Home

SignalForge Wiki

Welcome to the SignalForge documentation hub. SignalForge is a translational analytics platform that predicts how small molecules perturb gene expression and ranks candidate compounds against desired therapeutic signatures.

Quick Navigation

• Getting Started — Setup, installation, and first run
• Architecture Overview — System design and components
• API Reference — Complete API documentation
• Data Model — Database schema and data formats
• ML Pipeline — Model training, evaluation, and deployment
• Frontend Guide — UI components and user workflows
• Development — Contributing, local development, and testing
• Deployment — Production setup and operations
• FAQ & Troubleshooting — Common issues and solutions

What is SignalForge?

SignalForge turns transcriptomic reasoning into a reproducible, auditable, API-driven workflow. It enables researchers and drug discovery teams to:

1. Predict molecule-gene interactions — Input a small molecule and predict its regulatory effects on genes
2. Reverse-signature search — Upload a disease signature and rank compounds likely to invert it
3. Mechanism comparison — Compare how different compounds affect shared gene programs
4. Hit identification — Identify promising therapeutic candidates from large compound libraries

Key Features

Feature	Description
Fast predictions	Sub-second inference on millions of compound-gene pairs
Explainability	Pathway overlays and nearest-neighbor compound explanations
Multi-drug learning	Models trained on diverse perturbations (not single drugs)
Production-ready	Docker-containerized, versioned models, feature drift detection
Transparent	Clear preprocessing, realistic performance expectations

Project Status

Phase 4 (Complete) — Full product deployment with:

• Trained random forest and optional deep encoders
• FastAPI backend with eager model loading
• ~300-compound reverse-signature atlas from LINCS
• RDKit SMILES validation and inference provenance
• Docker containerization

Current Performance:

• Accuracy: 0.8736
• Macro F1: 0.8733

Technology Stack

Layer	Technology
Backend	Python 3.9+, FastAPI, scikit-learn, RDKit, PyTorch
Frontend	TypeScript, React, Tailwind CSS
Data	PostgreSQL, pandas, NumPy
Deployment	Docker, Docker Compose
ML	Random Forest, Neural Networks, Morgan fingerprints, GO embeddings

Repository Structure

SignalForge/
├── backend/              # FastAPI prediction service
├── frontend/             # React UI
├── ml/                   # Model training and evaluation
├── data/                 # Data schemas and ingestion
├── docker/               # Deployment configurations
├── docs/                 # Documentation and project brief
└── tests/                # Test suites

Getting Help

• Bug reports: Open an issue on GitHub
• Feature requests: Check existing issues or create a new one
• Questions: See FAQ & Troubleshooting
• Contributing: Read Development for guidelines

Latest Updates

• Model improvements: Upgraded to random forest with sample weighting (Phase 2)
• Multi-cell data: Expanded to full LINCS perturbation data (Phase 3)
• Production deployment: FastAPI + Docker containerization (Phase 4)
• Reverse search: ~300-compound atlas with SMILES validation

---

Last updated: September 2026
Maintained by: SignalForge team