💻 bioinformatics

Bridging the gap between omics and structural data: A framework for interpreting protein-RNA interaction specificity

This study presents a framework that integrates omics data with experimental protein-RNA structures to identify binding motif cores and benchmark AlphaFold3, revealing both its predictive promise and its current limitations, such as signs of memorization and challenges with alternative binding modes.

Original authors: Fauconnet, Y., Deng, S., Koundri, R., Pagani, M., Quignot, C., Antonio, D., Andreani, J.

Published 2026-09-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Fauconnet, Y., Deng, S., Koundri, R., Pagani, M., Quignot, C., Antonio, D., Andreani, J.

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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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 living cell, a constant, silent conversation takes place between two types of molecules: proteins and RNA. Proteins are the heavy lifters, the machines that build structures, speed up chemical reactions, and carry out the instructions of life. RNA acts as the messenger and the regulator, carrying genetic codes and deciding when those codes should be read. For life to function, these two must find each other and lock together with perfect precision. If a protein grabs the wrong piece of RNA, the cell's instructions can go haywire, leading to diseases like cancer or neurodegeneration. Scientists have long tried to map these meetings, taking detailed photographs of the molecules as they touch. However, capturing these moments is incredibly difficult. While there are hundreds of thousands of known protein structures, only a few thousand show proteins actually holding onto RNA. This scarcity leaves a massive gap in our understanding of how these interactions work and makes it hard to predict how they will behave in new situations.

To bridge this gap, a team of researchers at the Institute for Integrative Biology of the Cell in France set out to combine two different ways of looking at the problem. On one side, they had the rare, high-resolution 3D snapshots of proteins and RNA locked together. On the other, they had vast amounts of data from modern experiments that show which RNA sequences proteins prefer to bind with, even if the exact 3D shape isn't known. The researchers built a new method to check if these two sources of information agreed with each other. They created a scoring system that compares the specific RNA sequence seen in a 3D structure against the preferred sequences found in large-scale experiments. When the two matched well, the team identified a small, central region of the RNA sequence, which they called the "motif core," as the most critical part of the handshake. They found that these cores were not just random matches; they were the parts of the RNA that made the most physical contact with the protein and were surrounded by the most evolutionarily stable parts of the protein, suggesting these are indeed the true anchors of the interaction.

With this refined understanding of what a correct binding site looks like, the team turned their attention to a powerful new artificial intelligence tool called AlphaFold3. This software has recently revolutionized the field by predicting the 3D shapes of proteins and their complexes with high accuracy. The researchers wanted to see if AlphaFold3 could do the same for protein-RNA interactions and, more importantly, if it could predict the specific RNA sequence a protein would choose. They tested the system using a dataset of human proteins where they knew both the 3D structure and the preferred RNA sequences. They asked the AI to predict the structure using the exact RNA sequence from the experiment, a non-specific sequence that the protein shouldn't care about, and sequences containing the preferred binding motifs.

The results revealed a surprising limitation. When the AI was fed the exact RNA sequence it had likely seen during its training, it produced excellent predictions that matched the experimental structures. However, when the researchers gave it a different, non-specific RNA sequence, the AI still managed to fold the protein correctly and often placed the RNA in a similar position, even though the sequence was wrong. This suggests that the system is not truly learning the rules of how a specific protein recognizes a specific RNA sequence. Instead, it appears to be memorizing the shapes it has seen before. The AI was so good at recalling the training data that it could not easily distinguish between a specific, correct interaction and a generic, incorrect one. In cases where a protein could bind to RNA in two different ways, the AI tended to ignore the alternative and stick to the most common version it had memorized, failing to adapt to the specific input provided.

The study also highlighted the challenges of using these tools for proteins with multiple binding sites. In one specific case involving a protein called LIN28A, which has two different domains that can grab RNA, the AI consistently predicted the interaction at only one of the sites, regardless of the RNA sequence provided. It struggled to switch between the two possible modes of binding. The researchers concluded that while deep learning offers a promising path forward, it currently relies heavily on recognizing patterns from its training data rather than understanding the fundamental rules of molecular recognition. They emphasized that for these tools to become truly reliable for designing new therapies or understanding complex diseases, they need to be guided by more diverse data and perhaps constrained by the specific binding preferences that the researchers' new scoring method helps to identify. The team has made their findings and a web tool available to the scientific community, offering a way to visualize these interactions and test new predictions against the reality of experimental data.

Problem Statement
Protein-RNA interactions (PRIs) are fundamental to cellular processes such as gene regulation and viral infection, yet their structural characterization lags significantly behind that of proteins alone. While the Protein Data Bank (PDB) contains over 250,000 entries, only a few thousand represent protein-RNA complexes. This scarcity of high-quality experimental structures limits the training and performance of deep-learning-based structure prediction tools, particularly for RNA and protein-RNA interfaces. Current methods, including AlphaFold3 (AF3), exhibit lower success rates for protein-RNA interactions (~~40%) compared to protein-protein interactions (~~70%), partly due to the flexibility of RNA and the diverse nature of binding specificities, which range from strict sequence dependence to non-specific structural recognition.

