CarD-T: Carcinogen Detection via Transformers

Overview

CarD-T (Carcinogen Detection via Transformers) is a novel text analytics approach that combines transformer-based machine learning with probabilistic statistical analysis to efficiently nominate carcinogens from scientific texts. This model is designed to address the challenges faced by current systems in managing the burgeoning biomedical literature related to carcinogen identification and classification.

Model Details

Architecture: Based on Bio-ELECTRA, a 335 million parameter language model (sultan/BioM-ELECTRA-Large-SQuAD2)
Training Data: CarD-T-NER dataset containing 19,975 annotated examples from PubMed abstracts (2000-2024)
- Training set: 11,985 examples
- Test set: 7,990 examples
Task: Named Entity Recognition (NER) for carcinogen identification using BIO tagging
Performance:
- Precision: 0.894
- Recall: 0.857
- F1 Score: 0.875

Named Entity Labels

The model recognizes 4 entity types using BIO (Beginning-Inside-Outside) tagging scheme, resulting in 9 total labels:

Label ID	Label	Description
0	O	Outside any entity
1	B-carcinogen	Beginning of carcinogen entity
2	I-carcinogen	Inside carcinogen entity
3	B-negative	Beginning of negative/exculpatory evidence
4	I-negative	Inside negative evidence
5	B-cancertype	Beginning of cancer type/metadata
6	I-cancertype	Inside cancer type/metadata
7	B-antineoplastic	Beginning of anti-cancer agent
8	I-antineoplastic	Inside anti-cancer agent

Entity Type Descriptions:

carcinogen: Substances or agents implicated in carcinogenesis
negative: Exculpating evidence for potential carcinogenic entities
cancertype: Metadata including organism (human/animal/cell), cancer type, and affected organs
antineoplastic: Chemotherapy drugs and cancer-protective agents

Use Cases

Streamlining toxicogenomic literature reviews
Identifying potential carcinogens for further investigation
Augmenting existing carcinogen databases with emerging candidates
Extracting structured information from cancer research literature
Supporting evidence-based oncology research

Limitations

Identifies potential candidates, not confirmed carcinogens
Analysis limited to abstract-level information
May be influenced by publication trends and research focus shifts
Requires validation by domain experts for clinical applications

Installation

pip install transformers torch datasets

Usage

Basic Usage

from transformers import AutoTokenizer, AutoModelForTokenClassification
import torch

# Load model and tokenizer
model_name = "jimnoneill/CarD-T"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForTokenClassification.from_pretrained(model_name)

# Define label mappings
id2label = {
    0: "O",
    1: "B-carcinogen",
    2: "I-carcinogen",
    3: "B-negative",
    4: "I-negative",
    5: "B-cancertype",
    6: "I-cancertype",
    7: "B-antineoplastic",
    8: "I-antineoplastic"
}

Named Entity Recognition Pipeline

def predict_entities(text):
    # Tokenize input
    inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)
    
    # Get predictions
    with torch.no_grad():
        outputs = model(**inputs)
        predictions = outputs.logits.argmax(dim=2)
    
    # Convert tokens and predictions to entities
    tokens = tokenizer.convert_ids_to_tokens(inputs.input_ids[0])
    
    entities = []
    current_entity = None
    current_tokens = []
    
    for token, pred_id in zip(tokens, predictions[0]):
        pred_label = id2label[pred_id.item()]
        
        if pred_label == "O":
            if current_entity:
                entities.append({
                    "entity": current_entity,
                    "text": tokenizer.convert_tokens_to_string(current_tokens)
                })
                current_entity = None
                current_tokens = []
        elif pred_label.startswith("B-"):
            if current_entity:
                entities.append({
                    "entity": current_entity,
                    "text": tokenizer.convert_tokens_to_string(current_tokens)
                })
            current_entity = pred_label[2:]
            current_tokens = [token]
        elif pred_label.startswith("I-") and current_entity:
            current_tokens.append(token)
    
    # Don't forget the last entity
    if current_entity:
        entities.append({
            "entity": current_entity,
            "text": tokenizer.convert_tokens_to_string(current_tokens)
        })
    
    return entities

# Example usage
text = "Benzene exposure has been linked to acute myeloid leukemia, while vitamin D shows antineoplastic properties."
entities = predict_entities(text)
for entity in entities:
    print(f"{entity['entity']}: {entity['text']}")

Using with Hugging Face Pipeline

from transformers import pipeline

# Create NER pipeline
ner_pipeline = pipeline(
    "token-classification",
    model=model_name,
    aggregation_strategy="simple"
)

# Analyze text
text = "Studies show asbestos causes mesothelioma in humans, but aspirin may have protective effects."
results = ner_pipeline(text)

# Display results
for entity in results:
    print(f"{entity['entity_group']}: {entity['word']} (confidence: {entity['score']:.3f})")

Processing Scientific Abstracts

def analyze_abstract(abstract):
    """Analyze a scientific abstract for cancer-related entities."""
    entities = predict_entities(abstract)
    
    # Organize by entity type
    results = {
        "carcinogens": [],
        "protective_agents": [],
        "cancer_types": [],
        "negative_findings": []
    }
    
    for entity in entities:
        if entity['entity'] == "carcinogen":
            results["carcinogens"].append(entity['text'])
        elif entity['entity'] == "antineoplastic":
            results["protective_agents"].append(entity['text'])
        elif entity['entity'] == "cancertype":
            results["cancer_types"].append(entity['text'])
        elif entity['entity'] == "negative":
            results["negative_findings"].append(entity['text'])
    
    return results

# Example with a scientific abstract
abstract = """
Recent studies in male rats exposed to compound X showed increased incidence of 
hepatocellular carcinoma. However, concurrent administration of resveratrol 
demonstrated significant protective effects against liver tumor development. 
No carcinogenic activity was observed in female mice under similar conditions.
"""

analysis = analyze_abstract(abstract)
print("Analysis Results:")
for category, items in analysis.items():
    if items:
        print(f"\n{category.replace('_', ' ').title()}:")
        for item in items:
            print(f"  - {item}")

Training Configuration

The model was fine-tuned using the following configuration:

from transformers import TrainingArguments

training_args = TrainingArguments(
    output_dir="./card-t-model",
    learning_rate=2e-5,
    per_device_train_batch_size=16,
    per_device_eval_batch_size=16,
    num_train_epochs=5,
    weight_decay=0.01,
    evaluation_strategy="epoch",
    save_strategy="epoch",
    load_best_model_at_end=True,
    metric_for_best_model="f1",
    push_to_hub=True,
)

If you use this model in your research, please cite:

@article{oneill2024cardt,
  title={CarD-T: Interpreting Carcinomic Lexicon via Transformers},
  author={O'Neill, Jamey and Reddy, G.A. and Dhillon, N. and Tripathi, O. and Alexandrov, L. and Katira, P.},
  journal={MedRxiv},
  year={2024},
  doi={10.1101/2024.08.13.24311948}
}

License

This model is released under the Apache License 2.0, matching the license of the training dataset.

Acknowledgments

We thank the biomedical research community for making their findings publicly available through PubMed, enabling the creation of this model. Special thanks to the Bio-ELECTRA team for the base model architecture.

Contact

For questions, feedback, or collaborations:

Author: Jamey O'Neill
Email: [email protected]
Hugging Face: @jimnoneill
Dataset: CarD-T-NER

Disclaimer

This model is intended for research purposes only. It should not be used as a sole source for medical decisions or clinical diagnoses. Always consult with qualified healthcare professionals and validate findings through appropriate experimental methods.

jimnoneill
/

CarD-T