Abstract
Here, we integrate fine-tuned protein language models and protein-protein interaction databases to develop a Structure-agnostic Language Transformer & Peptide Prioritization (SaLT&PepPr) module that efficiently selects peptides from interaction interfaces, without the need for structural information. We experimentally fuse SaLT&PepPr-derived “guide” peptides to E3 ubiquitin ligase domains and reliably identify candidates that induce robust intracellular degradation of clinically-relevant targets, and exhibit high binding affinities, low off-targeting rates, and functional transcriptional effects.
Introduction
Targeted Protein Degradation (TPD) has emerged as a transformative therapeutic modality, often likened to "CRISPR for proteins." By utilizing molecules such as ubiquibodies or PROTACs, TPD redirects the cell's internal proteasomal machinery to destroy specific disease-linked proteins. However, a significant bottleneck remains: identifying the high-affinity binders required to "flag" these targets. Traditional binder design relies heavily on high-resolution 3D structural data from X-ray crystallography or cryo-EM, or exhaustive experimental screenings. This "structure hurdle" renders many proteins—particularly those that are intrinsically disordered or lack available structural data—effectively "undruggable."
To address this, we present SaLT&PepPr (Structure-agnostic Language Transformer & Peptide Prioritization), a sequence-only AI framework that designs peptide "guides" for protein degradation without requiring 3D structural information.
The SaLT&PepPr Architecture
The SaLT&PepPr framework leverages the power of protein language models (PLMs) to identify binding interfaces directly from amino acid sequences. The model is built upon Meta AI’s ESM-2 (650-million parameters), with the final layers fine-tuned to perform binary classification of protein-protein interaction (PPI) sites.
Instead of designing binders de novo, SaLT&PepPr utilizes a "partner" strategy. It scans proteins known to interact with a target—identified through databases like IMEx and BioGRID—to find specific peptide fragments responsible for the interaction. The model identifies "hotspots" (local maxima in binding probability) and extracts short peptide sequences. These AI-designed peptides are then fused to a modular E3 ubiquitin ligase (CHIPΔTPR), creating a "ubiquibody" (uAb) that uses the peptide as a guide to tag the target for degradation.
Experimental Results and Validation
The model was trained on a custom dataset derived from the Protein Data Bank (PDB) and benchmarked against state-of-the-art structure-based models using the ScanNet PPBS dataset. SaLT&PepPr achieved an AUROC of 0.77 on non-homologous test sets, demonstrating strong generalization to novel proteins.
In head-to-head experimental trials, SaLT&PepPr-derived guides consistently outperformed those designed using structure-prediction methods like AlphaFold2-Multimer. A key success was the degradation of endogenous β-catenin, a critical driver in colon cancer. Several SaLT&PepPr candidates achieved >60% degradation in DLD1 cells. Furthermore, these designed peptides exhibited high binding affinities ($K_d \approx 7-9$ nM) and exceptional specificity; mass spectrometry confirmed no significant off-target effects across approximately 6,700 quantified proteins.
Discussion and Future Directions
The success of SaLT&PepPr highlights that sequence-only models can capture the complex "grammar" of protein interfaces effectively enough to bypass the need for explicit 3D coordinates. This approach is not only faster—running inference in approximately one minute on a standard CPU—but also more versatile, excelling where traditional docking simulations fail due to protein disorder.
Future work will focus on optimizing the orientation of the peptide-ligase fusion and evolving the model from binary classification to energy-based gradient predictions. This architecture holds potential beyond TPD, including the identification of small-molecule binding pockets and the design of novel vaccine antigens.
Summary updated