CAPRINI-M: An AI-curated Cardiac-Specific Atlas of Protein Interactions in Mice

CAPRINI-M is an AI-curated web-based atlas that integrates literature mining, large language models, and AlphaFold3 structural predictions to provide a comprehensive, mechanistically annotated resource of cardiac protein-protein interactions in mice, offering superior performance and detailed thermodynamic insights compared to existing databases.

The Gist

Imagine your heart is a bustling, high-tech city. In this city, the buildings are cells, and the people running around are proteins. For the city to function, these people need to shake hands, hug, and work together in teams. These "handshakes" are called Protein-Protein Interactions (PPIs).

If the wrong people shake hands, or if the right people refuse to talk, the city gets chaotic. This is what happens in heart diseases like heart failure.

For a long time, scientists trying to map out who shakes hands with whom in the heart had to read thousands of research papers manually. It was like trying to map a city by reading every single newspaper article ever written about it, one by one. It was slow, expensive, and often missed the small but crucial details (like where exactly on the hand the handshake happened).

Enter CAPRINI-M. Think of it as a super-smart, AI-powered cartographer that has just finished drawing the most detailed map of the heart's "social network" ever created.

Here is how they built this map, explained in simple terms:

1. The AI Librarian (Reading the Books)

First, the team asked a super-smart AI (a Large Language Model called LLaMA) to read 9,105 scientific papers about heart biology.

• The Analogy: Imagine a librarian who can read a million books in a day. Instead of just skimming the titles, this librarian reads every sentence to find mentions of two proteins "interacting."
• The Result: The AI found 11,189 specific handshakes (interactions) that scientists had discovered in mice. It filtered out the noise, ignoring papers about cancer or lungs, and focused strictly on the heart.

2. The 3D Architect (Building the Models)

Knowing that two proteins interact is good, but knowing how they fit together is better.

• The Analogy: Imagine you know two people are holding hands, but you don't know if they are holding hands loosely, gripping tightly, or if one is holding the other's elbow.
• The Tool: The team used AlphaFold3, a cutting-edge AI that predicts what 3D shapes proteins look like. It built a 3D model for every single one of those 11,000+ handshakes.
• The Insight: This allowed them to see the "fingerprint" of the interaction. They could see exactly which parts of the proteins touched. This is crucial because sometimes a tiny change in a protein (like a mutation) can ruin the handshake, even if the proteins are still there.

3. The Thermometer (Measuring the Grip)

Once they had the 3D models, they wanted to know: How strong is this handshake?

• The Analogy: Some handshakes are firm and last a long time (strong bonds). Others are a quick, polite tap (weak bonds).
• The Tool: They calculated the "energy" of the interaction. In science, a lower energy number means a stronger, more stable bond. They gave every interaction a score. If Protein A and Protein B have a very low energy score, they are likely best friends who stick together. If the score is high, they might just be acquaintances who drift apart easily.

4. The Detective (Checking the Work)

AI can sometimes "hallucinate" (make things up). So, the team built a "Detective" (a Neural Network) to double-check the AI's work.

• The Analogy: Imagine a senior detective reviewing the librarian's notes. The detective looks at the 3D models and the energy scores and asks, "Does this handshake actually make sense biologically?"
• The Result: This detective helped filter out the fake handshakes and gave a "probability score" to every interaction, telling researchers how confident they can be that the interaction is real.

Why is this a Big Deal?

The researchers tested CAPRINI-M against other famous protein databases (like BioGRID and STRING).

• The Test: They asked, "If we look at a heart disease, which database helps us find the right genes and pathways faster?"
• The Winner: CAPRINI-M won. Because it was built specifically for the heart and included 3D structural details, it found the "heart of the matter" much better than the general databases, which are like general phone books containing numbers for everyone in the world, not just the heart city.

The Bottom Line

CAPRINI-M is a free, online tool where scientists can:

1. Search for any heart protein.
2. See who it interacts with.
3. Look at a 3D model of the handshake.
4. See how strong the bond is.

In short: Before, studying heart proteins was like trying to navigate a dark city with a blurry map. With CAPRINI-M, scientists now have a high-definition, 3D, GPS-enabled map that shows exactly who is holding hands with whom, how tight the grip is, and where the trouble spots are. This helps them design better drugs and understand heart disease much faster.

Technical Summary

1. Problem Statement

Protein-protein interactions (PPIs) are fundamental to cardiovascular disease (CVD) biology, yet existing knowledge is fragmented across literature and heterogeneous databases. Current general-purpose PPI resources suffer from several critical limitations:

• Bias: They are often skewed toward well-studied domains like cancer, lacking cardiac-specific nuance.
• Lack of Structural Detail: Most resources do not provide information on specific interaction interfaces or thermodynamic stability, which is crucial for understanding alternative splicing effects and disease variants.
• Manual Curation Bottleneck: Keeping databases up-to-date via manual curation is slow and labor-intensive.
• LLM Reliability: While Large Language Models (LLMs) offer a path to automate literature mining, they risk hallucinations and false positives/negatives without rigorous benchmarking.

