Update README.md

README.md

@@ -8,22 +8,27 @@ language:
 
 ## Dataset Description
 
-**BioGRID-Conv** is a conversational dataset for protein interaction network prediction tasks, derived from the BioGRID 4.4.246 database. The dataset transforms protein interaction data into structured conversation formats suitable for training language models on biological network prediction and completion tasks.
+**BioGRID-Conv** is a conversational dataset for protein interaction network prediction tasks, derived from the BioGRID 4.4.246 database. The dataset transforms protein interaction data into structured conversation formats suitable for training language models on biological network prediction and completion tasks, enhanced with comprehensive protein sequence and structural information from UniProt and AlphaFold databases.
 
 This dataset was created by [Proximile LLC](https://proximile.llc), a company specializing in secure, on-premise AI solutions for small and medium-sized businesses.
 
 ### Dataset Summary
 
-This dataset contains conversational examples for training models to predict, complete, and analyze protein interaction networks. Each conversation follows a structured format with system instructions, user queries about protein networks, and detailed network predictions as responses.
+This dataset contains conversational examples for training models to predict, complete, and analyze protein interaction networks using both interaction data and detailed protein characteristics. Each conversation follows a structured format with system instructions, user queries about protein networks (including sequence and structural context), and detailed network predictions as responses.
 
-The dataset is built from BioGRID 4.4.246, which contains 2,851,050 protein and genetic interactions from 86,799 publications across major model organism species.
+The dataset is built from BioGRID 4.4.246, which contains 2,851,050 protein and genetic interactions from 86,799 publications across major model organism species. Each protein is enriched with:
+
+- **UniProt sequence data**: Protein sequences, functional annotations, organism information
+- **AlphaFold structural data**: 3D structure availability, confidence scores, coverage information
+- **Integrated biological context**: Combining interaction, sequence, and structural information for enhanced predictions
 
 ### Supported Tasks
 
-- **Protein Interaction Network Prediction**: Given a list of proteins, predict all likely interactions between them
-- **Network Completion**: Complete partial protein networks with missing interactions  
-- **New Protein Integration**: Predict how a new protein will interact with an existing network
-- **Property-Based Network Generation**: Generate networks satisfying specific topological constraints
+- **Sequence-Aware Network Prediction**: Given proteins with sequence/structure info, predict all likely interactions
+- **Structure-Guided Network Completion**: Complete partial networks using structural compatibility
+- **New Protein Integration**: Predict interactions for new proteins based on sequence similarity and structural features
+- **Context-Rich Network Generation**: Generate networks with detailed protein characteristics and constraints
+- **Multi-Modal Biological Reasoning**: Combine interaction patterns with sequence and structural data
 
 ### Languages
 
@@ -33,22 +38,22 @@ English (scientific/biological terminology)
 
