Latent Bayesian Optimization via Autoregressive Normalizing Flows

Official PyTorch implementation of the "Latent Bayesian Optimization via Autoregressive Normalizing Flows". (ICLR 2025)

Installation

• Python 3.7

pip install -r requirements.yaml

We have included a zipped file of the pretrained SeqFlow's checkpoint on 1.27M molecules from the Guacamol dataset. To unzip the file and prepare it for use, please follow these steps in your terminal:

zip -s 0 ./model_weight/checkpoint.zip --out ./model_weight/checkpoint_all.zip
unzip ./model_weight/checkpoint_all.zip -d ./model_weight/

How to Run

Below are the commands for running the optimizer with different oracle budgets:

• With an Oracle Budget of 10,000

python -m BO.optimizer objective=guacamol objective.task_id=$task_id gpu=$gpu_num wandb=false generative_model=seqflow max_n_oracle_calls=10000 data.oracle_data_load_num=10000 acq_bsz=100 acq_topk=10 use_pretrain=true pretrain_name=1.27M_3 temperature=0.01 dacs_temp=400 train_oracle=200

• With an Oracle Budget of 70,000

python -m BO.optimizer objective=guacamol objective.task_id=$task_id gpu=$gpu_num wandb=false generative_model=seqflow max_n_oracle_calls=70000 data.oracle_data_load_num=10000 acq_bsz=100 acq_topk=10 use_pretrain=true pretrain_name=1.27M_3 temperature=0.01 dacs_temp=400 train_oracle=200

• With an Oracle Budget of 500

python -m BO.optimizer objective=guacamol objective.task_id=$task_id gpu=$gpu_num wandb=false generative_model=seqflow max_n_oracle_calls=500 data.oracle_data_load_num=100 acq_bsz=10 acq_topk=5 use_pretrain=true pretrain_name=1.27M_3 temperature=0.01 dacs_temp=400 train_oracle=10

Configuration Parameters

• gpu: GPU number to use.
• objective.task_id: Task identifier, corresponding to the tasks listed below.
• train_oracle: Number of oracle calls after which the generative model is trained.
• acq_topk: Number of query points $N_q$ for each trust region
• acq_bsz: Batch size of anchor points for each iteration.
• generative_model.noise: Standard deviation of the noise in the variational distribution q.
• generative_model.sim_coef: Coefficient for the similarity loss.

Guacamol Tasks

The following table lists the task id and their corresponding task names:

task_id	Full Task Name
adip	Amlodipine MPO
med2	Median molecules 2
osmb	Osimertinib MPO
pdop	Perindopril MPO
rano	Ranolazine MPO
zale	Zaleplon MPO
valt	Valsartan Smarts

Contact

If you have any questions, please create an issue on this repository or contact at [email protected].

Citation

@inproceedings{lee2025latent,
  title={Latent Bayesian Optimization via Autoregressive Normalizing Flows},
  author={Lee, Seunghun and Park, Jinyoung and Chu, Jaewon and Yoon, Minseo and Kim, Hyunwoo J},
  booktitle={The Thirteenth International Conference on Learning Representations},
  year={2025}
}