Datasets:

SeanWan05
/

MolAS-Active-Learning

# Legacy/internal README below
# **MolAS: Molecular Embedding–Based Algorithm Selection in Protein–Ligand Docking**

[![Journal of Cheminformatics](https://img.shields.io/badge/Journal%20of%20Cheminformatics-Published-2c7fb8.svg)](https://link.springer.com/article/10.1186/s13321-026-01168-8)
[![arXiv](https://img.shields.io/badge/arXiv-2512.02328-b31b1b.svg)](https://arxiv.org/abs/2512.02328)
[![Zenodo](https://img.shields.io/badge/Zenodo-10.5281%2Fzenodo.17760688-004d7f.svg)](https://zenodo.org/records/17760688)
[![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](LICENSE)

MolAS is a learning-based framework that predicts, for each protein–ligand pair, which docking or pose-generation method is most likely to succeed.
This repository contains the full training, evaluation, and analysis code used in the accompanying paper *Molecular Embedding–Based Algorithm Selection in Protein–Ligand Docking*.

---

## **Repository Layout**

```
molas/
│
├── main.py                 # Training / testing entry point (PyTorch Lightning)
├── main.sh                 # 5-fold and multi-GPU script (torchrun presets)
│
├── data/
│   ├── dataset.py          # Dataset construction from CSVs and precomputed graphs
│   ├── data_interface.py.  # LightningDataModule with K-Fold / split support
│   └── ...
│
├── model/                  # MolAS architectures and loss functions
│   └── ...
│
├── embedding_generation/   # Scripts for generating ESM and ChemBERTa embeddings (for completeness)
│
├── stats_test.py           # Fold-concatenated reporting and paired statistics
├── environment.yaml        # Python dependencies
└── README.md
```

---

## **Installation**

Create a new environment and install PyTorch **separately**.

```bash
# 1) Create env
conda env create -f environment.yaml
conda activate as_dock
# 2) Install PyTorch (choose a CUDA build)
pip install --index-url https://download.pytorch.org/whl/cu124 \
    torch torchvision torchaudio
# 3) Install remaining dependencies
pip install -r requirements.txt
```

---

## **Data Preparation**

Download the processed MolAS dataset and trained checkpoints from:

**Zenodo: https://zenodo.org/records/17760688**

Then update the `data_path` variables in `data/dataset.py`.

### **Expected directory structure**

```
{data_path}/
    ├── protein_graphs_esmc_600m/
    │     └── pyg_graph_{pdbid}_{ligandid}_esmc_600m.pt
    │
    ├── ligand_chemberta.pt
    │
    ├── {benchmark}_pb_ratio.csv     # PoseBusters validity (per-algorithm)
    └── {benchmark}_rmsd.csv         # RMSD targets (per-algorithm)
```

### **Active-learning matrices**

Active-learning runs use a separate reconstructed score matrix for each outer
cross-validation fold and AL seed:

```
AL/PORT-AS_cost_aware/
  {method}/
    fold{0..4}/
      seed{al_seed}/
        strict_score_matrices/
          strict_score_{percentage:03d}pct.csv
```

For outer fold `f`, the matrix contains only rows with `test_fold != f`.
Acquisition and reconstruction are performed separately on each fold's training
pool. A globally reconstructed matrix cropped to the training rows afterward is
not equivalent and is not used by the current pipeline. The matching
`split_exports/...csv` file defines both row membership and row order.

The split lifecycle is:

```
outer-fold training rows
→ fold-specific active learning and reconstruction
→ internal MolAS train/validation split
→ evaluation on the untouched outer test fold using ground-truth labels
```

### Notes on checkpoints

Log folder of checkpoints (named `lightning_logs`) should be placed inside the repository root, **parallel** to this README.md.

---

## **Quick Start**

### **5-Fold and Multi-GPU Training**

Use the preset runner:

```bash
bash main.sh
```

Configure:

* `benchmark` (e.g., `posex_self_docking`, `posex_cross_docking`, `astex_posex`, `moad`, `posebusters`)
* `relaxation` (`both | true | false`)
* `incl_columns` (algorithm list; ensure count matches `--num_classes`)
* `devices` (e.g., `"0,1,2,3"` for torchrun)
* `test=1` to run test-only mode and export reports
* `version=0` to set the initial version of the 5-version bundle for testing

#### **Version Directory**

To reproduce the results using the Zenodo-released checkpoints, set `version` according to the target benchmark and post-processing mode. Each entry refers to a contiguous 5-checkpoint bundle (`version`–`version+4`).

