💻 computer science

PPI2Text: Captioning Protein-Protein Interactions with Coordinate-Aligned Pair-Map Decoding

The paper introduces PPI2Text, a multimodal large language model that generates free-form, interpretable descriptions of protein-protein interactions from amino acid sequences by leveraging ESM3 encoders, a novel coordinate-aligned PaCo-RoPE positional encoding for pair maps, and a Qwen3 decoder, supported by a newly released 351k-pair dataset that demonstrates superior performance in both linguistic quality and biological factuality.

Original authors: Xiao Fei, Sarah Almeida Carneiro, Yang Zhang, Lawrence P. Petalidis, Achilleas Tsortos, Costas Bouyioukos, Michalis Vazirgiannis

Published 2026-05-12

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Xiao Fei, Sarah Almeida Carneiro, Yang Zhang, Lawrence P. Petalidis, Achilleas Tsortos, Costas Bouyioukos, Michalis Vazirgiannis

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the world of biology as a massive, bustling city. In this city, proteins are the workers, machines, and vehicles that keep everything running. Sometimes, two workers need to team up to get a job done. This teamwork is called a Protein-Protein Interaction (PPI).

For a long time, scientists trying to understand these teams had to use a very rigid, old-fashioned filing system. They would look at two proteins and simply check a box: "Yes, they work together" or "No, they don't." Sometimes they added a number to say how strong the teamwork was. But this was like describing a complex dance routine by just saying "they danced." It missed all the details: How did they hold hands? Why did they dance? What was the music?

The paper introduces a new tool called PPI2Text that changes the game. Instead of just checking a box, PPI2Text writes a free-flowing story explaining exactly how two proteins interact.

Here is how it works, broken down into simple parts:

1. The Data: Gathering the Clues

Before writing a story, you need facts. The researchers didn't just look at one source; they gathered clues from ten different biological databases (like a massive library of scientific notes).

The Problem: Some clues were very strong (like a high-quality photo of the two proteins holding hands), while others were weak (like a blurry rumor that they might be in the same room).
The Solution: They created a "Quality Filter." They threw away the blurry rumors and kept the high-quality evidence. They then used a smart AI assistant to write a draft story based only on the strong evidence, making sure the story didn't invent facts that weren't there. This resulted in a massive library of 351,000 protein interaction stories.

2. The Brain: How PPI2Text "Sees" the Interaction

Most AI models look at one protein at a time, like reading a biography of a single person. But PPI2Text is special because it looks at two proteins at once and visualizes their relationship as a 2D grid map.

The Map Analogy: Imagine Protein A is a row of houses on a street, and Protein B is another row of houses on a parallel street. The interaction happens where the houses face each other.
The Pair-Map: PPI2Text draws a giant grid connecting every house on Street A to every house on Street B. This map shows exactly which parts of the proteins are touching or influencing each other.
The "Coordinate-Aligned" Trick: Usually, when you turn a 2D map into a list of words for a computer, you lose the sense of where things are. PPI2Text invented a special "address system" (called PaCo-RoPE) that keeps the map's geometry intact. It ensures the AI knows that "House 5 on Street A" is still connected to "House 5 on Street B," even when the information is turned into text.

3. The Writer: Turning Maps into Stories

Once the AI has the map and the facts, it uses a powerful language engine (based on a model called Qwen3) to write the story.

It doesn't just say "Protein A and Protein B interact."
It says: "Protein A and Protein B form a stable team at the cell's command center. Protein A acts as a switch, while Protein B is the engine that powers the reaction. They lock together specifically at the top of Protein A to start the process."

4. Did It Work?

The researchers tested PPI2Text against other methods.

The Test: They asked the AI to describe interactions it had never seen before.
The Result: PPI2Text wrote stories that were not only grammatically correct but also factually accurate. When experts (and other AI judges) checked the stories against the original raw scientific evidence, PPI2Text got the details right much more often than the older methods. It successfully captured the "who, what, where, and how" of the interactions.

Summary

Think of previous methods as a checklist (Yes/No). PPI2Text is a biographer. It takes the raw, messy data of how proteins touch and turn it into a clear, readable narrative that explains the mechanics of life, one interaction at a time. This helps scientists understand complex biological systems without getting lost in rigid codes or numbers.

Technical Summary: PPI2Text

Problem Statement

Protein-protein interaction (PPI) modeling has traditionally been framed as a binary classification or multi-label regression task. While emerging multimodal large language models (LLMs) can generate free-text descriptions for single proteins, they lack the capability to describe the complex, conditional, and mechanistic nature of interactions between protein pairs. Existing computational approaches often rely on rigid schemas, reducing biological context to discrete labels or fixed relational structures. This limits the ability to capture how or why proteins interact, particularly regarding higher-order dependencies, non-linear relationships, and condition-specific mechanisms that emerge only through complex chains of biological events. Furthermore, biological PPI data is fragmented across heterogeneous sources (e.g., IntAct, UniProt, STRING), leading to inconsistent structures and a lack of unified, interpretable narratives.

Methodology

1. Dataset Construction: Evidence-Tiered Synthesis

To address the scarcity of free-text PPI annotations, the authors constructed PPI2Text-Dataset, a corpus of 351,000 protein pairs with free-form descriptions.

