HERCULES: an integrative deep-learning framework for predicting RNA-binding propensity and mutation effects at single-residue resolution

HERCULES is a unified, sequence-based deep-learning framework that integrates protein language models with physicochemical descriptors to simultaneously predict RNA-binding domains, assess global RNA-binding propensity, and evaluate the impact of single-residue mutations on RNA-binding ability with high accuracy and interpretability.

The Gist

Imagine your cell is a bustling, high-tech factory. Inside, there are millions of workers (proteins) and blueprints (RNA). For the factory to run smoothly, specific workers need to grab the blueprints, read them, and decide what to build. These workers are called RNA-Binding Proteins (RBPs).

Sometimes, a worker grabs the wrong blueprint, or a tiny typo in their uniform (a mutation) stops them from holding the blueprint at all. This can cause the factory to break down, leading to diseases like cancer or Alzheimer's.

Scientists have been trying to build a "detective tool" to find out:

1. Which workers are the blueprint grabbers?
2. Exactly where on their uniform do they hold the blueprint?
3. If we change a single thread in their uniform (a mutation), will they still be able to hold the blueprint?

Enter HERCULES.

What is HERCULES?

Think of HERCULES as a super-smart, two-brained detective that looks at a protein's "uniform" (its amino acid sequence) and instantly answers all three questions above. It doesn't need to see the protein in a 3D microscope; it just needs the text of the uniform's pattern.

The name stands for a fancy title, but you can think of it as a Hybrid Detective that uses two different ways of thinking to solve the case:

Brain 1: The "Big Picture" Reader (The Language Model)

Imagine you are reading a novel. Even if you don't know every single word perfectly, you can guess the plot, the setting, and the character's role just by looking at the whole story.

• How it works: HERCULES uses a "Protein Language Model" (trained on millions of protein stories). It reads the entire protein sequence at once.
• The Analogy: It's like a literary critic who knows that if a character wears a cape and carries a sword, they are likely a hero. Similarly, this brain looks at the whole protein and says, "Ah, this whole section looks like it belongs to a blueprint-grabbing team." It finds the general neighborhoods where the action happens.

Brain 2: The "Microscope" Specialist (The Physicochemical Module)

Now, imagine zooming in so close you can see the texture of the fabric. Is it sticky? Is it charged like static electricity? Is it slippery?

• How it works: This brain looks at the tiny chemical properties of individual amino acids (the threads). It knows that RNA is sticky and negative, so it looks for threads that are positive and sticky.
• The Analogy: This is like a forensic scientist checking the fabric for specific stains. It can tell you, "If you change this specific thread from red to blue, the fabric will no longer stick to the blueprint." This is crucial for spotting mutations.

How They Work Together

Most old detective tools were either good at the "Big Picture" (knowing a protein is a binder) but bad at finding the exact spot, OR they were good at the "Microscope" (finding the spot) but couldn't handle mutations well.

HERCULES combines them:

1. The Big Picture Brain says: "The action is happening in this general zone."
2. The Microscope Brain says: "And specifically, it's right here, and if you tweak this one thread, the whole thing falls apart."

They merge their notes to give you a single, perfect map.

Why is this a Big Deal?

The paper shows that HERCULES is the best detective we've ever had for three reasons:

1. It finds the hidden spots: It can spot "non-canonical" binders. Think of it as finding a secret agent who doesn't wear the standard uniform but still does the job. Old tools missed these; HERCULES catches them.
2. It predicts the damage of typos: If a protein has a mutation (a typo in the genetic code), HERCULES can predict if that typo will break the protein's ability to hold RNA. It got this right 87% of the time in tests.
3. It sees the "Ghost" interactions: The researchers tested HERCULES on proteins where they only knew one specific blueprint it could hold. But HERCULES, using its chemical knowledge, realized, "Hey, this protein could actually hold other types of blueprints too!" It proved that HERCULES understands the nature of the protein, not just the specific examples it was trained on.

The Bottom Line

HERCULES is like giving scientists a GPS and a crystal ball in one package.

• GPS: It tells you exactly where on the protein the RNA binding happens.
• Crystal Ball: It tells you what will happen if you change the protein's code.

This helps doctors and researchers understand diseases better and potentially design new drugs that can fix broken "blueprint grabbers" or create new tools to control how our cells read their instructions. And the best part? It's free for anyone to use!

Technical Summary

1. Problem Statement

RNA-binding proteins (RBPs) are critical regulators of RNA metabolism, and their dysfunction is linked to various diseases. While computational methods exist to identify RBPs, significant challenges remain:

• Localization: Accurately pinpointing RNA-binding domains (RBDs) at single-residue resolution is difficult, especially for intrinsically disordered regions and non-canonical RBDs.
• Mutation Effects: Quantifying how specific sequence variations (e.g., single amino acid substitutions) impact RNA-binding ability is an open problem.
• Limitations of Current Tools: Existing methods often rely on structural data (which is scarce and biased) or use global classification with post-hoc occlusion strategies (like HydRA) that lack sensitivity to subtle physicochemical changes. Purely structure-based predictors struggle with flexible regions and non-canonical interfaces.