| Benchmark     | Post-process        | `version` | Version bundle |
| ------------- | ------------------- | --------- | -------------- |
| MOAD-curated  | —                   | 5         | 5–9            |
| PoseX + Astex | Mixed               | 219       | 219–223        |
| PoseX-SD      | Mixed               | 5         | 5–9            |
| PoseX-CD      | Mixed               | 5         | 5–9            |
| PoseBusters   | Mixed               | 20        | 20–24          |
| PoseX + Astex | No                  | 209       | 209–213        |
| PoseX + Astex | Relaxation          | 214       | 214–218        |
| PoseX-SD      | No                  | 10        | 10–14          |
| PoseX-SD      | Relaxation          | 30        | 30–34          |
| PoseX-CD      | No                  | 30        | 30–34          |
| PoseX-CD      | Relaxation          | 25        | 25–29          |
| PoseBusters   | No                  | 10        | 10–14          |
| PoseBusters   | Energy minimisation | 15        | 15–19          |


### **Single-GPU Example**

```bash
python main.py \
    --benchmark posex_self_docking \
    --batch_size 16 \
    --lr 1e-3 \
    --max_epochs 100 \
    --devices 0 \
    --model MolAS \
    --num_node_features 1152 \
    --num_classes 24 \
    --incl_columns surfdock unimol diffdock diffdock_l gnina autodock glide alphafold3 moe dynamicbind rfaa fabind equibind boltz1x tankbind neuralplexer chai deepdock boltz ifd protenix diffdock_pocket DS interformer
```

### **Testing With a Checkpoint**

```bash
python main.py \
    --benchmark posex_self_docking \
    --devices 0 \
    --model MolAS \
    --num_node_features 1152 \
    --num_classes 24 \
    --incl_columns <algorithm list> \
    --test \
    --ckpt_path path/to/checkpoint.ckpt
```

---

## **Important Arguments**

* `--model`
  `MolAS` (default) or multi-GNN variants (e.g., `MolAS_GCN_GAT_GINE`, `MolAS_EGNN_GAT_GINE`, `MolASGT`).
  Multi-GNN variants require node features and adjacency consistent with provided graphs.

* `--incl_columns`
  Space-separated algorithm names. Must match `--num_classes`.

* `--use_ndcg_loss`, `--ndcg_loss_weight`, `--ndcg_k`
  Enable LambdaNDCG ranking loss.

* `--use_logistic_loss`, `--logistic_loss_weight`
  Pairwise logistic ranking objective.

* `--k_folds`, `--fold_num`
  Enable K-fold cross-validation.

* `--log_root`, `--log_name`, `--log_version`
  Control Lightning logging layout.

---

## **Acknowledgements**

MolAS uses embeddings from **ESM-C** and **ChemBERTa**, and builds on PyTorch, PyTorch Lightning, and PyTorch Geometric.

	# Legacy/internal README below
	# MolAS: Molecular Embedding–Based Algorithm Selection in Protein–Ligand Docking

	[![Journal of Cheminformatics](https://img.shields.io/badge/Journal%20of%20Cheminformatics-Published-2c7fb8.svg)](https://link.springer.com/article/10.1186/s13321-026-01168-8)
	[![arXiv](https://img.shields.io/badge/arXiv-2512.02328-b31b1b.svg)](https://arxiv.org/abs/2512.02328)
	[![Zenodo](https://img.shields.io/badge/Zenodo-10.5281%2Fzenodo.17760688-004d7f.svg)](https://zenodo.org/records/17760688)
	[![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](LICENSE)

	MolAS is a learning-based framework that predicts, for each protein–ligand pair, which docking or pose-generation method is most likely to succeed.
	This repository contains the full training, evaluation, and analysis code used in the accompanying paper Molecular Embedding–Based Algorithm Selection in Protein–Ligand Docking.

	---

	## Repository Layout

	```
	molas/
	│
	├── main.py # Training / testing entry point (PyTorch Lightning)
	├── main.sh # 5-fold and multi-GPU script (torchrun presets)
	│
	├── data/
	│ ├── dataset.py # Dataset construction from CSVs and precomputed graphs
	│ ├── data_interface.py. # LightningDataModule with K-Fold / split support
	│ └── ...
	│
	├── model/ # MolAS architectures and loss functions
	│ └── ...
	│
	├── embedding_generation/ # Scripts for generating ESM and ChemBERTa embeddings (for completeness)
	│
	├── stats_test.py # Fold-concatenated reporting and paired statistics
	├── environment.yaml # Python dependencies
	└── README.md
	```

	---

	## Installation

	Create a new environment and install PyTorch separately.

	```bash
	# 1) Create env
	conda env create -f environment.yaml
	conda activate as_dock

	# 2) Install PyTorch (choose a CUDA build)
	pip install --index-url https://download.pytorch.org/whl/cu124 \
	torch torchvision torchaudio

	# 3) Install remaining dependencies
	pip install -r requirements.txt
	```

	---

	## Data Preparation

	Download the processed MolAS dataset and trained checkpoints from:

	Zenodo: https://zenodo.org/records/17760688

	Then update the `data_path` variables in `data/dataset.py`.

