How magic-ai4med/map works
MAP predicts single-cell responses to drugs, including unprofiled drugs, using molecular and cellular representations informed by pharmacological knowledge. The documented workflow prepares Tahoe-100M, OP3 or SciPlex3 data, pre-trains knowledge encoders, trains the response model, then runs inference or evaluation. The graph includes the documented preprocessing and baseline branches; source excerpts give stronger implementation-level evidence for MAP training, evaluation and knowledge-encoder training than for the unsampled branches.
Main entry points: train.py MAP training · preprocess Preprocessing pipeline · train_resume.py Encoder pretraining · model.py Drug-gene model
Components 22
Data preparation
- Preprocessing pipeline
preprocess - Tahoe dataset
preprocess/data/tahoe.py
Knowledge encoders
- Knowledge graph data
MAP-KG/data/dataload.py - Drug-gene model
MAP-KG/model/model.py - Encoder pretraining
MAP-KG/train_resume.py - Molecular encoder
MAP-KG/model/mega_molbart/decoder.py
Response prediction
- Perturbation dataset
MAP/data/ds_multi_cell_lora_se.py - Cell-state encoder
MAP/model/se.py - Transformer layers
MAP/model/flash_transformer.py - MAP predictor
MAP/model/model.py - MAP training
MAP/train.py - Prediction demo
MAP/demo.py - Response evaluation
MAP/evaluate/eval_multi_cell.py - Evaluation metrics
MAP/evaluate/eval_utils.py
Baseline methods
- CRISP baseline
MAP/baseline/CRISP/train.py - PRnet baseline
MAP/baseline/PRnet/models/PRnet.py - XPert baseline
MAP/baseline/XPert/models/model_XPert.py - chemCPA baseline
MAP/baseline/chemCPA/chemCPA/model.py - Mean baseline
MAP/baseline/trainmean/eval_trainmean.py
Other components
- Single-cell datasets
- Researcher
- Pretrained weights
Connections 21
- Single-cell datasets → Preprocessing pipeline: prepares via
- Preprocessing pipeline → Tahoe dataset: loads Tahoe
preprocess/data/tahoe.py - Researcher → Encoder pretraining: starts pretraining
- Encoder pretraining → Knowledge graph data: loads batches
MAP-KG/train_resume.py - Encoder pretraining → Drug-gene model: trains
MAP-KG/train_resume.py - Drug-gene model → Molecular encoder: uses decoder
- Researcher → MAP training: starts training
- Preprocessing pipeline → Perturbation dataset: supplies inputs
- MAP training → Perturbation dataset: loads batches
MAP/train.py - MAP training → MAP predictor: trains
MAP/train.py - MAP predictor → Cell-state encoder: encodes cell state
MAP/model/model.py - Cell-state encoder → Transformer layers: uses layers
MAP/model/se.py - MAP training → Response evaluation: evaluates predictions
MAP/train.py - Response evaluation → Evaluation metrics: uses metrics
MAP/evaluate/eval_multi_cell.py - Researcher → Prediction demo: requests prediction
- Pretrained weights → Prediction demo: loads checkpoint
- Researcher → CRISP baseline: runs baseline
- Researcher → PRnet baseline: runs baseline
- Researcher → XPert baseline: runs baseline
- Researcher → chemCPA baseline: runs baseline
- Researcher → Mean baseline: runs baseline