MORSE: Agentic Multi-objective Molecular Optimization via Dynamic Routing of Property-specific Editor Networks

AI4Sci Korea 2026 — Extended Abstract
September 28–October 1, 2026 · Seoul Dragon City, Seoul, Republic of Korea

Getting Started

1. Clone the Repository

Property-specific Editor checkpoints are distributed through Git LFS. Install Git LFS before cloning the repository.

On Ubuntu or Debian:

sudo apt update
sudo apt install git-lfs

Alternatively, install Git LFS from conda-forge in an active Conda environment:

conda install -c conda-forge git-lfs

Initialize Git LFS and clone MORSE:

git lfs install
git clone https://github.com/jjjabcd/MORSE.git
cd MORSE

2. Environment Setup

Create a Python 3.11 Conda environment and install MORSE with all property scorers:

conda create -n morse python=3.11 -y
conda activate morse
pip install -e .

3. Download Checkpoints and Models

Download the Editor checkpoints from the project root:

git lfs pull --include="ckpts/editor/**/*.pt"

The released Router checkpoints are regular Git files and do not require an additional download.

Download the DRD2 classifier:

bash scripts/fetch_drd2_model.sh

Download the LLM state encoder from Qwen/Qwen2.5-7B-Instruct:

python scripts/fetch_encoder.py \
  --model Qwen/Qwen2.5-7B-Instruct \
  --cache-dir ckpts/llm

The BBBP scorer uses ADMET-AI and may download its model automatically on first use.

Checkpoint structure

ckpts/
├── editor/
│   ├── foundation/
│   │   ├── foundation_final.pt
│   │   ├── configs.csv
│   │   └── char2idx.csv
│   └── adapters/
│       ├── qed_lora.pt
│       ├── drd2_lora.pt
│       ├── plogp_lora.pt
│       └── bbbp_lora.pt
├── router/
│   ├── bdp/best.pt
│   ├── bdq/best.pt
│   └── bpq/best.pt
└── llm/                       # Local Hugging Face cache

4. Training

MORSE trains one Router for each objective combination. Run one of the following scripts:

# BBBP + DRD2 + penalized logP
bash scripts/bdp_train.sh

# BBBP + DRD2 + QED
bash scripts/bdq_train.sh

# BBBP + penalized logP + QED
bash scripts/bpq_train.sh

Training arguments

Arguments supplied after a training script override the corresponding values in the task config.

Argument	Value	Description
--data_dir	Directory path	Router training dataset directory
--checkpoints	Directory path	Editor checkpoint root
--encoder_model	Hugging Face model ID or local path	Frozen LLM state encoder
--llm_cache_dir	Directory path	LLM download/cache directory
--total_episodes	Positive integer	Number of training episodes
--batch_episodes	Positive integer	Episodes collected per training step
--max_length	Positive integer	Maximum encoder token length
--max_timesteps	Positive integer	Maximum Router steps per episode
--K	Positive integer	Number of Editor candidates when Editor beam search is disabled
--editor_num_beams	Positive integer	Editor decoding beam width; 1 uses sampling
--min_similarity	Float in [0, 1]	Similarity threshold used by the reward
--gamma	Float in [0, 1]	Reward discount factor
--lr	Positive float	Router learning rate
--grad_clip	Positive float	Gradient clipping norm
--value_coef	Non-negative float	Value loss coefficient
--entropy_coef	Non-negative float	Entropy bonus coefficient
--dropout	Float in [0, 1)	Router dropout probability
--policy_arch	no_numeric or original	Router policy input architecture
--ppo_epochs	Positive integer	PPO update epochs per collected batch
--clip_eps	Positive float	PPO clipping epsilon
--ppo_buffer_size	Positive integer	PPO replay-buffer capacity in training steps
--seed	Integer	Random seed
--router_device	cuda:N, cpu, or auto	LLM encoder and Router device
--qed_editor	cuda:N or cpu	QED Editor device
--drd2_editor	cuda:N or cpu	DRD2 Editor device
--plogp_editor	cuda:N or cpu	Penalized-logP Editor device
--bbbp_editor	cuda:N or cpu	BBBP Editor device
--bbbp_device	GPU index such as 0, or cpu	BBBP scorer device
--resume	Path to last.pt	Resume an interrupted training run
--save_trajectories	Flag; no value	Save timestep-level training trajectories
--wandb	Flag; no value	Enable Weights & Biases logging

The root entry point can also be used directly:

python train.py --config config/bdp.json

5. Evaluation

Evaluate the released Router checkpoint using sampled routing or greedy routing.

Sampled routing (the default in the released configs):

bash scripts/bdp_eval.sh
bash scripts/bdq_eval.sh
bash scripts/bpq_eval.sh

Greedy routing:

bash scripts/bdq_eval.sh --greedy

When --input is omitted, the evaluation script automatically evaluates the matching seen and unseen splits in data/TEST_multi_prop/.

The root entry point can also be used directly:

python evaluate.py \
  --config config/bdq.json \
  --checkpoint ckpts/router/bdq/best.pt

Output Directory Structure

Router training artifacts are saved under:

outputs/router/{TASK}/{ENCODER_NAME}/{TRAJECTORY_MODE}/

For the released configuration, the directory has the following structure:

outputs/router/bdp/Qwen2.5-7B-Instruct/none/
├── best.pt                    # Best Router policy
├── last.pt                    # Checkpoint for resuming training
├── config.json                # Resolved run configuration
├── train_history.csv          # Training metrics
└── eval/
    ├── sampled/               # Categorical sampling results
    └── greedy/                # Greedy routing results

Contact

For questions or issues, please contact:

• Email: rlawlsgurjh@gmail.com