	### Expected directory structure

	```
	{data_path}/
	├── protein_graphs_esmc_600m/
	│ └── pyg_graph_{pdbid}_{ligandid}_esmc_600m.pt
	│
	├── ligand_chemberta.pt
	│
	├── {benchmark}_pb_ratio.csv # PoseBusters validity (per-algorithm)
	└── {benchmark}_rmsd.csv # RMSD targets (per-algorithm)
	```

	### Active-learning matrices

	Active-learning runs use a separate reconstructed score matrix for each outer
	cross-validation fold and AL seed:

	```
	AL/PORT-AS_cost_aware/
	{method}/
	fold{0..4}/
	seed{al_seed}/
	strict_score_matrices/
	strict_score_{percentage:03d}pct.csv
	```

	For outer fold `f`, the matrix contains only rows with `test_fold != f`.
	Acquisition and reconstruction are performed separately on each fold's training
	pool. A globally reconstructed matrix cropped to the training rows afterward is
	not equivalent and is not used by the current pipeline. The matching
	`split_exports/...csv` file defines both row membership and row order.

	The split lifecycle is:

	```
	outer-fold training rows
	→ fold-specific active learning and reconstruction
	→ internal MolAS train/validation split
	→ evaluation on the untouched outer test fold using ground-truth labels
	```

	### Notes on checkpoints

	Log folder of checkpoints (named `lightning_logs`) should be placed inside the repository root, parallel to this README.md.

	---

	## Quick Start

	### 5-Fold and Multi-GPU Training

	Use the preset runner:

	```bash
	bash main.sh
	```

	Configure:

	* `benchmark` (e.g., `posex_self_docking`, `posex_cross_docking`, `astex_posex`, `moad`, `posebusters`)
	* `relaxation` (`both \| true \| false`)
	* `incl_columns` (algorithm list; ensure count matches `--num_classes`)
	* `devices` (e.g., `"0,1,2,3"` for torchrun)
	* `test=1` to run test-only mode and export reports
	* `version=0` to set the initial version of the 5-version bundle for testing

	#### Version Directory

	To reproduce the results using the Zenodo-released checkpoints, set `version` according to the target benchmark and post-processing mode. Each entry refers to a contiguous 5-checkpoint bundle (`version`–`version+4`).

	\| Benchmark \| Post-process \| `version` \| Version bundle \|
	\| ------------- \| ------------------- \| --------- \| -------------- \|
	\| MOAD-curated \| — \| 5 \| 5–9 \|
	\| PoseX + Astex \| Mixed \| 219 \| 219–223 \|
	\| PoseX-SD \| Mixed \| 5 \| 5–9 \|
	\| PoseX-CD \| Mixed \| 5 \| 5–9 \|
	\| PoseBusters \| Mixed \| 20 \| 20–24 \|
	\| PoseX + Astex \| No \| 209 \| 209–213 \|
	\| PoseX + Astex \| Relaxation \| 214 \| 214–218 \|
	\| PoseX-SD \| No \| 10 \| 10–14 \|
	\| PoseX-SD \| Relaxation \| 30 \| 30–34 \|
	\| PoseX-CD \| No \| 30 \| 30–34 \|
	\| PoseX-CD \| Relaxation \| 25 \| 25–29 \|
	\| PoseBusters \| No \| 10 \| 10–14 \|
	\| PoseBusters \| Energy minimisation \| 15 \| 15–19 \|


	### Single-GPU Example

	```bash
	python main.py \
	--benchmark posex_self_docking \
	--batch_size 16 \
	--lr 1e-3 \
	--max_epochs 100 \
	--devices 0 \
	--model MolAS \
	--num_node_features 1152 \
	--num_classes 24 \
	--incl_columns surfdock unimol diffdock diffdock_l gnina autodock glide alphafold3 moe dynamicbind rfaa fabind equibind boltz1x tankbind neuralplexer chai deepdock boltz ifd protenix diffdock_pocket DS interformer
	```

	### Testing With a Checkpoint

	```bash
	python main.py \
	--benchmark posex_self_docking \
	--devices 0 \
	--model MolAS \
	--num_node_features 1152 \
	--num_classes 24 \
	--incl_columns <algorithm list> \
	--test \
	--ckpt_path path/to/checkpoint.ckpt
	```

	---

	## Important Arguments

	* `--model`
	`MolAS` (default) or multi-GNN variants (e.g., `MolAS_GCN_GAT_GINE`, `MolAS_EGNN_GAT_GINE`, `MolASGT`).
	Multi-GNN variants require node features and adjacency consistent with provided graphs.

	* `--incl_columns`
	Space-separated algorithm names. Must match `--num_classes`.

	* `--use_ndcg_loss`, `--ndcg_loss_weight`, `--ndcg_k`
	Enable LambdaNDCG ranking loss.

	* `--use_logistic_loss`, `--logistic_loss_weight`
	Pairwise logistic ranking objective.

	* `--k_folds`, `--fold_num`
	Enable K-fold cross-validation.

	* `--log_root`, `--log_name`, `--log_version`
	Control Lightning logging layout.

	---

	## Acknowledgements

	MolAS uses embeddings from ESM-C and ChemBERTa, and builds on PyTorch, PyTorch Lightning, and PyTorch Geometric.