How microsoft/mu-protein works
μProtein combines protein-language-model pretraining, μFormer fitness prediction, and μSearch exploration. A typical workflow trains or loads a PMLM, fits a fitness predictor to mutation data, exposes that predictor as a landscape oracle, and searches for promising sequences. The graph also preserves the documented pretraining and modeling stages. Some internal links are based on SOURCE INDEX references or README descriptions rather than fully sampled implementations; those limits are marked with dashed edges where appropriate.
Main entry points: trainer.py Model trainer · models/pmlm.py PMLM model · tasks/pmlm.py Pretraining tasks · explorer.py Search loop
Components 21
PMLM pretraining
- Sequence tokenizer
pmlm/src/protein/tokenizers.py - Training datasets
pmlm/src/protein/datasets.py - Transformer encoder
pmlm/src/protein/models/encoder.py - Attention layers
pmlm/src/protein/models/multihead_attention.py - PMLM model
pmlm/src/protein/models/pmlm.py - Prediction heads
pmlm/src/protein/models/prediction.py - Pretraining objectives
pmlm/src/protein/criterions/pmlm.py - Pretraining tasks
pmlm/src/protein/tasks/pmlm.py
Fitness modeling
- Training interface
mu-former/main.py - Mutation datasets
mu-former/src/data.py - Model vocabulary
mu-former/src/vocab.py - Fitness predictor
mu-former/src/model.py - Fitness objective
mu-former/src/criterion.py - Model trainer
mu-former/src/trainer.py - Learning-rate schedule
mu-former/src/lr_scheduler.py
Landscape search
- Muformer landscape
mu-search/src/flexs/flexs/landscapes/muformer.py - Search loop
mu-search/src/flexs/flexs/explorer.py - Random strategy
mu-search/src/flexs/flexs/baselines/explorers/random.py - Sequence mutation
mu-search/src/flexs/flexs/utils/sequence_utils.py - Measured sequences
Other components
- Protein engineer
Connections 24
- Protein engineer → Training interface: configures training
mu-former/main.py - Training interface → Model trainer: starts training
mu-former/main.py - Model trainer → Mutation datasets: loads data
mu-former/src/trainer.py - Model trainer → Fitness predictor: trains model
mu-former/src/trainer.py - Model trainer → Fitness objective: selects loss
mu-former/src/trainer.py - Model trainer → Learning-rate schedule: configures schedule
mu-former/src/trainer.py - Model trainer → Model vocabulary: uses vocabulary
mu-former/src/trainer.py - Protein engineer → Pretraining tasks: pretrains model
README.md - Pretraining tasks → Training datasets: uses datasets
- Training datasets → Sequence tokenizer: tokenizes sequences
pmlm/src/protein/datasets.py - Pretraining tasks → PMLM model: trains model
- PMLM model → Transformer encoder: uses encoder
pmlm/src/protein/models/pmlm.py - Transformer encoder → Attention layers: uses layers
- PMLM model → Prediction heads: uses heads
pmlm/src/protein/models/pmlm.py - Pretraining tasks → Pretraining objectives: computes objective
- PMLM model → Training interface: provides pretrained model
README.md - Model trainer → Muformer landscape: supplies trained model
README.md - Protein engineer → Search loop: initiates search
README.md - Search loop → Random strategy: runs strategy
- Random strategy → Sequence mutation: generates mutants
mu-search/src/flexs/flexs/baselines/explorers/random.py - Random strategy → Muformer landscape: scores candidates
mu-search/src/flexs/flexs/baselines/explorers/random.py - Search loop → Muformer landscape: measures candidates
- Search loop → Measured sequences: records measurements
- Muformer landscape → Fitness predictor: loads predictor
mu-search/src/flexs/flexs/landscapes/muformer.py