Step-by-step guide to extracting biomedical entities from text with transformer models on the platform.

You want to pull drugs, genes, and diseases out of scientific text; this walks you through named entity extraction from a PubMed abstract in a few steps — pick the tool, supply your text, and read the entities back.

Quick entity extraction with .bloom* platform

The Transformers tool uses domain-specific models like BioBERT and PubMedBERT to extract drugs, genes, diseases, and other biomedical entities from scientific text. This guide walks through extracting entities from a PubMed abstract.

Step-by-step NER workflow

Access the Transformers tool

Navigate to Transformers in the .bloom* tool selector, or ask Ivy directly:

"Extract entities from this PubMed abstract"
"Run NER on this text using BioBERT"

Prepare your text input

Option A: Direct text entry (fastest)

Paste an abstract or paragraph into the text input field
Example: "Gefitinib, an EGFR tyrosine kinase inhibitor, showed significant activity in non-small cell lung cancer patients with EGFR mutations."

Option B: PubMed lookup

Enter one or more PubMed IDs (e.g., PMID:12345678)
.bloom* fetches the abstract text automatically

Option C: File upload

Upload a text file with one document per line
Or upload a CSV with a text column

Option D: From previous jobs

Import text from literature search or web search results

Configure the NLP task

Basic settings (recommended for first run):

Task: Named Entity Recognition (NER) - extracts entity mentions and types
Model: PubMedBERT (best accuracy for PubMed text)
Keep default entity types (Drug, Gene, Disease, Chemical)

Advanced settings (via gear icon):

Task: Choose NER, Text Classification, Zero-Shot Classification, Question Answering, Summarization, or Embedding Generation
Model: Enter a Hugging Face model ID — BioBERT, PubMedBERT, SciBERT, ChemBERTa, and BioGPT are common biomedical choices
Entity types: Select which entity types to extract (Drug, Gene, Disease, Chemical, Mutation, Cell Line)
Confidence threshold: Minimum confidence for entity recognition (default: 0.5)

Submit and monitor the job

Click Extract Entities (or the appropriate action for your selected task)
.bloom* shows real-time progress through tokenization and inference
Typical completion: 5-15 seconds per abstract
Batch processing handles multiple documents in parallel

Review and analyze results

NER results:

Annotated text with highlighted entities color-coded by type
Entity table listing each mention with type, span, and confidence score
Entity frequency summary showing the most frequently mentioned entities

Text classification results (if selected):

Predicted class label with confidence
Class probability distribution across all categories

Question answering / summarization results (if selected):

Extracted answer span or generated summary, with a confidence score for QA

Embedding results (if selected):

Dense vector representation of the input text
Similarity scores if comparing multiple documents

Download options:

JSON: Structured entity annotations with positions and confidence
CSV: Entity table with columns for text, type, start, end, and score
Annotated text: HTML with highlighted entities

Common NLP scenarios

Single abstract analysis

Paste an abstract about your target or therapeutic area
Run NER to extract all mentioned entities
Review the entity frequency table for key players
Use extracted entities to inform database searches (UniProt, PubChem)

Batch literature mining

Collect 50-100 PubMed IDs on a topic of interest
Run NER on all abstracts in batch mode
Aggregate entity frequencies across all abstracts
Identify the most frequently mentioned genes, drugs, and diseases
Build an entity co-occurrence network for knowledge discovery

Drug-target interaction extraction

Select NER as the task and run it on abstracts mentioning drug-target relationships
Review the co-occurring drug and gene/protein entities in each abstract
Use question answering (e.g., "what does gefitinib inhibit?") to pull out the specific relationship in context
Feed confirmed interactions into Target Passport for validation

Document similarity clustering

Select Embedding Generation as the task
Input a collection of abstracts or documents
Download per-document embeddings
Cluster documents by embedding similarity
Identify thematic groups for systematic review

Quality control checklist

Before submission

Text is in English biomedical language
Documents are within the 512-token limit (or will be chunked)
Selected model matches the text domain (PubMedBERT for PubMed, ChemBERTa for chemistry)
Appropriate task selected (NER, RE, Classification, or Embedding)

After extraction

Extracted entities are recognized and correctly typed
Confidence scores are reasonable (most entities above threshold)
Entity boundaries are correct (no partial words or merged entities)
Known entities in the text are captured (spot-check)

Next steps after extraction

Knowledge integration:

Use extracted gene names to query UniProt for protein information
Use extracted drug names to query PubChem for chemical structure
Feed entity lists into Target Passport or Compound Passport for enrichment

Literature analysis:

Build entity co-occurrence matrices for relationship discovery
Track entity mention frequency over time for trend analysis
Cluster documents by semantic similarity using generated embeddings

Pipeline automation:

Chain Transformers with PubMed search for automated literature monitoring
Feed extracted entities into database lookup tools for structured data enrichment
Build custom literature intelligence pipelines for ongoing project support

Transformers quickstart