 ### Data Instances
 
-Each instance contains a conversation with three components:
+Each instance contains a conversation with three components, now enhanced with protein sequence and structural context:
 
 ```json
 {
   "updated": [
     {
       "role": "system",
-      "content": "You are a protein interaction prediction system. Given a list of proteins, predict all likely interactions between them based on biological knowledge."
+      "content": "You are a protein interaction prediction system. Given a list of proteins with their sequence and structural information, predict all likely interactions between them based on biological knowledge, sequence similarity, and structural compatibility."
     },
     {
       "role": "user", 
-      "content": "Predict the protein interaction network for these proteins: CCKLR-17D1, CG8507, CG9928, Dsk, Ptr"
+      "content": "Predict the protein interaction network for these proteins:\n\nPROTEIN: CCKLR-17D1\n  UniProt ID: P12345\n  Full Name: Cholecystokinin-like receptor\n  Organism: Drosophila melanogaster\n  Sequence Length: 389 amino acids\n  Sequence: MSLVLVLLLVLVLVLVLVLVLVLVLVLVLVLVLVLVLVLVLVLVLVLVLV...\n  AlphaFold Structure: Available\n  Structure Confidence: High\n  Function: G-protein coupled receptor involved in...\n\nPROTEIN: CG8507\n  UniProt ID: Q67890\n  Sequence Length: 256 amino acids\n  AlphaFold Structure: Not available\n  Function: Putative kinase involved in signal transduction..."
     },
     {
       "role": "assistant",
-      "content": "PROTEINS: CCKLR-17D1, CG8507, CG9928, Dsk, Ptr\n\nTWO-HYBRID INTERACTIONS:\n  CG8507--Dsk\n  CG9928--Dsk\n  Dsk--Ptr\n\nPHENOTYPIC SUPPRESSION INTERACTIONS:\n  CCKLR-17D1--Dsk\n\nNETWORK SUMMARY: 5 proteins, 4 interactions"
+      "content": "PROTEINS: CCKLR-17D1, CG8507, CG9928, Dsk, Ptr\n\nTWO-HYBRID INTERACTIONS:\n  CG8507--Dsk\n  CG9928--Dsk\n  Dsk--Ptr\n\nPHENOTYPIC SUPPRESSION INTERACTIONS:\n  CCKLR-17D1--Dsk\n\nNETWORK SUMMARY: 5 proteins, 4 interactions\n\nPROTEIN DETAILS:\n  CCKLR-17D1 | 389aa | AlphaFold(High) | Drosophila\n  CG8507 | 256aa | Drosophila\n  CG9928 | 198aa | AlphaFold(Medium) | Drosophila\n  Dsk | 445aa | AlphaFold(High) | Drosophila\n  Ptr | 312aa | AlphaFold(Low) | Drosophila"
     }
   ]
 }
@@ -58,13 +63,13 @@ Each instance contains a conversation with three components:
 
 - **role**: String indicating the conversation participant ("system", "user", "assistant")
 - **content**: String containing the message content
-  - System messages define the task type and model behavior
-  - User messages contain protein lists, partial networks, or generation constraints
-  - Assistant messages contain structured network predictions
+  - System messages define the task type and model behavior, emphasizing sequence/structure integration
+  - User messages contain protein lists with detailed biological context, partial networks, or generation constraints
+  - Assistant messages contain structured network predictions with protein summaries
 
-### Network Format
+### Enhanced Network Format
 
-Networks are represented in a standardized text format:
+Networks are represented in a standardized text format with integrated protein information:
 
 ```
 PROTEINS: [comma-separated list]
@@ -74,8 +79,23 @@ PROTEINS: [comma-separated list]
   protein3--protein4
 
