SGRPO

This repository contains the code for Supergroup Relative Policy Optimization (SGRPO) across three biomolecular post-training settings:

de novo small-molecule design
pocket-based small-molecule design
de novo protein design

For pocket-based design, evaluation is reported with AutoDock Vina docking metrics such as vina_dock_mean. The training-time scalar utility still uses unidock_score as the internal field name for the corresponding Vina-derived high-is-better proxy.

Recommended Repo-Relative Layout

All commands below assume:

you clone the repository as SGRPO
you run commands from the repository root
you place assets at the repo-relative paths shown below

SGRPO/
  checkpoints/
    genmol_v2_v1.0/
      model_v2.ckpt
    genmol_p_v1.0/
      5500.ckpt
  configs/
    public/
      protein_denovo_eval.yaml
      protein_denovo_train.yaml
      small_molecule_denovo_eval.yaml
      small_molecule_denovo_train.yaml
      small_molecule_pocket_tasks.tsv
      small_molecule_pocket_train.yaml
  data/
    crossdocked/
      processed/
        crossdocked_v1.1_rmsd1.0_pocket10_processed_final.lmdb
        crossdocked_pocket10_pose_split.pt
        crossdocked_pocket10_pocket_prefix_manifest.pt
      raw/
        test_set/
          ...
  runs/
    cpgrpo_denovo/
      sgrpo_denovo/
        checkpoint-002000/
          model.ckpt
    cpgrpo_denovo_pocket_prefix/
      sgrpo_pocket/
        checkpoint-001000/
          model.ckpt
    progen2_official/
    progen2_models/
      openfold_official/
      python_overlay/
      prot_bert_bfd/
      proteinsol_official/
      temberture_official/
    progen2_sgrpo/
      sgrpo_protein/
        checkpoint-000100/
          config.json
          generation_config.json
          model.safetensors
  scripts/
    exps/
      lead/
        docking/
          qvina02
  sgrpo-main-results/
    ...

sgrpo-main-results/ is git-ignored. It is the default workspace for sweep outputs, aggregated JSON files, and plotting inputs.

Environment

The final experiments used Linux x86-64, Python 3.10, PyTorch 2.6.0 with the CUDA 12.4 runtime, Transformers 4.52.4, Accelerate 1.13.0, and DeepSpeed 0.16.7. The supported recreation path is env/setup.sh; it installs the molecular, pocket-based, and protein dependencies used by the training and sweep pipelines.

The host must provide an NVIDIA driver compatible with CUDA 12.4. A local CUDA toolkit is not required for the small-molecule workflows, but it is required later when compiling the OpenFold extension for the protein workflow.

git clone <YOUR_GITHUB_URL>/SGRPO.git
cd SGRPO

source "$(conda info --base)/etc/profile.d/conda.sh"
export CONDA_ENV_NAME=sgrpo
export INSTALL_MM_PROGEN2_DEPS=1
bash env/setup.sh
conda activate "${CONDA_ENV_NAME}"

If you prefer an explicit prefix instead of a named conda environment:

source "$(conda info --base)/etc/profile.d/conda.sh"
export CONDA_ENV_PREFIX="$(pwd)/.conda/sgrpo"
bash env/setup.sh
conda activate "${CONDA_ENV_PREFIX}"

env/setup.sh performs the following steps:

creates a fresh Python 3.10 environment with pip==23.3.1
installs the base repo requirements from env/requirements.txt
installs the repo in editable mode
installs the Accelerate, DeepSpeed, PyTorch Geometric, and CUDA-12.4-compatible torch-scatter versions used by the final runs
installs the extra packages needed by the pocket-based and ProGen2 pipelines
pins setuptools<81 so that wandb==0.13.5 remains importable

Verify the recreated base environment before downloading model and dataset assets:

python - <<'PY'
import accelerate
import deepspeed
import torch
import transformers

expected = {
    "torch": "2.6.0",
    "transformers": "4.52.4",
    "accelerate": "1.13.0",
    "deepspeed": "0.16.7",
}
actual = {
    "torch": torch.__version__.split("+")[0],
    "transformers": transformers.__version__,
    "accelerate": accelerate.__version__,
    "deepspeed": deepspeed.__version__,
}
if actual != expected:
    raise RuntimeError(f"Environment version mismatch: expected={expected}, actual={actual}")
print(actual)
PY

Uni-Dock is intentionally installed into its own Conda environment, and the protein workflow uses a repo-local Python overlay plus a repo-local CUDA 12.4 compiler toolchain. Those two additions are described in their respective asset sections below; they are not hidden dependencies of the base environment.

