GenPTM: A Generalizable Framework for Protein Post-Translational Modification Information Extraction from the Scientific Literature

Technical Summary: GenPTM – A Generalizable Framework for Protein Post-Translational Modification Information Extraction

Problem Statement
Protein post-translational modifications (PTMs) are critical regulatory mechanisms in cellular processes, yet existing PTM databases suffer from limited coverage and infrequent updates due to their heavy reliance on manual curation. While the scientific literature contains vast amounts of PTM knowledge, current automated information extraction (IE) tools are largely restricted to specific, well-studied PTM types (e.g., phosphorylation, ubiquitination). Developing separate systems for the hundreds of known PTM types is not feasible, particularly for less-frequently reported modifications where annotated training data is scarce. There is a distinct need for a general-purpose, PTM-agnostic framework capable of extracting modified proteins and their specific amino acid sites directly from biomedical literature without requiring PTM-specific model retraining.

Methodology
The authors propose GenPTM, a unified IE framework that addresses lexical variation across different PTM types through a generalized text representation strategy. The methodology consists of three primary components:

1. Generalized Instance Creation:
   The core innovation is the replacement of PTM-specific lexical items with generic placeholders to create a PTM-agnostic input format. The system identifies two primary textual patterns for PTM events:

   • Explicit mention of the modification name (e.g., "acetylation").
   • Mention of the attached chemical group (e.g., "acetyl" or "ubiquitin").
     These specific terms are normalized to generic placeholders: "MODIFICATION" (or "GROUP" for chemical moieties) and "GROUP". Furthermore, candidate protein entities are replaced with the f-term "PROTEIN" and site mentions with "SITE". All candidate entities are enclosed in special boundary markers ([E]...[/E]) to guide the model's attention. This strategy forces the model to learn the relational structure between modification cues and entities rather than memorizing specific protein names or PTM terminologies.

2. Transformer-Based Classification:
   The framework utilizes BiomedBERT, a transformer model pre-trained on PubMed abstracts. The task is formulated as a binary classification problem where the model determines if a marked entity (protein or site) is truly modified within a sentence. The model extracts the [CLS] token embedding, passes it through fully connected layers, and applies a softmax activation to predict "modified" or "not modified." The model is fine-tuned on a corpus derived from five major PTM types (Ubiquitination, Sumoylation, Neddylation, Acetylation, and Phosphorylation).

3. Inference and Post-Processing:
   During inference, the pipeline processes unseen abstracts by generating instances for the target PTM. A post-processing module is essential to reconstruct the final output:

   • Coreference Resolution: Resolves pronouns or referential phrases (e.g., "these sites") to their antecedent proteins or sites, potentially looking across sentence boundaries.
   • Site-to-Protein Association: Links predicted sites to their substrate proteins using grammar-based rules derived from dependency parsing and part-of-speech tags, ensuring the output is a complete <protein, site> pair.

Key Contributions

• A Generalizable Framework: The authors propose a novel pipeline capable of detecting any specific PTM type using a single, unified model, overcoming the limitation of PTM-specific tools.
• PTM-Agnostic Representation: The introduction of a placeholder-based representation strategy allows the model to generalize across diverse PTM types by abstracting away specific lexical variations.
• Comprehensive Corpus Construction: A large, manually annotated corpus was created covering 13 distinct PTM types, including well-studied, frequent, and low-frequency modifications.
• Extensive Evaluation: The framework was evaluated on 13 PTM types, demonstrating effectiveness not only on the five types used for training but also on eight unseen types, including rare modifications like Citrullination and AMPylation.

Results
GenPTM was evaluated at the abstract level across three criteria: <protein, site> pair extraction, protein-only extraction, and site-only extraction.

• Overall Performance: The system achieved micro-averaged F1-scores ranging from 92% to 96% across all evaluated PTM categories.
• Generalization: The model maintained high performance on Other Frequent PTMs (OF-PTMs), achieving F1-scores of approximately 91–98% for protein-only and site-only extractions, and 96.77% for <protein, site> pairs.
• Low-Frequency PTMs (LF-PTMs): Performance on less-reported PTMs (e.g., Citrullination, AMPylation) showed a slight decrease, with F1-scores around 76% for pairs and ~89% for protein/site-only extractions. The authors attribute this decline primarily to limitations in the Named Entity Recognition (NER) component and the scarcity of complex sentence structures in the limited available literature, rather than the core prediction model.
• Comparison: The authors note that GenPTM outperforms previous tools (e.g., MPTM, EAGL, TEES, RLIMS-P) which typically report F1-scores in the 50–80% range or are restricted to specific PTMs. GenPTM achieved a 91.30% F1-score for <protein, site> pairs across multiple PTMs, compared to the ~53% peak reported in the BioNLP Shared Task 2011.

Significance and Claims
The paper claims that GenPTM provides a viable solution for PTM-agnostic information extraction, enabling the automated discovery of PTM knowledge for over 100 distinct modification types. By validating the hypothesis that replacing specific modification terms with generic placeholders allows a model to learn shared textual patterns, the framework offers a scalable alternative to building separate models for each PTM.

The authors emphasize that while the current framework excludes complex cases like Glycosylation (due to diverse structural terms) and Methylation (due to ambiguity with DNA methylation), the approach establishes a new direction for scalable, automated PTM knowledge discovery. The tool is designed to support database curation efforts, facilitate the creation of new PTM-specific resources, and provide contextual evidence for researchers, thereby helping to bridge the gap between the rapidly growing scientific literature and static biological databases.