LCR Annotation and Validation Pipeline

This pipeline is designed to fetch, parse, and validate research papers concerning Low Complexity Regions (LCRs) to automate LCR annotation.

---

Requirements

Python Libraries

The scripts require the following Python packages:

• PyMuPDF (fitz): Used for high-fidelity text extraction from PDF files.[cite: 1]
• httpx: Used for asynchronous HTTP requests to the NCBI and PMC APIs.[cite: 3]
• lxml: Required for parsing XML data from PubMed Central.[cite: 3]
• requests: Used for synchronous API calls to UniProt and Europe PMC.[cite: 1, 5]
• asyncio: Manages asynchronous fetching processes.[cite: 3]

System Tools

• SEG: A local command-line tool for identifying low-complexity segments in protein sequences.[cite: 5]
• CAST: A local command-line tool for detecting compositionally biased regions in sequences.[cite: 5]

---

Script Descriptions

pmc_regex_miner.py

Fetches LCR-related papers from the PubMed Central API using search terms like "tandem repeat" and "intrinsically disordered," filtering results with a comprehensive regex list.[cite: 3]

training_data_pdf_parsing.py

Extracts structured text from local PDF files using a state-machine approach to identify section headers and clean out journal artifacts (e.g., "Article in Press") but has to be updated, doesnt really work now :(.[cite: 1]

lcr_classifier.py

Performs binary classification on gathered papers to identify which ones explicitly mention "LCR" or "low-complexity" in the text versus those that do not.[cite: 4]

lcr_location_finder.py

Uses multiple regex patterns to extract coordinate ranges from paper text, including standard mentions like "residues 331-369" and protein-specific notations.[cite: 2]

fetch_uniprot_check_seq.py

Identifies UniProt Accession Numbers in the paper text, downloads the corresponding FASTA sequence, and cross-references author-declared LCRs with predictions from SEG and CAST.[cite: 5]

---

LABBOOK

26/03/2026

We created a way to fetch relevant papers via PubMed Central API, and later sort them by regex as mentioning LCR, or not mentioning LCR[cite: 6]

We research models and decided to use BioClinicalModernBERT, as well as later add BioLINKBERT[cite: 6]

I downloaded around 50 papers from the publikacje excel and started parsing through them with training_data_pdf_parsing.py, but decided to skipt to automatizing fetch_uniprot_check_seq.py[cite: 6]

---

27/03/2026

I started working on the fetch_uniprot_check_seq.py[cite: 6]

I came back to fixing the pdf positive articles parser[cite: 6]

We decided that we need to broaden our positive set, because it is biased with Rna-binding papers. Also, we need to create a negative set of papers without a function annotation[cite: 6]