One-Time Assets

1. De novo small-molecule checkpoint

The pretrained de novo initialization checkpoint is provided through the official NGC release:

ngc registry resource download-version "nvidia/clara/genmol_v2:1.0"

After the NGC download completes, copy model_v2.ckpt into the repo-relative path expected by the public config:

mkdir -p checkpoints/genmol_v2_v1.0
cp /path/to/downloaded/model_v2.ckpt checkpoints/genmol_v2_v1.0/model_v2.ckpt

This creates:

checkpoints/genmol_v2_v1.0/model_v2.ckpt

2. Pocket-based small-molecule assets

The pocket-based setup needs:

the initialization checkpoint
a processed CrossDocked LMDB
the corresponding split file
a pocket-prefix manifest
the raw CrossDocked test-set structure tree for Vina evaluation
a working unidock executable on PATH
a Vina-compatible executable at scripts/exps/lead/docking/qvina02

The pocket-based initialization checkpoint is available in the XinwuYe/SGRPO Hugging Face repository.

Download it directly into the repo-relative path used by the public config:

mkdir -p checkpoints/genmol_p_v1.0
curl -fL \
  'https://huggingface.co/XinwuYe/SGRPO/resolve/main/checkpoints/genmol_p_v1.0/5500.ckpt' \
  -o checkpoints/genmol_p_v1.0/5500.ckpt

This creates:

checkpoints/genmol_p_v1.0/5500.ckpt

The remaining pocket-based assets are fetched from their official sources and are not mirrored by this repository.

CrossDocked processed files and docking test set

Official sources:

processed LMDB / split / test_set.zip: TargetDiff data folder https://drive.google.com/drive/folders/1j21cc7-97TedKh_El5E34yI8o5ckI7eK?usp=share_link
original CrossDocked release: https://bits.csb.pitt.edu/files/crossdock2020/

The following commands download the official TargetDiff data folder to a temporary directory, copy the required files into the repo-relative paths expected by the public configs, and unpack the Vina docking test set:

python -m pip install gdown

mkdir -p data/crossdocked/processed data/crossdocked/raw

gdown --folder "https://drive.google.com/drive/folders/1j21cc7-97TedKh_El5E34yI8o5ckI7eK?usp=share_link" \
  -O .cache/targetdiff_crossdocked

cp "$(find .cache/targetdiff_crossdocked -name 'crossdocked_v1.1_rmsd1.0_pocket10_processed_final.lmdb' -print -quit)" \
  data/crossdocked/processed/crossdocked_v1.1_rmsd1.0_pocket10_processed_final.lmdb

cp "$(find .cache/targetdiff_crossdocked -name 'crossdocked_pocket10_pose_split.pt' -print -quit)" \
  data/crossdocked/processed/crossdocked_pocket10_pose_split.pt

cp "$(find .cache/targetdiff_crossdocked -name 'test_set.zip' -print -quit)" \
  data/crossdocked/raw/test_set.zip

unzip -q data/crossdocked/raw/test_set.zip -d data/crossdocked/raw

This should leave the public-config-required files at:

data/crossdocked/processed/crossdocked_v1.1_rmsd1.0_pocket10_processed_final.lmdb
data/crossdocked/processed/crossdocked_pocket10_pose_split.pt
data/crossdocked/raw/test_set/

Then build the pocket-prefix manifest:

python scripts/build_crossdocked_pocket_prefix_manifest.py \
  --crossdocked_lmdb_path data/crossdocked/processed/crossdocked_v1.1_rmsd1.0_pocket10_processed_final.lmdb \
  --crossdocked_split_path data/crossdocked/processed/crossdocked_pocket10_pose_split.pt \
  --output_path data/crossdocked/processed/crossdocked_pocket10_pocket_prefix_manifest.pt

This creates:

data/crossdocked/processed/crossdocked_pocket10_pocket_prefix_manifest.pt

If you want to regenerate the processed files from the original CrossDocked release instead of using the official TargetDiff preprocessed artifacts, the official v1.1 archive is:

mkdir -p data/CrossDocked2020_v1.1

curl -L https://bits.csb.pitt.edu/files/crossdock2020/v1.1/CrossDocked2020_v1.1.tgz \
  -o data/CrossDocked2020_v1.1/CrossDocked2020_v1.1.tgz

tar -xzf data/CrossDocked2020_v1.1/CrossDocked2020_v1.1.tgz \
  -C data/CrossDocked2020_v1.1

The CrossDocked2020 data files published at bits.csb.pitt.edu/files/crossdock2020/ are released under CC0 1.0.