Source Aggregation: Raw evidence was aggregated from ten curated biological databases (IntAct, PubMed, UniProt, 3did, Pfam, STRING, SIGNOR, Reactome, CORUM, ComplexPortal).
Evidence-Tiered Filtering: A unified evidence score $E(r)$ was computed for each pair, integrating interaction-level signals (direct experimental support) and contextual signals (biological coherence). Pairs were clustered into three tiers (T1–T3) using K-means. T1 (weakly supported) was discarded; T3 (strongly curated) was retained; T2 was subsampled using homology-aware strategies (MMSeq2) to ensure diversity without redundancy.
Controlled Synthesis: A labeler LLM (Gemini) was used to synthesize descriptions. Crucially, a tier-controlled generation framework was employed to constrain the output based on evidence strength. Four orthogonal constraints were enforced in the prompts:
- Descriptive Granularity: Length constraints based on confidence.
- Epistemic Strength: Verb selection (assertive vs. hedged) based on evidence.
- Mechanistic Attribution: Restriction of mechanistic claims when structural/curated evidence is absent.
- Silence Policy: Preventing functional descriptions for under-annotated entities.

2. Model Architecture: PPI2Text

PPI2Text is a multimodal LLM designed to generate free-text PPI captions from amino acid sequences.

Single-Protein Encoding: The model uses a frozen ESM3 encoder to generate per-residue embeddings for each protein ( $P_A, P_B$ ). These embeddings are compressed via a strided 1D convolution (factor of 4) to reduce sequence length while preserving local information.
Pair-Map Construction: To capture residue-level interactions, the model constructs a 2D Pair Map.
- Bidirectional cross-attention blocks exchange information between the two protein representations.
- For every pair of compressed residues $(i, j)$ , a feature vector is formed by concatenating the embeddings and their Hadamard product: $M_{ij} = \phi_{pair}([H^A_i \| H^B_j \| H^A_i \odot H^B_j])$ .
- This 3D tensor is adaptively pooled to a fixed $32 \times 32$ grid (1024 tokens) and projected into the decoder's hidden space.
Coordinate-Aligned Decoding (PaCo-RoPE): The core innovation is PaCo-RoPE (Pair-Coordinated Rotary Positional Encoding). Standard 1D positional encodings would flatten the 2D pair map, decoupling it from the original residue indices. PaCo-RoPE extends the standard RoPE to a 3D coordinate system $(p_T, p_\theta, p_\phi)$ $(p_{T}, p_{θ}, p_{ϕ})$ :
- Text tokens use $(n, n, n)$ .
- Protein A tokens use $(n, \lfloor i/4 \rfloor, 0)$ .
- Protein B tokens use $(n, 0, \lfloor j/4 \rfloor)$ .
- Pair-map tokens use coordinates derived from their grid position $(m, n)$ , aligned with the residue axes of both proteins.
- This design allows the decoder to attend to the 2D interaction structure while maintaining alignment with the specific residue positions of both partners, all without additional parameters.
Training: The model is fine-tuned using Qwen3 as the decoder (frozen base, LoRA adapters) and ESM3 as the encoder (frozen). Training uses cross-entropy loss on response tokens with a teacher-forcing protocol.

Key Contributions

Free-Text PPI Modeling Framework: A novel task formulation shifting PPI modeling from structured labels to human-readable, free-form descriptions, enabling richer semantic interpretation.
Large-Scale Dataset: The release of PPI2Text-Dataset (351k pairs) with high-quality, evidence-tiered descriptions synthesized from ten major biological databases.
Interaction-Aware Architecture: The PPI2Text model, which utilizes a 2D pair-map representation and PaCo-RoPE to align multimodal components (single-protein sequences and interaction maps) within a language decoder.
Empirical Validation: Demonstration of superior performance over baselines in both linguistic metrics and factuality metrics evaluated against raw biological evidence.

Results

The model was evaluated on two test sets: a temporal holdout (PPIs annotated after May 2025) and a C3-hard split (strict protein-level decontamination ensuring no homologous proteins in training/test).

Linguistic Metrics: PPI2Text outperformed baselines (including Seq+Qwen3, MINT+Qwen3, and ablation variants) on BLEU, ROUGE, and BERTScore.
Factuality Metrics: Using an LLM-as-a-judge (Claude-Opus-4.7) to evaluate against raw evidence, PPI2Text achieved the highest scores in Entity Grounding, Interaction Topology, and Mechanism Fidelity.
Ablation Studies:
- Removing the protein encoder (Seq+Qwen3) resulted in poor performance, highlighting the need for evolutionary/structural signals.
- Removing either single-protein representations or pair-map tokens degraded performance, indicating both are necessary.
- Replacing PaCo-RoPE with standard 1D RoPE significantly reduced performance, confirming the necessity of coordinate-aligned encoding for 2D interaction maps.
- Removing bidirectional cross-attention (No-Cross) lowered performance, showing its role in constructing the pair map.
Generalization: PPI2Text maintained robust performance on the C3-hard split, outperforming baselines even when generalizing to entirely unseen protein families.

Significance

The paper claims to present the first framework for generating free-form, human-readable descriptions of protein-protein interactions. By moving beyond rigid annotation schemas, the approach models interaction semantics in natural language, offering richer and more interpretable representations. The work establishes a foundation for future research in knowledge extraction and reasoning, demonstrating that interaction-aware encoding can capture fine-grained residue-level relationships. The authors emphasize that the model's ability to preserve biological entities and mechanisms, validated against raw evidence, suggests it yields more reliable and biologically meaningful predictions than previous classification-based methods. The release of the dataset and model code aims to encourage the development of more expressive and interpretable models in the field.