Multi-Objective-Guided Discrete Flow Matching for Controllable Biological Sequence Design
arXiv Paper: https://arxiv.org/abs/2505.07086
Designing biological sequences that satisfy multiple, often conflicting, functional and biophysical criteria remains a central challenge in biomolecule engineering. While discrete flow matching models have recently shown promise for efficient sampling in high-dimensional sequence spaces, existing approaches address only single objectives or require continuous embeddings that can distort discrete distributions. We present Multi-Objective-Guided Discrete Flow Matching (MOG-DFM), a general framework to steer any pretrained discrete-time flow matching generator toward Pareto-efficient trade-offs across multiple scalar objectives. At each sampling step, MOG-DFM computes a hybrid rank-directional score for candidate transitions and applies an adaptive hypercone filter to enforce consistent multi-objective progression. We also trained two unconditional discrete flow matching models, PepDFM for diverse peptide generation and EnhancerDFM for functional enhancer DNA generation, as base generation models for MOG-DFM. We demonstrate MOG-DFM's effectiveness in generating peptide binders optimized across five properties (hemolysis, non-fouling, solubility, half-life, and binding affinity), and in designing DNA sequences with specific enhancer classes and DNA shapes. In total, MOG-DFM proves to be a powerful tool for multi-property-guided biomolecule sequence design.
Usage
0. Conda Environment
conda create -n mog-dfm python=3.9
conda install pytorch==2.4.0 torchvision==0.19.0 torchaudio==2.4.0 pytorch-cuda=12.4 -c pytorch -c nvidia
pip install fair-esm transformers xgboost datasets torchdiffeq
To use Deep DNAshape, please create another conda environment following the guidance of its repository.
1. PepDFM and EnhancerDFM training and evaluation
The pretrained weights for PepDFM and EnhancerDFM are available in the ckpt directory.
We also provide the complete training and evaluation code for both models.
2. Multi-Objective Guided Generation
2.0 Score Models
The pretrained weights for the score models (hemolysis, non-fouling, solubility, half-life, binding affinity, and enhancer class) are available in the classifier_ckpt directory.
(The checkpoint binding_affinity_unpooled.pt will be uploaded soon!)
Prediction scripts for each score model are provided in the classifier_code directory.
2.1 Peptide Generation Task
Example command for peptide generation guided by multiple objectives (hemolysis, non-fouling, solubility, half-life, and binding affinity):
python PepDFM_multi_objective_generation.py --is_peptide True --T 100 --n_samples 5 --n_batches 10 --length 10 --target_protein GSHMIEPNVISVRLFKRKVGGLGFLVKERVSKPPVIISDLIRGGAAEQSGLIQAGDIILAVNDRPLVDLSYDSALEVLRGIASETHVVLILRGPEGFTTHLETTFTGDGTPKTIRVTQPLGPPTKAV
Note that the hemolysis model outputs one minus the actual hemolysis score, and the half-life model outputs the base-10 logarithm of the half-life in hours.
The guidance settings and their importance weights can be found and modified in PepDFM_multi_objective_generation.py
2.2 Enhancer DNA Generation Task
Example command for enhancer DNA generation guided by the enhancer class and DNA shape:
python EnhancerDFM_multi_objective_generation.py --is_peptide False --T 800 --n_samples 5 --n_batches 10 --length 100 --target_enhancer_class 0 --target_DNA_shape HelT
The guidance settings and their importance weights can be found and modified in EnhancerDFM_multi_objective_generation.py
To use this repository, you agree to abide by the MOG-DFM License.