Uni-Dock

Official source:

repository: https://github.com/dptech-corp/Uni-Dock

Official installation path:

conda create -y -n unidock_env unidock -c conda-forge
conda run -n unidock_env unidock --help
export PATH="$(conda info --base)/envs/unidock_env/bin:${PATH}"

The public training config expects a unidock executable on PATH when you launch the pocket-based workflow.

QuickVina 2

Official sources:

repository: https://github.com/QVina/qvina
compile instructions: https://qvina.github.io/compilingQvina2.html

The public evaluation command expects a qvina02 executable at scripts/exps/lead/docking/qvina02.

The quickest official route is to download the upstream binary and rename it in place:

mkdir -p scripts/exps/lead/docking

curl -L https://raw.githubusercontent.com/QVina/qvina/master/bin/qvina2.1 \
  -o scripts/exps/lead/docking/qvina02

chmod +x scripts/exps/lead/docking/qvina02

3. De novo protein assets

The ProGen2 setup needs:

the official ProGen2 tokenizer and checkpoint tree
a Python overlay containing pinned transformers and adapters
a compiled OpenFold extension installed into that overlay
ProtBERT
TemBERTure
Protein-Sol

These assets are fetched from their official sources and are not mirrored by this repository. The commands below download each asset into its expected repo-relative location with the expected on-disk name.

ProGen2 official release

Official source:

repository: https://github.com/enijkamp/progen2
checkpoints: https://storage.googleapis.com/sfr-progen-research/checkpoints/

python scripts/setup_progen2_official.py \
  --output-dir runs/progen2_official \
  --model progen2-small

This creates:

runs/progen2_official/tokenizer.json
runs/progen2_official/prompts_unconditional.txt
runs/progen2_official/checkpoints/progen2-small/

OpenFold

Official source:

repository: https://github.com/aqlaboratory/openfold

git clone https://github.com/aqlaboratory/openfold.git \
  runs/progen2_models/openfold_official

python scripts/setup_progen2_python_overlay.py \
  --overlay-dir runs/progen2_models/python_overlay

conda create -y \
  -p runs/progen2_models/cuda_toolchain_cu124 \
  -c nvidia/label/cuda-12.4.1 \
  cuda-nvcc cuda-cudart-dev cuda-libraries-dev

export CUDA_HOME="$(pwd)/runs/progen2_models/cuda_toolchain_cu124"
export PATH="${CUDA_HOME}/bin:${PATH}"
export LD_LIBRARY_PATH="${CUDA_HOME}/lib64:${CUDA_HOME}/targets/x86_64-linux/lib:${LD_LIBRARY_PATH:-}"

test -x "${CUDA_HOME}/bin/nvcc"
test -f "${CUDA_HOME}/include/cusparse.h"

python scripts/setup_openfold_extension.py \
  --source-dir runs/progen2_models/openfold_official \
  --overlay-dir runs/progen2_models/python_overlay

The OpenFold extension is compiled against the active PyTorch environment and must therefore be rebuilt if the PyTorch or CUDA minor version changes.

TemBERTure

Official source:

repository: https://github.com/ibmm-unibe-ch/TemBERTure

python scripts/setup_temberture_official.py \
  --output-dir runs/progen2_models/temberture_official

This populates:

runs/progen2_models/temberture_official/

Protein-Sol

Official source:

download endpoint: https://protein-sol.manchester.ac.uk/cgi-bin/utilities/download_sequence_code.php

python scripts/setup_proteinsol_official.py \
  --output-dir runs/progen2_models/proteinsol_official

This downloads the official zip bundle and extracts it to:

runs/progen2_models/proteinsol_official/protein-sol-sequence-prediction-software/

ProtBERT-BFD

Official source:

model card: https://huggingface.co/Rostlab/prot_bert_bfd

huggingface-cli download Rostlab/prot_bert_bfd \
  --local-dir runs/progen2_models/prot_bert_bfd

This downloads the official Hugging Face snapshot to:

runs/progen2_models/prot_bert_bfd/

For ProGen2 training and evaluation, the OpenFold overlay must be importable:

export PYTHONPATH="$(pwd)/runs/progen2_models/python_overlay:$(pwd)/src:${PYTHONPATH:-}"

4. Canonical SGRPO evaluation checkpoints

The public evaluation configs use fixed repo-relative checkpoint paths. To reproduce the canonical SGRPO evaluation results without retraining, download the three released SGRPO checkpoints from the XinwuYe/SGRPO Hugging Face repository.

