How ersilia-os/eos7w6n works
This repository packages the GROVER molecular-embedding model. The principal workflow accepts compound input, generates GROVER fingerprints and appends RDKit descriptors, then writes fixed-width 5,000-feature output. The framework also includes self-supervised pretraining and supervised fine-tuning/evaluation. Relationships are limited to README statements and SOURCE INDEX references; unsampled implementation details are not inferred.
Main entry points: moldataset.py Molecule dataset · models.py GROVER encoder · pretrain.py Self-supervised pretraining · fingerprint.py Fingerprint task
Components 19
Embedding runtime
- CLI entry
model/framework/code/main.py - Fingerprint task
model/framework/code/task/fingerprint.py - RDKit descriptors
- GROVER checkpoint
model/checkpoints/grover_large.pt
Molecule representation
- Molecule features
model/framework/code/grover/data/molfeaturegenerator.py - Molecule dataset
model/framework/code/grover/data/moldataset.py - Molecule graph
model/framework/code/grover/data/molgraph.py - Feature scaling
model/framework/code/grover/data/scaler.py - Task labels
model/framework/code/grover/data/task_labels.py - Atom bond vocabulary
model/framework/code/grover/data/torchvocab.py
Model learning
- GROVER encoder
model/framework/code/grover/model/models.py - Transformer layers
model/framework/code/grover/model/layers.py - Self-supervised pretraining
model/framework/code/task/pretrain.py - Training engine
model/framework/code/task/grovertrainer.py - Fine-tuning
model/framework/code/task/train.py - Prediction
model/framework/code/task/predict.py
Model support
- Training utilities
model/framework/code/grover/util - Argument parsing
model/framework/code/grover/util/parsing.py
Other components
- Compound caller
Connections 22
- Compound caller → CLI entry: submits compounds
README.md - CLI entry → Fingerprint task: dispatches fingerprinting
model/framework/code/main.py - CLI entry → Argument parsing: parses options
model/framework/code/main.py - Fingerprint task → Molecule dataset: loads molecules
model/framework/code/task/fingerprint.py - Molecule dataset → Molecule features: generates features
model/framework/code/grover/data/moldataset.py - Molecule dataset → Feature scaling: uses scaler
model/framework/code/grover/data/moldataset.py - GROVER encoder → Molecule graph: uses graph batches
model/framework/code/grover/model/models.py - GROVER encoder → Transformer layers: uses layers
model/framework/code/grover/model/models.py - Fingerprint task → GROVER encoder: generates embeddings
model/framework/code/task/fingerprint.py - GROVER checkpoint → GROVER encoder: loads weights
README.md - Fingerprint task → RDKit descriptors: appends descriptors
README.md - RDKit descriptors → Compound caller: contributes 200 features
README.md - Fingerprint task → Compound caller: writes 5,000 features
README.md - Self-supervised pretraining → Molecule dataset: loads training data
model/framework/code/task/pretrain.py - Self-supervised pretraining → Atom bond vocabulary: loads vocabularies
model/framework/code/task/pretrain.py - Self-supervised pretraining → GROVER encoder: trains encoder
model/framework/code/task/pretrain.py - Self-supervised pretraining → Training engine: delegates training
model/framework/code/task/pretrain.py - Fine-tuning → Molecule dataset: loads datasets
model/framework/code/task/train.py - Fine-tuning → Training utilities: uses metrics and scheduling
model/framework/code/task/train.py - Prediction → Molecule dataset: loads input molecules
model/framework/code/task/predict.py - Prediction → GROVER encoder: runs model
model/framework/code/task/predict.py - Self-supervised pretraining → Task labels: uses pretraining labels
model/framework/code/task/pretrain.py