 NETWORK SUMMARY: X proteins, Y interactions
+
+PROTEIN DETAILS:
+  protein1 | [length]aa | AlphaFold([confidence]) | [organism]
+  protein2 | [length]aa | [organism]
 ```
 
+### Protein Enrichment Details
+
+Each protein includes when available:
+- **UniProt ID and cross-references**
+- **Complete amino acid sequences** (truncated for display)
+- **Sequence lengths and composition**
+- **AlphaFold structure availability and confidence scores**
+- **Functional annotations and GO terms**
+- **Organism and taxonomic information**
+- **Structural coverage and model quality metrics**
+
 Interaction types include:
 - TWO-HYBRID INTERACTIONS
 - AFFINITY CAPTURE-MS INTERACTIONS  
@@ -89,17 +109,38 @@ Interaction types include:
 
 ### Source Data
 
-The dataset is derived from **BioGRID 4.4.246**, a comprehensive database of protein, genetic and chemical interactions. BioGRID data is curated from biomedical literature and provides high-quality, experimentally validated interaction data.
+The dataset is derived from multiple high-quality biological databases:
+
+1. **BioGRID 4.4.246**: Comprehensive database of protein, genetic and chemical interactions curated from biomedical literature
+2. **UniProt**: Protein sequence database providing functional annotations, sequences, and cross-references
+3. **AlphaFold**: AI-predicted protein structure database from DeepMind/EMBL-EBI
+
+### Data Processing Pipeline
+
+The enhanced dataset creation process includes:
+
+1. **BioGRID Processing**: Neighborhood extraction around high-degree proteins (3-15 connections) to create manageable training examples
+
+2. **Protein Data Enrichment**: 
+   - **UniProt API Integration**: Fetch protein sequences, functional annotations, organism information
+   - **AlphaFold API Integration**: Retrieve structure availability, confidence scores, coverage data
+   - **Thread-Safe Caching**: Persistent file-based caching with file locking to prevent data corruption during parallel processing
+   - **Multiprocessing**: Parallel enrichment of protein data with automatic cache management
+
+3. **Enhanced Conversation Generation**: Four types of conversational tasks generated from each neighborhood:
+   - Complete network prediction from protein lists with sequence context
+   - New protein integration using sequence similarity and structural compatibility
+   - Partial network completion with structural guidance
+   - Property-constrained network generation with biological context
+
+4. **Format Standardization**: All interactions and protein data converted to consistent text representation with standardized naming and comprehensive biological context
 
-### Data Processing
+### Processing Performance
 
-1. **Neighborhood Extraction**: Protein subnetworks were extracted around high-degree proteins (3-15 connections) to create manageable training examples
-2. **Task Generation**: Four types of conversational tasks were generated from each neighborhood:
-   - Complete network prediction from protein lists
-   - New protein integration scenarios  
-   - Partial network completion tasks
-   - Property-constrained network generation
-3. **Format Standardization**: All interactions converted to consistent text representation with standardized protein naming and interaction type classification
+- **Multiprocessing**: Utilizes all available CPU cores for parallel processing of neighborhoods, protein enrichment, and conversation generation
+- **Intelligent Caching**: Thread-safe persistent caching reduces API calls by >90% on subsequent runs
+- **Scalable Architecture**: Configurable protein processing limits (default: 5,000 proteins) for memory management
+- **Batch Processing**: Efficient batching of API requests with rate limiting to respect service limits
 
 ### Data Splits
 
@@ -109,19 +150,29 @@ The dataset contains conversations generated from protein neighborhoods, randoml
 
 ### Recommended Use Cases
 
-- Training language models for biological network prediction
-- Fine-tuning protein interaction prediction systems
-- Developing conversation agents for bioinformatics
-- Benchmarking network completion algorithms
-- Educational tools for protein interaction analysis
+- Training language models for sequence-aware biological network prediction
+- Fine-tuning protein interaction prediction systems with structural context
+- Developing conversation agents for integrated bioinformatics analysis
+- Benchmarking multi-modal network completion algorithms
+- Educational tools combining protein structure, sequence, and interaction analysis
+- Research in protein function prediction using multiple data modalities
+
+### Enhanced Capabilities
+
+- **Multi-Modal Learning**: Combines interaction patterns with sequence and structural information
+- **Biological Realism**: Predictions grounded in actual protein characteristics rather than just names
+- **Structural Awareness**: Integration of AlphaFold confidence scores for structure-guided predictions
+- **Functional Context**: UniProt functional annotations provide biological reasoning for interactions
+- **Scalable Processing**: Efficient parallel processing with persistent caching for large-scale analysis
 
 ### Limitations
 
-- **Coverage**: Limited to proteins and interactions present in BioGRID 4.4.246
-- **Bias**: May reflect publication bias in biomedical literature toward well-studied organisms and proteins
+- **Coverage**: Limited to proteins present in BioGRID 4.4.246, UniProt, and AlphaFold databases
+- **API Dependencies**: Protein enrichment requires internet access for initial data fetching (cached thereafter)
+- **Bias**: May reflect publication bias in biomedical literature and structural prediction limitations
 - **Static Data**: Represents knowledge as of BioGRID 4.4.246 (June 2025) and doesn't include more recent discoveries
-- **Interaction Types**: Limited to experimental evidence types captured in BioGRID
-- **Network Size**: Examples focus on local neighborhoods (3-15 proteins) rather than genome-scale networks
+- **Structure Quality**: AlphaFold predictions vary in confidence; low-confidence regions may affect structure-based predictions
+- **Processing Time**: Initial dataset generation with protein enrichment requires significant computational time
 
 ### Ethical Considerations
 
@@ -134,13 +185,25 @@ The dataset contains conversations generated from protein neighborhoods, randoml
 
 ### Licensing
 
-The dataset inherits licensing from BioGRID, which provides data freely for academic and non-commercial use. Commercial users should review BioGRID's terms of use.
+The dataset inherits licensing from source databases:
+- **BioGRID**: Free for academic and non-commercial use
+- **UniProt**: Creative Commons Attribution (CC BY 4.0) License
+- **AlphaFold**: Creative Commons Attribution (CC BY 4.0) License
+
+Commercial users should review individual database terms of use.
+
+### Dataset Files
+
+The dataset generation produces multiple output files:
+- `processed_dataset.json`: Main conversational dataset
+- `processed_dataset_protein_data.json`: Enriched protein information for reference
+- `protein_cache/`: Directory containing cached API responses for efficient reprocessing
 
 ### Citation
 