The following commands place the checkpoints at the exact repo-relative paths used by the public evaluation configs:

De novo small-molecule SGRPO checkpoint

mkdir -p runs/cpgrpo_denovo/sgrpo_denovo/checkpoint-002000
curl -fL \
  'https://huggingface.co/XinwuYe/SGRPO/resolve/main/checkpoints/sgrpo_main/genmol_denovo_sgrpo_rewardsum_loo_2000/model.ckpt' \
  -o runs/cpgrpo_denovo/sgrpo_denovo/checkpoint-002000/model.ckpt

Pocket-based small-molecule SGRPO checkpoint

mkdir -p runs/cpgrpo_denovo_pocket_prefix/sgrpo_pocket/checkpoint-001000
curl -fL \
  'https://huggingface.co/XinwuYe/SGRPO/resolve/main/checkpoints/sgrpo_main/mmgenmol_sgrpo_unidock_rewardsum_loo_1000/model.ckpt' \
  -o runs/cpgrpo_denovo_pocket_prefix/sgrpo_pocket/checkpoint-001000/model.ckpt

De novo protein SGRPO checkpoint

mkdir -p runs/progen2_sgrpo/sgrpo_protein/checkpoint-000100
base='https://huggingface.co/XinwuYe/SGRPO/resolve/main/checkpoints/sgrpo_main/progen2_sgrpo_gw08_rewardsum_loo_100'
curl -fL "$base/config.json" \
  -o runs/progen2_sgrpo/sgrpo_protein/checkpoint-000100/config.json
curl -fL "$base/generation_config.json" \
  -o runs/progen2_sgrpo/sgrpo_protein/checkpoint-000100/generation_config.json
curl -fL "$base/model.safetensors" \
  -o runs/progen2_sgrpo/sgrpo_protein/checkpoint-000100/model.safetensors
curl -fL "$base/trainer_state.pt" \
  -o runs/progen2_sgrpo/sgrpo_protein/checkpoint-000100/trainer_state.pt

Public Configs

The files under configs/public/ are ready to use without editing if you follow the repo-relative layout above:

configs/public/small_molecule_denovo_train.yaml
configs/public/small_molecule_pocket_train.yaml
configs/public/protein_denovo_train.yaml
configs/public/small_molecule_denovo_eval.yaml
configs/public/small_molecule_pocket_tasks.tsv
configs/public/protein_denovo_eval.yaml

These public configs use fixed repo-relative output directories so that the training and evaluation commands line up directly. If you need a different layout or want multiple concurrent runs of the same workflow, copy a file out of configs/public/ and edit the copy.

Train the Canonical SGRPO Models

1. De novo small-molecule design

accelerate launch \
  --config_file configs/accelerate_zero2.yaml \
  --num_processes 8 \
  --main_process_port 29501 \
  scripts/train_cpgrpo_denovo.py \
  --config configs/public/small_molecule_denovo_train.yaml

This writes checkpoints to:

runs/cpgrpo_denovo/sgrpo_denovo/

2. Pocket-based small-molecule design

accelerate launch \
  --config_file configs/accelerate_zero2.yaml \
  --num_processes 8 \
  --main_process_port 29502 \
  scripts/train_cpgrpo_denovo_pocket_prefix.py \
  --config configs/public/small_molecule_pocket_train.yaml

This writes checkpoints to:

runs/cpgrpo_denovo_pocket_prefix/sgrpo_pocket/

3. De novo protein design

export PYTHONPATH="$(pwd)/runs/progen2_models/python_overlay:$(pwd)/src:${PYTHONPATH:-}"

accelerate launch \
  --config_file configs/accelerate_ddp_8gpu.yaml \
  --num_processes 8 \
  --main_process_port 29503 \
  scripts/train_progen2_sgrpo.py \
  configs/public/protein_denovo_train.yaml

This writes checkpoints to:

runs/progen2_sgrpo/sgrpo_protein/

The supported evaluation checkpoint for the protein workflow is checkpoint-000100.

