How rathanpaida/drug_discovery_research works
TNK1 is a research platform for predicting kinase activity and generating, evaluating, and prioritizing candidate molecules. The documented workflow runs from pan-kinome data and molecular features through transfer learning, uncertainty calibration, profiling, reinforcement-learning generation, novelty checks, and docking toward a wet-lab contract. The graph uses the root package as the canonical implementation; the tree also contains a parallel TO_RUN_ON_GPU copy. Workflow edges rely on README documentation, while internal model and scorer links use the provided source index and excerpts.
Main entry points: pipeline.py Transfer Training · scorer.py Candidate Scoring · run_rl.py Policy Training · features.py Molecular Features
Components 24
Data and Features
- Activity Data
tnk1/data/activity.py - Data Splits
tnk1/data/splits.py - Molecular Features
tnk1/baselines/features.py - Stage Context
tnk1/context.py
Modeling and Uncertainty
- Model Data
tnk1/models/data.py - Kinase Features
tnk1/models/kinase_features.py - Transfer Training
tnk1/models/pipeline.py - Kinome Model
tnk1/models/model.py - Training Engine
tnk1/models/trainer.py - Model Ensemble
tnk1/uq/ensemble.py - Uncertainty Metrics
tnk1/uq/metrics.py - Conformal Calibration
tnk1/uq/calibration.py
Candidate Evaluation
- Candidate Scoring
tnk1/rl/scorer.py - Selectivity Profile
tnk1/selectivity/selectivity.py - ADMET Profile
tnk1/admet/admet.py - Candidate Filters
tnk1/admet/filters.py - Novelty Audit
tnk1/novelty/verifier.py - Molecular Docking
scripts/dock_40_molecules_tnk1.py - Wet-Lab Contract
tnk1/active_learning/contract.py
Molecule Generation
- Policy Training
rl_min/run_rl.py - DAP Algorithm
rl_min/dap.py - Policy Model
rl_min/model.py - SMILES Vocabulary
rl_min/vocab.py
Other components
- Researcher
Connections 25
- Researcher → Activity Data: provides data
README.md - Activity Data → Molecular Features: feeds features
README.md - Molecular Features → Transfer Training: supports training
README.md - Data Splits → Transfer Training: partitions data
tnk1/models/pipeline.py - Stage Context → Kinase Features: loads features
tnk1/context.py - Stage Context → Molecular Features: loads descriptors
tnk1/context.py - Transfer Training → Model Data: builds datasets
tnk1/models/pipeline.py - Transfer Training → Kinome Model: creates model
tnk1/models/pipeline.py - Transfer Training → Training Engine: runs epochs
tnk1/models/pipeline.py - Transfer Training → Kinase Features: uses features
tnk1/models/pipeline.py - Transfer Training → Uncertainty Metrics: reports metrics
tnk1/models/pipeline.py - Transfer Training → Data Splits: uses folds
tnk1/models/pipeline.py - Transfer Training → Model Ensemble: feeds ensemble
README.md - Model Ensemble → Conformal Calibration: calibrates predictions
README.md - Conformal Calibration → Candidate Scoring: informs scoring
README.md - Candidate Scoring → Selectivity Profile: uses profile
tnk1/rl/scorer.py - Candidate Scoring → ADMET Profile: uses profile
tnk1/rl/scorer.py - Candidate Scoring → Candidate Filters: applies filters
tnk1/rl/scorer.py - Candidate Scoring → Policy Training: scores candidates
README.md - Policy Training → DAP Algorithm: trains policy
rl_min/run_rl.py - Policy Training → Policy Model: uses model
rl_min/run_rl.py - Policy Training → SMILES Vocabulary: uses vocabulary
rl_min/run_rl.py - Policy Training → Novelty Audit: generated molecules audited
README.md - Novelty Audit → Molecular Docking: passes candidates
README.md - Molecular Docking → Wet-Lab Contract: supports prioritization
README.md