Protify Documentation Hub

July 28, 2026 · View on GitHub

Protify is a low-code platform for training and evaluating protein language models on property-prediction and benchmarking tasks. You can run it via CLI, YAML config, or a Tk GUI, with optional cloud compute (Modal) and precomputed embeddings. This hub is the entry point for all Protify documentation.

For the full project story, installation options, and high-level usage, see the main README in the repository root.

Quick start

Install: pip install -e . from the repo root (or use the Docker image). With Docker: run from repo root using the Getting started Docker section (no need to cd into src/protify).

Minimal CLI run (one model, one dataset, probe-only):

py -m src.protify.main --model_names ESM2-8 --data_names DeepLoc-2 --num_epochs 2

YAML run: Edit src/protify/yamls/base.yaml and run:

py -m src.protify.main --yaml_path src/protify/yamls/base.yaml

Results and logs go to results/ and logs/ by default. For full install options and where outputs go, see Getting started.

Documentation map

Section	Description	Link
Getting started	Installation (pip, venv, Docker), entry points, first CLI and YAML runs, output locations	getting_started.md
Configuration	All CLI argument groups, YAML config (base.yaml), merge behavior, model_names vs model_paths, full examples	cli_and_config.md
Data	DataArguments, supported datasets, data_dirs, loading flow, column normalization, translation flags, dataset classes	data.md
Models and embeddings	Base models (model_names / model_paths), get_base_model and get_tokenizer, EmbeddingArguments, Embedder flow, pooling, SQL vs PTH	models_and_embeddings.md
Probes and training	Probe types (linear, transformer, lyra), ProbeArguments, TrainerArguments, training flows, vectorized multi-seed linear probes, preflight planning, validation, num_runs, save/export	probes_and_training.md
Model components	For developers: attention, attention_utils, transformer, mlp; which probes use which	model_components.md
ProteinGym	ProteinGymRunner, scoring methods, run_proteingym_zero_shot, CLI args, compare_scoring_methods, benchmark	proteingym.md
Visualization	create_plots (TSV format, six PNGs), regression vs classification metrics, ci_plots, CLI	visualization.md
Embedding clustering	Cluster saved vector or matrix embeddings, optional scaling, algorithm comparison, assignments, metrics, and diagnostic plots	clustering.md
Logging and replay	MetricsLogger, results TSV shape, log_metrics, LogReplayer, replay flow	logging_and_replay.md
Hyperparameter optimization	W&B sweep, sweep.yaml, HyperoptModule, CLI args, best-config rerun	hyperparameter_optimization.md
Modal	Tk GUI Modal tab and modal_backend vs legacy Gradio app, config and path rewriting, deploy and run	modal.md
GUI	Tk GUI tabs, settings and full_args, flow (session, data, embeddings, trainer), background tasks	gui.md
Resources	Listing supported models and datasets (CLI and programmatic)	resource_listing.md
Testing	testing_suite layout, how to run pytest, Docker note	testing.md

Architecture overview

flowchart LR
  CLI[CLI / YAML]
  GUI[Tk GUI]
  MP[MainProcess]
  DataM[DataMixin]
  Embed[Embedder]
  Probe[get_probe]
  Train[TrainerMixin]
  Log[MetricsLogger]
  CLI --> MP
  GUI --> MP
  MP --> DataM
  MP --> Embed
  MP --> Probe
  MP --> Train
  MP --> Log
  DataM --> Embed
  Embed --> Train
  Probe --> Train
  Train --> Log
  Log --> results[results / plots]

Entry points: main.py (CLI/YAML) and gui.py (Tk). Both use the same MainProcess class.
MainProcess composes: DataMixin (data loading), Embedder (embeddings), get_probe (probe construction), TrainerMixin (training loop), and MetricsLogger (session log and results TSV).
Config is a single namespace from argparse plus optional YAML merge; see Configuration.

Quick start

Documentation map

Architecture overview

See also