How tiger-ai-lab/disprotedit works
Project type and purpose
• DisProtEdit is a machine-learning framework (“protein editing framework”) with:
– A data pipeline ingesting and preprocessing the SwissProtDis dataset
– A dual-channel training pipeline (structural vs. functional representations)
– An editing/inference component that applies learned latent edits via text prompts
– Supporting artifacts: benchmarks, website, CI
(minimal) file structure
• .github/workflows/static.yml → Continuous-integration (static checks, linting)
• README.md, LICENSE → Documentation / licensing
• (Absence of src/ suggests code is in a different branch or hidden; infer standard ML project layout)
Extract architecture insights from README
• Core modules: dataset creation (SwissProtDis), alignment & regularization losses (MMD), disentangled latent encoder, text-based editing API
• Outputs: trained model checkpoints, editing benchmarks
• External touchpoints: arXiv, project website, future Hugging Face dataset
Main components
A. Data Layer
– SwissProtDis Dataset (raw protein sequences + GPT-4o-generated text descriptions)
– Data Preprocessing (tokenization, text-sequence alignment)
B. Training Pipeline
– Dual-Channel Encoder (Structural-Channel, Functional-Channel)
– Loss Modules (Alignment Loss, Uniformity Loss, Angular MMD Loss)
C. Model Artifacts
– Disentangled Latent Representations
– Checkpoints & Serialized Models
D. Editing / Inference Service
– Text-Prompt Interface (CLI or REST API)
– Sequence Editor (applies edits in latent space → generates new protein sequences)
E. Evaluation & Benchmarking
– Editing Benchmark Module (under construction)
– Metrics Dashboard
F. External Integrations
– Website (Docs + Demo)
– arXiv / Publication Link
– (Future) Hugging Face Dataset Hosting
G. CI/CD
– GitHub Actions (static analysis via .github/workflows/static.yml)
Relationships & Data Flow
• Within Training Pipeline: connect Encoder sub-modules to Loss Modules → Model Artifacts
• Artifacts feed both the Inference Service and the Evaluation Module
• Inference Service ↔ User / API Clients
• CI/CD monitors code changes → triggers linting/tests → deploys website & docs
Architectural patterns & principles
• Modular pipeline: clear separation of data ingestion, model training, inference, evaluation
• Dual-channel supervision: two parallel encoder streams with shared latent space
• Plug-and-play losses: each MMD/loss module “plugs in” via dependency injection
• Service-oriented: inference exposed as a standalone service (CLI/REST)
• Data layer: Python, BioPython (presumed), GPT-4o for text decomposition
• Training: PyTorch or JAX (inferred), CUDA/GPU acceleration
• API: FastAPI/Flask (if REST), Python CLI
• CI/CD: GitHub Actions (static analysis, future testing)
• Hosting: GitHub Pages (website), Hugging Face (dataset)