2. Methodology: The HERCULES Framework

HERCULES (Hybrid framEwoRk for RNA-binding domain loCalization and mUtation anaLysis using physicochemical and languagE modelS) is a unified, sequence-based framework that integrates two complementary components to predict RNA-binding propensity and mutation effects.

A. Component 1: Fine-Tuned Protein Language Model (PLM)

• Base Model: Uses ProteinBERT, a pre-trained protein language model.
• Training: Fine-tuned on a curated dataset of human RBPs vs. non-RBPs (binary classification).
• Residue-Level Output: Instead of relying on occlusion-based perturbation, HERCULES leverages the attention heads of the fine-tuned model. By averaging attention values across layers and heads, it generates a global, context-aware residue-level profile that captures long-range dependencies and domain architecture.

B. Component 2: Physicochemical Descriptor Model

• Data Source: Curated dataset of experimentally validated RNA-binding-disrupting mutations from UniProt.
• Feature Engineering: Computes 82 residue-level physicochemical descriptors (e.g., hydrophobicity, charge, disorder, structural propensity) derived from established sets (cleverSuite, catGRANULE 2.0).
• Feature Selection: Uses Elastic Net regularized logistic regression to select the most informative descriptors (retaining 30 features).
• Mutation Scoring: Employs Linear Discriminant Analysis (Fisher's method) to create an optimal linear combination of features. This model predicts the direction and magnitude of the effect of a mutation (Wild-Type vs. Mutant difference vectors), specifically targeting deleterious changes.

C. Integration Strategy

• Profile Fusion: The attention-derived profile (global context) and the physicochemical profile (local sensitivity) are combined via a weighted linear sum:
  $$\text{Final Profile} = L \cdot \text{Attention} + \alpha \cdot \text{Smooth}(\text{Physicochemical})$$
  • $L$: Protein length (to scale attention values).
  • $\alpha$: Optimized weighting factor (0.2).
  • $\text{Smooth}$: Sliding window smoothing.
• Global Score: A Multilayer Perceptron (MLP) integrates the global PLM score and the mean physicochemical score to produce a final global RNA-binding propensity score.

3. Key Contributions

1. Unified Framework: HERCULES is the first sequence-only model to simultaneously perform global RBP classification, residue-level RBD localization, and mutation effect prediction.
2. Mechanistic Interpretability: By explicitly modeling physicochemical descriptors, the framework offers insights into why a mutation is deleterious (e.g., changes in charge or disorder), rather than treating the protein as a black box.
3. Handling Non-Canonical Regions: The model excels at identifying RNA-binding regions in intrinsically disordered proteins (IDPs), where structure-based methods typically fail.
4. Robustness to Data Bias: The framework demonstrates superior generalization when tested on augmented datasets involving AlphaFold3-predicted complexes, suggesting it learns intrinsic binding propensities rather than overfitting to specific experimental structures.

4. Key Results

Global Classification

• On an independent test set, HERCULES discriminates RBPs from non-RBPs with an AUROC of 0.86 and AUPRC of 0.86.
• While HydRA showed slightly higher AUROC in strict binary classification, HERCULES demonstrated higher sensitivity in identifying annotated RBPs across diverse organisms, favoring the detection of conditionally active regions.

Residue-Level Localization

• Benchmarking: Evaluated on 562 human proteins with Pfam-annotated RBDs (<40% identity to training set).
• Performance: HERCULES significantly outperformed state-of-the-art tools (HydRA and DRNApred) in AUPRC ratio to random and Positive Likelihood Ratio (PLR).
• Domain Types: It achieved high accuracy across canonical (e.g., RRMs), non-canonical, and putative RBDs. Notably, it is the only method that clearly resolved multiple RRM domains in proteins like TDP-43 and FUS.
• Disordered Regions: Performance improved with increasing intrinsic disorder content, whereas competitors (DRNApred) degraded.

Mutation Effect Prediction

• Using a curated dataset of 393 experimentally validated mutations, HERCULES correctly classified 87–88% of deleterious variants.
• It successfully predicted the spatial localization of functional disruption, showing a sharp drop in propensity scores specifically at mutation sites (e.g., in TDP-43).

Structural Validation (PDB & AlphaFold3)

• Tested on 176 experimentally resolved protein-RNA complexes.
• Augmentation Test: When contact annotations were augmented with AlphaFold3-predicted complexes involving G-quadruplex RNAs, HERCULES was the only method to improve its AUROC. Competitors (trained on PDB data) saw performance degradation, indicating HERCULES generalizes better to diverse RNA interaction contexts.

5. Significance and Impact

• Rational Design: The ability to quantify mutation effects and localize binding sites enables the rational design of RNA aptamers and engineered proteins.
• Sequence-First Approach: As sequence data outpaces structural data, HERCULES provides a scalable, high-accuracy method for annotating the "RNA-binding proteome" without requiring 3D structures.
• Interpretability: The explicit use of physicochemical descriptors bridges the gap between deep learning predictions and biochemical intuition, offering a mechanistic understanding of RNA-protein interactions.
• Accessibility: The tool is available as an open-source Python package and a user-friendly web server, facilitating broad adoption in the research community.

In summary, HERCULES represents a significant advancement in computational biology by unifying global context (via PLMs) with local biochemical sensitivity (via physicochemical descriptors), offering a robust, interpretable, and generalizable solution for studying RNA-protein interactions.