Twin Peaks: Dual-Head Architecture for Structure-Free Prediction of Protein-Protein Binding Affinity and Mutation Effects
Fuente:
arXiv
Salvato in:
| Autori principali: | , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2025
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866909810528616448 |
|---|---|
| author | Dey, Supantha Chowdhury, Ratul |
| author_facet | Dey, Supantha Chowdhury, Ratul |
| contents | We present a novel dual-head deep learning architecture for protein-protein interaction modeling that enables simultaneous prediction of binding affinity ($ΔG$) and mutation-induced affinity changes ($ΔΔG$) using only protein sequence information. Our approach offers a significant advancement over existing methods by employing specialized prediction heads that operate on a shared representation network, allowing direct and optimized prediction of both values. To ensure robust generalization, we integrated complementary datasets from SKEMPI v2 and PDBbind with a rigorous protein domain-based splitting strategy that prevents information leakage between training and validation sets. Our architecture combines transformer-based encoders with a novel cross-attention mechanism that processes paired protein sequences directly, without requiring any structural information. The network embeds input sequences using ESM3 representations, then employs a learnable sliced window embedding layer to manage variable-length sequences efficiently. A multi-layer transformer encoder with bidirectional self-attention captures intra-protein patterns, while cross-attention layers enable explicit modeling of interactions between protein pairs. This shared representation network feeds into separate $ΔG$ and $ΔΔG$ prediction heads, allowing task-specific optimization while leveraging common features. The model achieves $ΔΔG$ validation of Pearson correlation at 0.485, while maintaining strong $ΔG$ predictions (Pearson: 0.638). While existing approaches require protein structure data and binding interface information, our model eliminates these constraints. This provides a critical advantage for the numerous proteins with unknown structures or those challenging to crystallize, such as viral and intrinsically disordered proteins. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2509_22950 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Twin Peaks: Dual-Head Architecture for Structure-Free Prediction of Protein-Protein Binding Affinity and Mutation Effects Dey, Supantha Chowdhury, Ratul Quantitative Methods We present a novel dual-head deep learning architecture for protein-protein interaction modeling that enables simultaneous prediction of binding affinity ($ΔG$) and mutation-induced affinity changes ($ΔΔG$) using only protein sequence information. Our approach offers a significant advancement over existing methods by employing specialized prediction heads that operate on a shared representation network, allowing direct and optimized prediction of both values. To ensure robust generalization, we integrated complementary datasets from SKEMPI v2 and PDBbind with a rigorous protein domain-based splitting strategy that prevents information leakage between training and validation sets. Our architecture combines transformer-based encoders with a novel cross-attention mechanism that processes paired protein sequences directly, without requiring any structural information. The network embeds input sequences using ESM3 representations, then employs a learnable sliced window embedding layer to manage variable-length sequences efficiently. A multi-layer transformer encoder with bidirectional self-attention captures intra-protein patterns, while cross-attention layers enable explicit modeling of interactions between protein pairs. This shared representation network feeds into separate $ΔG$ and $ΔΔG$ prediction heads, allowing task-specific optimization while leveraging common features. The model achieves $ΔΔG$ validation of Pearson correlation at 0.485, while maintaining strong $ΔG$ predictions (Pearson: 0.638). While existing approaches require protein structure data and binding interface information, our model eliminates these constraints. This provides a critical advantage for the numerous proteins with unknown structures or those challenging to crystallize, such as viral and intrinsically disordered proteins. |
| title | Twin Peaks: Dual-Head Architecture for Structure-Free Prediction of Protein-Protein Binding Affinity and Mutation Effects |
| topic | Quantitative Methods |
| url | https://arxiv.org/abs/2509.22950 |