If you rerun one of the public training configs after its output directory already exists, either remove the existing output directory first or copy the config and change output_dir. The training scripts are fail-fast and will not silently merge runs.

Reproduce the Canonical SGRPO Sweeps

1. De novo small-molecule paired ((\rho,\tau)) sweep

python scripts/eval_denovo_sgrpo.py \
  --config configs/public/small_molecule_denovo_eval.yaml

This evaluates the six supported operating points:

(randomness, temperature) = (0.1, 0.5)
(0.3, 0.8)
(0.5, 1.1)
(0.7, 1.4)
(0.9, 1.7)
(1.0, 2.0)

Outputs are written under:

sgrpo-main-results/small_molecule_denovo/paired_sweep/

2. Pocket-based paired ((\rho,\tau)) sweep with Vina docking

This workflow:

generates molecules at the six paired operating points
docks those generations with AutoDock Vina
aggregates the six operating points into one JSON summary

The public task manifest is:

configs/public/small_molecule_pocket_tasks.tsv

Create the output workspace:

mkdir -p \
  sgrpo-main-results/small_molecule_pocket/generation \
  sgrpo-main-results/small_molecule_pocket/docking \
  sgrpo-main-results/small_molecule_pocket/docking_cache \
  sgrpo-main-results/small_molecule_pocket/aggregate

Run generation:

tail -n +2 configs/public/small_molecule_pocket_tasks.tsv | while IFS=$'\t' read -r task_id model_name sweep_type sweep_value randomness temperature checkpoint_path output_path; do
  python scripts/sample_denovo_pocket_prefix.py \
    --checkpoint_path "${checkpoint_path}" \
    --manifest_path data/crossdocked/processed/crossdocked_pocket10_pocket_prefix_manifest.pt \
    --split test \
    --num_pockets 100 \
    --num_samples_per_pocket 16 \
    --generation_batch_size 1600 \
    --generation_temperature "${temperature}" \
    --randomness "${randomness}" \
    --min_add_len 60 \
    --seed 42 \
    --device cuda \
    --bf16 \
    --output_path "${output_path}"
done

Run docking:

tail -n +2 configs/public/small_molecule_pocket_tasks.tsv | while IFS=$'\t' read -r task_id model_name sweep_type sweep_value randomness temperature checkpoint_path generated_rows_path; do
  output_dir="sgrpo-main-results/small_molecule_pocket/docking/${model_name}/${sweep_type}_${sweep_value}"
  mkdir -p "${output_dir}"
  python scripts/dock_pocket_prefix_generated_rows.py \
    --generated_rows_path "${generated_rows_path}" \
    --manifest_path data/crossdocked/processed/crossdocked_pocket10_pocket_prefix_manifest.pt \
    --split test \
    --crossdocked_root data/crossdocked/raw/test_set \
    --qvina_path scripts/exps/lead/docking/qvina02 \
    --docking_cache_dir sgrpo-main-results/small_molecule_pocket/docking_cache \
    --output_rows_path "${output_dir}/docking.records.jsonl" \
    --output_summary_path "${output_dir}/docking.summary.json" \
    --num_workers 64 \
    --docking_modes vina_dock \
    --docking_exhaustiveness 8 \
    --docking_num_cpu 1 \
    --docking_num_modes 10 \
    --docking_timeout_gen3d 30 \
    --docking_timeout_dock 100 \
    --docking_box_size 20.0 20.0 20.0 \
    --progress_every 100
done

Aggregate the six points:

python scripts/aggregate_mmgenmol_sweep_results.py \
  --tasks_path configs/public/small_molecule_pocket_tasks.tsv \
  --docking_root sgrpo-main-results/small_molecule_pocket/docking \
  --output_dir sgrpo-main-results/small_molecule_pocket/aggregate \
  --output_prefix paired_sweep \
  --expected_num_tasks 6 \
  --expected_rows_per_task 1600 \
  --expected_num_pockets 100 \
  --expected_samples_per_pocket 16 \
  --docking_mode vina_dock \
  --plot_name_prefix small_molecule_pocket \
  --plot_title_prefix "Pocket-Based Design" \
  --qed_weight 0.3 \
  --sa_score_weight 0.2 \
  --drugclip_score_weight 0.0

The key docking metric in the aggregate output is vina_dock_mean.

