GDV-Scout: Gene, Disease, and Variant NER

GDV-Scout is a 108M-parameter biomedical named-entity recognition model for extracting genes or gene products, diseases, and sequence variants from English biomedical text. It is a BertForTokenClassification checkpoint using BIO labels and a BioLinkBERT-base encoder.

The model follows BioRED's relatively strict entity-boundary convention. It is intended for precise spans in sentence-level biomedical text. If broad BC2GM-style gene mentions (including more gene families, complexes, and generic mentions) are required, a model trained specifically for that annotation convention may be more appropriate.

Model details

Property	Value
Architecture	BERT token classifier (BertForTokenClassification)
Encoder lineage	michiyasunaga/BioLinkBERT-base
Parameters	107,647,495
Task	Biomedical named-entity recognition
Language	English
Input unit	One sentence
Training/evaluation length	Up to 192 subword tokens
Label scheme	BIO
Entity types	Disease, GeneOrGeneProduct, Variant

The checkpoint was produced in two stages:

1. BioLinkBERT-base was fine-tuned as an NER student on reconciled majority-vote silver labels from a six-run LLM ensemble.
2. That trained student, including its token-classification head, was continually fine-tuned on the human-annotated BioRED training set.

BioRED Gene annotations are mapped to GeneOrGeneProduct. BioRED Chemical, Species, and CellLine annotations are treated as O because they are outside this model's target schema.

Labels

ID	Label
0	O
1	B-Disease
2	I-Disease
3	B-GeneOrGeneProduct
4	I-GeneOrGeneProduct
5	B-Variant
6	I-Variant

Intended use

Use this model to identify candidate gene/gene-product, disease, and variant mentions in English biomedical sentences, such as PubMed titles and abstracts. Typical downstream uses include literature triage, biomedical search, and candidate generation for entity normalization or relation extraction.

This model performs mention detection only. It does not normalize mentions to database identifiers, determine relations between entities, or establish that a biomedical claim is true. It should not be used by itself for diagnosis, treatment decisions, or other clinical decision-making.

Usage

from transformers import pipeline

checkpoint = "nobal/gdv-scout"
extract_entities = pipeline(
    task="token-classification",
    model=checkpoint,
    tokenizer=checkpoint,
    aggregation_strategy="simple",
)

text = "The EGFR L858R variant predicts response in non-small cell lung cancer."
entities = extract_entities(text)

for entity in entities:
    print(entity)

For documents, split the text into sentences first and keep each input below 192 subword tokens. Longer inputs are outside the training setup; tokenize them with truncation=True, max_length=192 before calling the model directly.

Evaluation

BioRED test set

The saved checkpoint was re-evaluated locally on the converted BioRED test split on 2026-09-02. Scores use seqeval entity-level evaluation: entity type and the complete BIO boundary must match. The test split contains 1,108 sentences and 2,054 target entities.

Entity type	Precision	Recall	F1	Support
Disease	0.8049	0.8442	0.8241	860
GeneOrGeneProduct	0.8712	0.8876	0.8793	1,014
Variant	0.6840	0.8056	0.7398	180
Micro average	0.8249	0.8622	0.8431	2,054
Macro average	0.7867	0.8458	0.8144	2,054

The associated project report summarizes the original run as approximately 0.85 micro-F1. The table above gives the metrics reproduced directly from this packaged checkpoint and is therefore the recommended result to cite for these files.

External benchmarks

The following scores were computed from saved predictions using case-insensitive, document-level normalized mention matching. “Relaxed” permits token-set containment in either direction to accommodate different tokenization and boundary conventions. These numbers are not directly comparable to the exact BioRED scores above.

Benchmark	Target type	Precision	Recall	Relaxed F1
BC2GM test	GeneOrGeneProduct	0.916	0.729	0.812
NCBI-disease test	Disease	0.909	0.771	0.834
BC5CDR-disease test	Disease	0.873	0.896	0.884

Training data

The continual fine-tuning stage uses the standard BioRED train/validation/test partitions, converted from token/BIO records to sentence-level character spans. Only BioRED's Gene, Disease, and Variant labels are retained as target entities.

Split	Sentences	Disease	GeneOrGeneProduct	Variant
Train	4,374	3,388	3,848	660
Validation	1,134	905	948	191
Test	1,108	860	1,014	180

See the BioRED paper for dataset construction and annotation details.

Training procedure

Entity spans were aligned to BioLinkBERT subword offsets. Special tokens use an ignore index of -100; other tokens are optimized with standard token-classification loss. The best checkpoint was selected by validation entity-level F1.

Hyperparameter	Value
Epochs	3
Learning rate	2e-5
Train batch size	16
Evaluation batch size	32
Weight decay	0.01
Warmup ratio	0.1
LR schedule	Linear
Random seed	13
Maximum sequence length	192

Limitations

• The model is trained for English biomedical prose and may perform poorly on clinical notes, patents, non-English text, or general-domain text.
• It follows BioRED's strict span convention. Different corpora may define gene families, complexes, generic mentions, and entity boundaries differently.
• Chemical, species, and cell-line entities are deliberately not extracted.
• Variant performance is less reliable than gene and disease performance, and the BioRED variant test subset is comparatively small.
• Long text should be sentence-segmented. Naive truncation can silently omit entities.
• Predictions can contain false positives and missed entities. Human review and entity normalization are recommended for high-stakes or production applications.
• No explicit cross-corpus PMID overlap audit between the earlier silver-label corpus and BioRED was recorded; benchmark results should be interpreted with that caveat.

Citation

If you use the training dataset, cite BioRED:

@article{luo2022biored,
  title = {BioRED: a rich biomedical relation extraction dataset},
  author = {Luo, Ling and Lai, Po-Ting and Wei, Chih-Hsuan and Arighi, Cecilia N. and Lu, Zhiyong},
  journal = {Briefings in Bioinformatics},
  volume = {23},
  number = {5},
  pages = {bbac282},
  year = {2022},
  doi = {10.1093/bib/bbac282}
}

The encoder is based on BioLinkBERT:

@inproceedings{yasunaga2022linkbert,
  title = {LinkBERT: Pretraining Language Models with Document Links},
  author = {Yasunaga, Michihiro and Leskovec, Jure and Liang, Percy},
  booktitle = {Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics},
  year = {2022}
}