Methodology
The authors developed a framework to bridge the gap between structural data and high-throughput omics data to better interpret PRI specificity and benchmark prediction tools.

Dataset Construction: The study compiled a dataset of 294 human protein-RNA interfaces from the PDB that also possessed independent RNA-binding preference data from in vitro assays (RNAcompete and RNA Bind-n-Seq) and in vivo CLIP data. Proteins with only CLIP data were excluded.
Agreement Scoring Workflow: A novel scoring system ( $S$ $S$ ) was designed to quantify the concordance between the RNA sequence observed in a 3D structural interface and the binding preferences derived from omics data.
- Frequency Matrices: Binding preferences were processed into position-specific mono- and dinucleotide frequency matrices from aligned 5-mers.
- Scoring Mechanism: A sliding window approach aligned the structural RNA sequence against the frequency matrix. The score $S$ (where $S=0$ is perfect agreement and negative values indicate poorer agreement) was calculated based on the likelihood of observing specific dinucleotides at given positions, incorporating pseudocounts to handle sparse data.
- Motif Core Identification: The workflow identified a "motif core" (four consecutive nucleotides) within the structural interface that yielded the maximum agreement score.
Structural and Evolutionary Validation: The authors validated the biological relevance of the identified motif cores by analyzing:
- Atomic Pair (AP) Contacts: The density of atomic contacts within 5 Å of the protein.
- Evolutionary Conservation (EC): The conservation of amino acids contacting the motif core compared to the rest of the interface.
AlphaFold3 Benchmarking: The authors used the filtered dataset to test AF3's ability to predict specific interfaces. They generated AF3 models for 280 unique protein-RNA pairs using four different RNA input conditions:
- PDB-RNA: The exact RNA sequence from the experimental structure.
- PolyU: A non-specific poly-U sequence of the same length.
- RNAC/RBNS: Sequences embedding the consensus binding motif from in vitro experiments, flanked by random nucleotides.
- Performance was evaluated using backbone RMSD, fraction of native contacts ( $f_{nat}$ ), and CAPRI metrics (High, Medium, Acceptable, Incorrect).

Key Results

Agreement Scoring Performance: The majority of the 294 complexes showed near-zero agreement scores for in vitro data (RNAC/RBNS), while CLIP data showed lower consistency. A relevancy threshold ( $T = \mu - \sigma = -2.25$ ) was established, filtering the dataset to 159 (RNAC) and 169 (RBNS) high-confidence interfaces.
Validation of Motif Cores: The identified motif cores were structurally and evolutionarily significant. In >90% of cases, motif cores exhibited higher atomic pair contact densities than the rest of the interface. Similarly, amino acids contacting the motif cores were more evolutionarily conserved than those contacting non-core interface regions.
AlphaFold3 Specificity and Memorization:
- AF3 achieved high structural accuracy for protein backbones (median RMSD ~1.79 Å) regardless of the RNA input.
- However, RNA backbone accuracy was lower (median RMSD ~5.1 Å).
- Crucially, models generated with the exact PDB-RNA input achieved significantly higher $f_{nat}$ and CAPRI "High" quality scores (21%) compared to models using PolyU, RNAC, or RBNS inputs (3–6% High quality).
- This disparity suggests that AF3's ability to predict precise interface contacts relies heavily on the exact RNA sequence seen during training, indicating signs of memorization rather than true sequence-level generalization.
Case Studies:
- SNRPA: AF3 robustly reproduced the dominant binding mode (1AUD) across all inputs but failed to recover an alternative binding mode (6XH3) when the PDB-RNA input was used, converging instead to the dominant mode.
- LIN28A: The protein possesses two distinct binding domains (ZKD and CSD). AF3 consistently predicted the CSD-mediated interaction regardless of input but struggled to reproduce the ZKD-mediated interaction, highlighting difficulties with proteins having multiple binding sites.

Significance and Claims
The paper claims to provide a quantitative framework for integrating structural and omics data to identify biologically relevant interaction motifs ("motif cores") and to expose current limitations in deep learning-based protein-RNA structure prediction.

Resource: The authors released a web resource (https://bioi2.i2bc.paris-saclay.fr/rnaprotci/) allowing interactive exploration of the dataset, agreement scores, and structural features.
Limitations of Current AI: The study evidences that while AF3 produces high-quality backbone structures, it struggles with RNA sequence specificity. The superior performance on exact training sequences suggests the model may be memorizing specific complexes rather than generalizing the rules of RNA recognition.
Future Directions: The authors suggest that future improvements in protein-RNA prediction may require incorporating omics-derived motifs as constraints (e.g., using tools like Protenix) or building enriched multiple sequence alignments (MSAs) that include binding preference information. They emphasize the need for more diverse structural and omics data to move beyond memorization toward genuine specificity prediction.

The authors remain modest, noting that their dataset is largely redundant with AF3's training data and that the current benchmark is restricted to human proteins with specific data availability, limiting broad generalization claims.

More like this