3. De novo protein temperature sweep

export PYTHONPATH="$(pwd)/runs/progen2_models/python_overlay:$(pwd)/src:${PYTHONPATH:-}"
export PROGEN2_TASKS_PATH=sgrpo-main-results/protein_denovo/temperature_sweep/tasks.tsv

python scripts/progen2_sweep_pipeline.py build-tasks \
  --config configs/public/protein_denovo_eval.yaml

Run generation for each task id listed in PROGEN2_TASKS_PATH:

for task_id in $(tail -n +2 "${PROGEN2_TASKS_PATH}" | cut -f1); do
  python scripts/progen2_sweep_pipeline.py generate-task \
    --config configs/public/protein_denovo_eval.yaml \
    --task-id "${task_id}"
done

Run the packed GPU rewards:

python scripts/progen2_sweep_pipeline.py score-packed-gpu-reward \
  --config configs/public/protein_denovo_eval.yaml \
  --reward-name naturalness

python scripts/progen2_sweep_pipeline.py score-packed-gpu-reward \
  --config configs/public/protein_denovo_eval.yaml \
  --reward-name stability

Run the point-wise rewards and diversity jobs:

for task_id in $(tail -n +2 "${PROGEN2_TASKS_PATH}" | cut -f1); do
  python scripts/progen2_sweep_pipeline.py score-point-reward-task \
    --config configs/public/protein_denovo_eval.yaml \
    --reward-name foldability \
    --task-id "${task_id}"

  python scripts/progen2_sweep_pipeline.py score-point-reward-task \
    --config configs/public/protein_denovo_eval.yaml \
    --reward-name developability \
    --task-id "${task_id}"

  python scripts/progen2_sweep_pipeline.py score-point-diversity-task \
    --config configs/public/protein_denovo_eval.yaml \
    --task-id "${task_id}"
done

Aggregate the final sweep:

python scripts/progen2_sweep_pipeline.py aggregate \
  --config configs/public/protein_denovo_eval.yaml

Outputs are written under:

sgrpo-main-results/protein_denovo/temperature_sweep/

Acknowledgements

This repository builds on ideas, code, and assets from the following upstream projects:

Name		Name	Last commit message	Last commit date
assets		assets
configs		configs
data		data
env		env
prompts		prompts
scripts		scripts
src		src
tests		tests
vis-code		vis-code
.gitignore		.gitignore
README.md		README.md
VERSION		VERSION
pyproject.toml		pyproject.toml

Navigation Menu

Repository navigation

About

Repository files navigation

SGRPO

Recommended Repo-Relative Layout

Environment

One-Time Assets

1. De novo small-molecule checkpoint

2. Pocket-based small-molecule assets

CrossDocked processed files and docking test set

Uni-Dock

QuickVina 2

3. De novo protein assets

ProGen2 official release

OpenFold

TemBERTure

Protein-Sol

ProtBERT-BFD

4. Canonical SGRPO evaluation checkpoints

De novo small-molecule SGRPO checkpoint

Pocket-based small-molecule SGRPO checkpoint

De novo protein SGRPO checkpoint

Public Configs

Train the Canonical SGRPO Models

1. De novo small-molecule design

2. Pocket-based small-molecule design

3. De novo protein design

Reproduce the Canonical SGRPO Sweeps

1. De novo small-molecule paired ((\rho,\tau)) sweep

2. Pocket-based paired ((\rho,\tau)) sweep with Vina docking

3. De novo protein temperature sweep

Acknowledgements

About

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

SGRPO

Recommended Repo-Relative Layout

Environment

One-Time Assets

1. De novo small-molecule checkpoint

2. Pocket-based small-molecule assets

CrossDocked processed files and docking test set

Uni-Dock

QuickVina 2

3. De novo protein assets

ProGen2 official release

OpenFold

TemBERTure

Protein-Sol

ProtBERT-BFD

4. Canonical SGRPO evaluation checkpoints

De novo small-molecule SGRPO checkpoint

Pocket-based small-molecule SGRPO checkpoint

De novo protein SGRPO checkpoint

Public Configs

Train the Canonical SGRPO Models

1. De novo small-molecule design

2. Pocket-based small-molecule design

3. De novo protein design

Reproduce the Canonical SGRPO Sweeps

1. De novo small-molecule paired ((\rho,\tau)) sweep

2. Pocket-based paired ((\rho,\tau)) sweep with Vina docking

3. De novo protein temperature sweep

Acknowledgements

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages