer was remarkably good at this task. For most of the proteins tested, the model gave a much higher confidence score when the protein was paired with its correct DNA target than when it was paired with a random sequence. This meant the computer could act as a filter, sorting the good matches from the bad ones. The researchers noticed that the proteins that were easiest to predict were also the ones the computer could distinguish most clearly. When the computer was very confident about the shape of the protein-DNA complex, it was almost always the correct match. This discovery gave them a reliable signal to use for a much larger search.

With this signal in hand, the team launched a massive screen to find the binding rules for every transcription factor in the zebrafish. They started with a library of potential DNA sequences, derived from regions of the genome that are known to be active in developing fish embryos. They then paired every single zebrafish transcription factor with these candidate sequences, running millions of computer simulations to see which combinations formed the most stable and confident structures. For each protein, they kept only the top matches—the ones where the computer was most certain the protein and DNA belonged together.

To turn these individual snapshots into a usable map, the researchers used another tool to analyze the shapes of the best matches. This tool looked at the specific atoms where the protein touched the DNA and calculated the pattern of preferences. It asked: does this protein like the letter A at the first position, or the letter G? By combining the results from many different high-confidence predictions for the same protein, they built a detailed profile, or a "motif," that describes exactly what DNA sequence that protein seeks out. They repeated this process for over 2,500 transcription factors, creating a database called ZebraMotif.

The new database covers nearly three times as many zebrafish proteins as any previous collection. Before this work, experimental data existed for only about one-third of the zebrafish proteins, and even those were often just guesses based on human data. The new map provides direct, structure-based predictions for the vast majority of the remaining proteins. When the researchers tested their predictions against the few proteins where the true answer was already known from experiments, the computer-generated maps matched the real-world data about half the time for the primary prediction, and even better when they looked at the top few possibilities. This level of accuracy is impressive considering the computer had never seen the experimental data for these specific fish proteins.

The researchers also checked if these new maps made biological sense. They looked at whether the proteins predicted to bind a certain DNA sequence were actually active in the same cells where that DNA was open and accessible. They found that the new maps showed a strong connection between the protein's presence and the accessibility of its target sites, similar to the connection seen in the older, experimentally derived maps. This suggests that the computer predictions are not just mathematical artifacts, but reflect real biological rules. Furthermore, when they used these new maps to reconstruct the gene regulatory networks of developing fish embryos, the new data filled in missing links, showing how different proteins work together to control development in ways that were invisible before.

This approach does not require any new lab experiments to get started. It only needs the genetic code of the organism and a list of potential DNA sequences to test. Because of this, the method can be applied to any species, even those where no experimental binding data exists. The researchers demonstrated this by creating a complete map for the zebrafish, a species that has been a model for development for decades but lacked this specific knowledge. While the method is not perfect and works best for proteins that fold into stable shapes, it opens a door to understanding the genetic logic of life across the entire tree of life, turning a task that once seemed impossible into a routine computational exercise.

Drowning in papers in your field?

Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.

Try Digest →