BIB

We propose BIB: BIdirectional Learning for Offline Model-based Biological Sequence Design, which focuses on designing biological sequences to maximize some sequence score.

Installation

The environment of BIB can be installed by running the following commands:

pip install torch
pip install transformers

Reproducing Performances on DNA Tasks

We consider two DNA tasks: TFBind8(r) and TFBind10(r). Take the TFBind8(r) task as an example and we can run:

cd dna
python -u BIB.py --task TFBind8-Exact-v0 --task_mode oracle

This command will generate the offline dataset and perform some precomputations.

Then we can run the following command to obtain the experimental results of BIB where gamma_learn=1 activates the Adaptive-$\gamma$ module.

python -u BIB.py --task TFBind8-Exact-v0 --eta_learn 0 --gamma_learn 1  --mode BDI

Further ablation studies can verify the effectiveness of forward mapping, backward mapping and the Adaptive-$\gamma$ module.

python -u BIB.py --task TFBind8-Exact-v0 --eta_learn 0 --gamma_learn 0  --mode backward
python -u BIB.py --task TFBind8-Exact-v0 --eta_learn 0 --gamma_learn 0  --mode forward
python -u BIB.py --task TFBind8-Exact-v0 --eta_learn 0 --gamma_learn 0  --mode BDI

Last but not least, we can activate the Adaptive-$\eta$ module by setting eta_learn=1 to verify its effectiveness.

python -u BIB.py --task TFBind8-Exact-v0 --eta_learn 1 --gamma_learn 1  --mode BDI

The commands for TFBind10(r) are similar.

Reproducing Performances on Protein Tasks

We consider three protein tasks: avGFP, AAV and E4B. Take the avGFP task as an example and we can run:

cd protein
python -u BIB.py --task avGFP --task_mode oracle
python -u BIB.py --task avGFP --eta_learn 0 --gamma_learn 1  --mode BDI
python -u BIB.py --task avGFP --eta_learn 0 --gamma_learn 0  --mode backward
python -u BIB.py --task avGFP --eta_learn 0 --gamma_learn 0  --mode forward
python -u BIB.py --task avGFP --eta_learn 0 --gamma_learn 0  --mode BDI
python -u BIB.py --task avGFP --eta_learn 1 --gamma_learn 1  --mode BDI

The commands for AAV and E4B are similar.

Acknowledgements

We thank the pre-trained DNA-BERT model (https://github.com/jerryji1993/DNABERT) for our DNA tasks and the pre-trained Prot-T5 model and Prot-BERT model (https://github.com/agemagician/ProtTrans) for our protein tasks.