2. Methodology

The authors developed CAPRINI-M, an end-to-end AI-powered framework that transforms scattered cardiobiology literature into a structurally and thermodynamically annotated knowledge base. The workflow consists of six key stages:

A. Literature Mining and Relation Extraction (RE)

• Corpus: 9,105 full-text manuscripts from PubMed Central were retrieved and filtered to retain 7,548 cardiac-relevant papers.
• LLM Strategy: The open-source LLaMA-3.3 70B model was used for extraction.
• Benchmarking: Strategies were optimized using the RegulaTome corpus (6,435 annotated PPIs). The best-performing approach combined example-based prompting (Pos/Neg examples) with spaCy entity extraction and synonym expansion to handle protein name variability.
• Output: 11,189 PPIs were extracted from 4,503 papers.

B. Structural Modeling (AlphaFold3)

• Pipeline: A high-throughput pipeline processed all extracted mouse protein pairs using AlphaFold3.
• Selection: Five alternative 3D structures were generated per pair; the top-ranked structure (based on internal confidence scores) was retained.
• Feature Extraction: From the predicted complexes, the authors extracted interface descriptors (contact statistics, buried surface area) and confidence metrics (ipTM, pTM, PAE).

C. Thermodynamic Analysis

• Energy Calculation: Using AMBER force fields and MM/GBSA (via MMPBSA.py), the authors calculated approximate binding free energies ($\Delta G$) for each complex.
• Purpose: These values serve as proxies for binding strength and stability, allowing for the prioritization of competitive interactions.

D. Interaction Probability Prediction (Machine Learning)

• Model Training: A neural network (NN) classifier was trained on a leakage-minimized human PPI benchmark (DIGGER dataset) and applied to mouse PPIs, leveraging the cross-species generalizability of protein language models.
• Features: The model used a multimodal approach, concatenating ESM3 (protein language model) embeddings with top 12 AlphaFold3-derived features (identified via SHAP analysis, e.g., ipSAE, pDockQ, contact patches).
• Performance: The SHAP12_ESM3 model achieved the highest performance (F1 = 0.888, ROC-AUC = 0.961).

E. Web Application

• The database is hosted as a Shiny for Python web application, allowing users to search, filter, and visualize 3D structures (via 3Dmol.js) and thermodynamic data.

3. Key Contributions

1. First Cardiac-Specific Mouse PPI Atlas: A curated resource of 11,189 interactions among 4,255 unique proteins, specifically tailored to cardiovascular biology.
2. Interface-Resolved Annotation: Unlike standard PPI databases, CAPRINI-M provides residue-level interaction interfaces and thermodynamic stability estimates ($\Delta G$).
3. Validated AI Pipeline: A rigorous benchmarking framework demonstrating that optimized LLM prompting + AlphaFold3 + Multimodal NN classifiers yields high-precision interaction predictions.
4. Open Access: Full availability of the web tool, source code, and raw prediction data via GitHub and Zenodo.

4. Results

Performance Benchmarks

• LLM Extraction: The "PosExamples" strategy (without pre-supplied entities) achieved the best F1 score (0.717) for PPI extraction.
• PPI Prediction: The multimodal AF3+ESM3 model significantly outperformed sequence-only and structure-only baselines, achieving an accuracy of 0.890 and F1 of 0.888 on the human test set.
• Atlas Statistics:
  • Average interaction probability: 40.61%.
  • 32.40% of interactions have a >50% probability of forming a complex.
  • Average $\Delta G$: -20.74 kJ/mol.

Systems-Level Validation

• Pathway Enrichment: When used for network reconstruction in a murine heart failure model (TACOMA dataset), CAPRINI-M significantly outperformed general resources (BioGRID and STRING).
  • CAPRINI-M identified 6 significant pathways (BH-FDR < 0.1), whereas BioGRID and STRING identified 0.
  • Enrichment scores were significantly higher for CAPRINI-M (paired t-test $p < 0.01$).

Experimental Concordance

• The predicted $\Delta G$ rankings correlated with experimental literature in four case studies:
  • HIF/ARNT: Predicted tighter binding for HIF1A-ARNT vs. HIF2A-ARNT matched experimental ITC data.
  • Notch1: Predicted higher affinity for DLL4 over DLL1 matched experimental affinity measurements.
  • GJA1/Cx43: Predicted competitive preference of SRC over TJP1 aligned with biochemical competition data.
  • BAG3: Predicted thermodynamic favorability for HSPB8 over HSPA8 supported observed subcellular targeting behaviors.

5. Significance

CAPRINI-M represents a paradigm shift in how domain-specific interactomes are constructed. By integrating literature mining, structural biology (AlphaFold3), and machine learning, it provides:

• Mechanistic Insight: Researchers can now analyze specific binding interfaces and how mutations or alternative splicing might disrupt them, rather than just binary "interaction/no interaction" data.
• Prioritization: The $\Delta G$ and probability scores allow researchers to filter thousands of potential interactions down to the most thermodynamically stable and biologically plausible candidates for experimental validation.
• Translational Potential: The resource is designed to integrate with the LINDA framework (for splicing analysis), enabling future studies on cell-type-specific network rewiring in cardiac diseases.

Limitations & Future Work:
The authors acknowledge that coverage is limited to open-access literature, $\Delta G$ values are approximations, and the model relies on human-trained benchmarks. Future iterations (CAPRINI-H) will expand to human data, incorporate isoform-specific modeling, and refine kinetic estimates.