MolCRAFT Series for Drug Design: MolPilot

Welcome to the official repository for the MolCRAFT series of projects! This series focuses on developing and improving deep learning models for structure-based drug design (SBDD) and molecule optimization (SBMO). Our goal is to create molecules with high binding affinity and plausible 3D conformations.

This repository contains the source code for the following projects:

• MolCRAFT: Structure-Based Drug Design in Continuous Parameter Space (ICML'24)
• MolJO: Empower Structure-Based Molecule Optimization with Gradient Guided Bayesian Flow Networks (ICML'25)
• MolPilot: Piloting Structure-Based Drug Design via Modality-Specific Optimal Schedule (ICML'25)

📜 Overview

The MolCRAFT series addresses critical challenges in generative models for SBDD, including modeling molecular geometries, handling hybrid continuous-discrete spaces, and optimizing molecules against protein targets. Each project introduces novel methodologies and achieves state-of-the-art performance on relevant benchmarks.

🧭 Navigation

Folder	TL, DR	Description
MolCRAFT	Unified Space for Molecule Generation	MolCRAFT is the first SBDD generative model based on Bayesian Flow Network (BFN) operating in the unified continuous parameter space for different modalities, with variance reduction sampling strategy to generate high-quality samples with more than 10x speedup.
MolJO	Gradient-Guided Molecule Optimization	MolJO is a gradient-based Structure-Based Molecule Optimization (SBMO) framework derived within BFN. It employs joint guidance across continuous coordinates and discrete atom types, alongside a backward correction strategy for effective optimization.
MolPilot	Optimal Scheduling	MolPilot enhances SBDD by introducing a VLB-Optimal Scheduling (VOS) strategy for the twisted multimodal probability paths, significantly improving molecular geometries and interaction modeling, achieving 95.9% PB-Valid rate.

---

🚀 MolPilot

Official implementation of ICML 2025 "Piloting Structure-Based Drug Design via Modality-Specific Optimal Schedule".

We propose VLB-Optimal Scheduling (VOS) and demonstrate its generality on the popular diffusion-based models (TargetDiff, with the code in targetdiff folder) and BFN-based models (our MolPilot).

In fact, VOS can be easily integrated into other frameworks, with only minor changes w.r.t. training:

# Example: TargetDiff molopt_score_model.py

class ScorePosNet3D(nn.Module):
  def get_diffusion_loss(...):
    ##### Original Training Loss #####
    time_step, pt = self.sample_time(num_graphs, protein_pos.device, self.sample_time_method)
    # Xt = a.sqrt() * X0 + (1-a).sqrt() * eps
    ligand_pos_perturbed = a_pos.sqrt() * ligand_pos + (1.0 - a_pos).sqrt() * pos_noise  # pos_noise * std

    ##### VOS Generalized Loss #####
    time_step_v, pt = self.sample_time(num_graphs, protein_pos.device, self.sample_time_method)
    # Vt = a * V0 + (1-a) / K
    log_ligand_v0 = index_to_log_onehot(ligand_v, self.num_classes)
    ligand_v_perturbed, log_ligand_vt = self.q_v_sample(log_ligand_v0, time_step_v, batch_ligand)
    kl_v = self.compute_v_Lt(log_v_model_prob=log_v_model_prob, log_v0=log_ligand_v0,
                             log_v_true_prob=log_v_true_prob, t=time_step_v, batch=batch_ligand)

The optimal test-time noise schedule can be obtained by first storing the gridded loss surface values, and then running the dynamic programming script in test/test_geodesic_budget.py.

Environment

It is highly recommended to install via docker if a Linux server with NVIDIA GPU is available.

Otherwise, you might check README for env for further details of docker or conda setup.

Prerequisite

A docker with nvidia-container-runtime enabled on your Linux system is required.

• This repo provides an easy-to-use script to install docker and nvidia-container-runtime, in ./docker run sudo ./setup_docker_for_host.sh to set up your host machine.
• For details, please refer to the install guide.

Install via Docker

We highly recommend you to set up the environment via docker, since all you need to do is a simple make command.

cd ./docker
make

---

Data

We use the same data as TargetDiff. Data used for training / evaluating the model should be put in the data folder by default, and accessible in the data Google Drive folder.

To train the model from scratch, download the lmdb file and split file into data folder:

• crossdocked_v1.1_rmsd1.0_pocket10_processed_final.lmdb
• crossdocked_pocket10_pose_split.pt

To evaluate the model on the test set, download and unzip the test_set.zip into data folder. It includes the original PDB files that will be used in Vina Docking.

data:
  name: pl # [pl, pl_tr] where tr means offline-transformed

---

Training

python train_bfn_twisted.py --exp_name ${EXP_NAME} --revision ${REVISION} --config_file configs/crossdock_train_test.yaml  --time_decoupled

where the default values should be set the same as:

python train_bfn_twisted.py --sigma1_coord 0.05 --beta1 1.5 --beta1_bond 1.5 --lr 5e-4 --time_emb_dim 0  --self_condition --epochs 30 --batch_size 16 --max_grad_norm Q --scheduler plateau --destination_prediction True --use_discrete_t True --num_samples 10 --sampling_strategy end_back_pmf --sample_num_atoms ref --ligand_atom_mode add_aromatic

Debugging

python train_bfn_twisted.py --no_wandb --debug --epochs 1

Sampling

We provide the pretrained MolPilot checkpoint here.

Sampling for pockets in the testset

To sample for CrossDock, set the CONFIG to configs/crossdock_train_test.yaml. For PoseBusters, set it to configs/posebusters_test.yaml.

# Sample with time scheduler
python train_bfn_twisted.py --config_file ${CONFIG} --ckpt_path ${CKPT_PATH} --time_scheduler_path ${TIME_SCHEDULER} --test_only --exp_name ${EXP_NAME} --revision ${REVISION} --num_samples ${NUM_MOLS_PER_POCKET} --sample_steps 100 --eval_batch_size ${BATCH_SIZE}

Sampling from pdb file

To sample from a whole protein pdb file, we need the corresponding reference ligand to clip the protein pocket (a 10A region around the reference position).

python sample_for_pocket.py --protein_path ${PDB_PATH} --ligand_path ${SDF_PATH} --time_scheduler_path ${TIME_SCHEDULER} --num_samples ${NUM_MOLS_PER_POCKET}

Evaluation

Evaluating meta files

We provide our samples as molpilot_ref_vina_docked.pt on CrossDock in the sample Google Drive folder.

Citation

@article{qiu2025piloting,
  title={Piloting Structure-Based Drug Design via Modality-Specific Optimal Schedule},
  author={Qiu, Keyue and Song, Yuxuan and Fan, Zhehuan and Liu, Peidong and Zhang, Zhe and Zheng, Mingyue and Zhou, Hao and Ma, Wei-Ying},
  journal={ICML 2025},
  year={2025}
}