- The paper introduces an innovative PPI modeling approach that generates free-form captions using coordinate-aligned pair-map decoding.
- A multimodal encoder–decoder architecture, combining frozen ESM3 with the novel PaCo-RoPE, captures residue-level interactions with high mechanistic fidelity.
- Experimental results show superior performance with improved BLEU and mechanism fidelity metrics, validating the model's robust generalization on stringent tests.
PPI2Text: Free-Form Captioning of Protein–Protein Interactions with Residue-Level Pair-Map Decoding
Motivation and Context
Traditional computational approaches to protein–protein interaction (PPI) modeling have predominantly constrained interaction prediction to classification or regression tasks, producing discrete outcomes or limited annotated structures. These approaches lack the expressivity required to encapsulate the mechanistic, contextual, and high-order dependencies that govern biological interactions in vivo. The motivation for this work is twofold: first, to address the rigidity and interpretability deficit resulting from structured representations, and second, to leverage the advances in multimodal LLMs for generating human-readable, evidence-grounded free-text descriptions of PPIs.
This paradigm shift introduces new opportunities for hypothesis generation, improved integration with literature-derived knowledge bases, and enhanced interpretability for downstream bioinformatics and systems biology tasks. However, free-text modeling of PPIs from sequences remains unexplored due to data quality, model architecture, and alignment constraints.
Methodology
Evidence-Tiered Dataset Construction
A major contribution of this work is the design and release of the PPI2Text-Dataset, an evidence-tiered corpus comprising 351k high-confidence PPI pairs, each paired with a free-form natural language description. The dataset is constructed by systematic aggregation from ten expert-curated biological repositories including IntAct, UniProt, PubMed, STRING, Reactome, 3did, SIGNOR, CORUM, ComplexPortal, and Pfam. Each candidate PPI is scored along orthogonal evidence axes measuring interaction mapping, mechanistic detail, literature redundancy, curated source validation, and contextual coherence. K-means clustering on evidence scores stratifies the data into three quality tiers, with only T2–T3 PPIs retained after homology-aware filtering (using MMSeq2) to maximize diversity and factual reliability while minimizing hallucinations. Textual captions are synthesized under strict evidence-controlled prompting via Gemini, constrained by granularity, epistemic strength, mechanistic attribution, and silence policy, yielding descriptions that dynamically encode confidence and context.
PPI2Text Model Architecture
The proposed PPI2Text model is a multimodal encoder–decoder system. Each protein in a PPI is embedded using a frozen ESM3 encoder, which incorporates per-residue sequence, structural, and biophysical features. The two protein representations are linked through bidirectional cross-attention to enhance partner awareness and to model feature exchange before pairwise interaction mapping. Critically, a residue-level PairMap is constructed: for every residue pair, a high-dimensional, non-scalar feature is synthesized through a concatenation and Hadamard product of embedding vectors, projecting to a 2D grid that captures interaction topology with maximal contextual and mechanistic fidelity.
This pair-map grid is further compressed via adaptive average pooling to a fixed-size sequence of 1024 pseudo-tokens, which are then injected—together with single-protein summary representations—into the prompt of a frozen Qwen3 LLM decoder.
Coordinate-Aligned Rotary Positional Encoding (PaCo-RoPE)
Integrating 2D pairwise interaction information into an autoregressive LLM (which natively expects 1D input) is nontrivial. To preserve spatial semantics, the authors introduce PaCo-RoPE—a novel positional encoding scheme that aligns protein residue axes with corresponding grid coordinates in the pair-map and mediates token placement in the context of both proteins and text. This 3D-interleaved RoPE structure encodes the vertical (protein A), horizontal (protein B), and sequence positional axes, ensuring that the model does not collapse pair-map information into a sequence but rather attends to it along aligned structural axes. This is critical for mechanistic fidelity and fundamentally improves generalization to OOD protein pairs.
Fine-tuning is performed with LoRA adaptation at the decoder, while core encoders and the LM backbone remain frozen to maintain pretraining priors and computational efficiency.
Experimental Design and Results
Experiments use both a realistic temporal holdout (annotated post-May 2025) and a stringent C3-hard OOD split (no homologous proteins in train/test), with pair-level and protein-level decontamination, respectively. Baselines include models with no protein encoder, state-of-the-art sequence-based encoders (MINT), single-modality ablations, removal of cross-attention, and ablation of the pair-map and PaCo-RoPE.
Performance metrics include standard lexical and semantic metrics (BLEU, ROUGE, BERTScore/BioBERTScore) computed against synthesized references, but more importantly, direct factual evaluation against curated raw evidence using an LLM-as-a-judge setup measuring entity grounding, interaction topology, and mechanistic fidelity.
PPI2Text achieves the highest scores across all metrics and outperform all baselines and ablations. In the C3-hard split, where no train/test overlap in protein identity is present and generalization is strictly compositional, PPI2Text maintains strong performance—an improvement of ~2 points in BLEU-4 and over 1 point in mechanism fidelity metric compared to the strongest baseline. Notably, flattening the pair map or using standard 1D positional encoding substantially harms mechanism fidelity and entity grounding, directly validating the necessity of the coordinate-aligned pair-map and cross-modality attention. The full model demonstrates robust generalization, high biological faithfulness, and convergence between automated and expert-curated evaluation.
Theoretical and Practical Implications
This work fundamentally extends the expressivity of PPI modeling by producing free-form, evidence-aligned descriptions that encapsulate molecular mechanisms, regulatory contexts, and biological contingency, thereby bridging the gap between rigid computational predictions and the complex narratives found in biomedical literature. This aligns with ascendant multimodal LLMs in the biomedical domain but uniquely adapts them for high-granularity pairwise interaction decoding.
The compositional and interpretable nature of the model enables flexible integration into downstream pipelines, including automatic literature mining, pathway curation, drug target prioritization, and text-based hypothesis generation. The architectural innovations around pairwise grid encoding and positional alignment can be generalized to other multienitity biological interaction problems, including transcript–protein or multi-substrate enzyme mechanisms.
Limitations and Future Directions
The model presupposes known-positive PPI input, delegating binary interaction detection to upstream tools. Current evaluation is limited to computational and curated evidence; wet-lab validation of novel predictions is a requisite step for assessing biological impact. Integration of structure-level contact information and extension to higher-order complexes or competitive/collaborative interaction networks is suggested for future work. The framework’s plug-and-play architectural design will facilitate the rapid adoption of improved residue encoders or future multimodal protein LLMs.
Conclusion
PPI2Text introduces and validates a sequence-to-caption paradigm for PPI modeling, leveraging a coordinate-aligned multimodal encoder–decoder architecture. By synthesizing high-confidence, evidence-aligned free-text PPI descriptions at scale, the approach provides researchers with a new modality for interpretability and integration, supporting a spectrum of bioinformatic discovery pipelines. The paired public dataset and codebase further establish this work as a methodological foundation for future advances in biolanguage modeling and generative biology research.
