How aitrics/freed works
FREED trains fragment-based molecular generators with reinforcement learning: a user selects a target and training variant, the learner proposes fragment edits in a molecule environment, and docking supplies target-dependent reward. Generated molecules can then be assessed for drug-likeness and hit enrichment. The README documents PE, BU, PER, curiosity, vanilla SAC, and PPO; the tree provides their implementation paths. Some orchestration and configuration wiring is not sampled, so uncertain links are omitted.
Main entry points: molecule_graph.py Molecule environment · run_rl.py Training entry · core_motif.py Graph policy · sac_motif_freed_bu.py FREED BU
Components 19
RL training
- Training entry
run_rl.py - SAC variants
sac_motif.py - FREED PE
sac_motif_freed_pe.py - FREED BU
sac_motif_freed_bu.py - Prioritized replay
sac_motif_per.py - PPO baseline
ppo_motif.py - Curiosity variants
sac_motif_freed_pe.py
Molecule generation
- Molecule environment
gym_molecule/envs/molecule_graph.py - Graph policy
core_motif.py - Value policy
core_motif_vbased.py - MC graph model
core_motif_mc.py - Fragment vocabularies
gym_molecule/dataset/motifs_91.txt
Docking
- Vina docking
gym_molecule/envs/docking_simple.py - QVina executable
bin/qvina02 - Target receptors
ReLeaSE_Vina/docking
Results evaluation
- Molecule quality filters
rl_final_results/libs/filters.py - Hit enrichment
rl_final_results/libs/enrichment.py - Generated molecule files
rl_final_results
Other components
- Researcher
Connections 28
- Researcher → Training entry: starts training
- Training entry → SAC variants: selects learner
- Training entry → FREED PE: selects PE
- Training entry → FREED BU: selects BU
- Training entry → Prioritized replay: selects PER
- Training entry → PPO baseline: selects PPO
- Training entry → Curiosity variants: selects curiosity
- SAC variants → Molecule environment: steps environment
- FREED PE → Molecule environment: steps environment
- FREED BU → Molecule environment: steps environment
- Prioritized replay → Molecule environment: steps environment
- PPO baseline → Molecule environment: steps environment
- Curiosity variants → Molecule environment: steps environment
- SAC variants → Graph policy: uses actor critic
- FREED PE → Graph policy: uses actor critic
- FREED BU → Graph policy: uses actor critic
- Prioritized replay → Graph policy: uses actor critic
- PPO baseline → Value policy: uses value policy
- Graph policy → Molecule environment: queries observations
- Value policy → Molecule environment: builds candidates
- MC graph model → FREED BU: supports uncertainty
- Molecule environment → Fragment vocabularies: loads fragments
- Molecule environment → Vina docking: scores molecules
- Vina docking → QVina executable: runs docking
- Vina docking → Target receptors: uses receptor files
- Training entry → Generated molecule files: writes molecules
- Generated molecule files → Molecule quality filters: supplies SMILES
- Generated molecule files → Hit enrichment: supplies docking scores