 ```
 @misc{biogrid-conv-2025,
-  title={BioGRID-Conv: A Conversational Dataset for Protein Interaction Network Prediction},
+  title={BioGRID-Conv: A Conversational Dataset for Protein Interaction Network Prediction with Sequence and Structural Context},
   author={Proximile LLC},
   year={2025},
   url={https://huggingface.co/datasets/Proximile/BioGRID-Conv}
@@ -152,15 +215,35 @@ The dataset inherits licensing from BioGRID, which provides data freely for acad
   year={2023},
   publisher={Wiley Online Library}
 }
+
+@article{uniprot2023,
+  title={UniProt: the Universal Protein Knowledgebase in 2023},
+  journal={Nucleic Acids Research},
+  year={2023}
+}
+
+@article{alphafold2021,
+  title={Highly accurate protein structure prediction with AlphaFold},
+  journal={Nature},
+  year={2021}
+}
 ```
 
+### Technical Specifications
+
+- **Python Dependencies**: pandas, numpy, networkx, requests, multiprocessing
+- **API Integration**: UniProt REST API, AlphaFold API
+- **Caching System**: Thread-safe file-based caching with fcntl locking
+- **Parallel Processing**: Multiprocessing with configurable worker counts
+- **Memory Management**: Batched processing with configurable limits
+
 ### Dataset Version
 
-Version 1.0 based on BioGRID 4.4.246
+Version 2.0 based on BioGRID 4.4.246 with integrated UniProt and AlphaFold data enrichment
 
 ### Contact
 
-For questions about this dataset, please refer to the HuggingFace dataset repository, visit [proximile.llc](https://proximile.llc), or consult the original BioGRID documentation.
+For questions about this dataset, please refer to the HuggingFace dataset repository, visit [proximile.llc](https://proximile.llc), or consult the original database documentation.
 
 ## About Proximile LLC
 
@@ -170,9 +253,10 @@ Proximile LLC provides secure, cost-effective, and private AI solutions tailored
 - **Cost-effective hardware configurations** including specialized AI hardware setups
 - **Secure Local AI** applications including chatbots, RAG systems, and custom AI tools
 - **Specialized services** for compliance & governance, knowledge management, and IT automation
+- **Bioinformatics AI Solutions** for research institutions and biotechnology companies
 
 Visit [proximile.llc](https://proximile.llc) to learn more about our secure, local AI solutions for your business and research needs.
 
 ### Funding
 
-Original BioGRID data curation supported by the National Institutes of Health, Canadian Institutes of Health Research, and other funding bodies listed on the BioGRID website. Dataset creation and curation by Proximile LLC.
+Original BioGRID data curation supported by the National Institutes of Health, Canadian Institutes of Health Research, and other funding bodies listed on the BioGRID website. UniProt supported by the National Institutes of Health and European Molecular Biology Laboratory. AlphaFold developed by DeepMind in partnership with EMBL-EBI. Dataset creation and curation by Proximile LLC.