FoldBack: Protein Folding Code-guided Specific Peptide Design through 3D Dynamic Feedback

Peptides often exert therapeutic effects over diseases by binding to target proteins. Protein language model (PLM) is a promising perspective for peptide drug design. However, most of PLMs are difficult to consider the specificity of peptide to given target proteins. On the other hand, PLMs often produce peptides with unrealistic 3D molecular structures. Therefore, this work proposes FoldBack that is a unified framework to design peptides with target specificity and practical structure. To be specific, the sequence and structure of targets are treated as context of peptide generation process. Based on this intuition, we developed multimodal Adapter tuning to enable PLMs to generate target-aware peptides. In addition, 3D structure of peptides is evaluated and dynamically fed back to the multimodal Adapter by reinforcement learning technologies, thus preventing the formation of unusual structures. Finally, a systematic assessment pipeline is established, which integrates the peptide-target affinity, structure quality, physicochemical properties, diversity and novelty analysis. Results suggest that FoldBack achieves SOTA performance in various evaluation criteria comparing with baselines. Overall, FoldBack provides an effective framework for targeted peptide drug discovery.

1.Environment Setup

For the FoldBack model, please refer to the dependencies specified in the environment.yml file.
Since ESMC has higher Python version requirements, an additional environment is needed. Please ensure you use a newer Python version (e.g., Python 3.12 or above) to properly obtain sequence embeddings.

Installation of ESMFold can be found in the official documentation.

The FoldBack model weights can be downloaded from this link.

2.Example Workflow

First, structural features of the receptor are extracted using get_receptor_emb.py from the ProteinMPNN module in the receptor_processing directory.

For example:

python get_receptor_emb.py \
  --pdb-path ../../example/receptor.pdb \
  --out-path ../../example/receptor_structure_emb.pkl

Then, sequence features of the receptor are extracted using get_receptor_emb.py from the ESMC module located in the receptor_processing directory.

For example:

python get_receptor_emb.py \
  --fasta-path ../../example/receptor.fasta \
  --out-path ../../example

These representations are subsequently used to guide the generation of compatible peptide sequences.

You can view the full list of command-line arguments and their descriptions for any Python script by running it with the --help flag.

For example:

usage: get_receptor_emb.py [-h] [--fasta-path FASTA_PATH] [--out-path OUT_PATH]

options:
  -h, --help            show this help message and exit
  --fasta-path FASTA_PATH
                        Path to the fasta file which seq features will be extracted.
  --out-path OUT_PATH   Path to the output PKL file where the extracted features will be saved.

Second, you can generate peptide sequences conditioned on a precomputed receptor feature file (.pkl). You are free to specify the peptide length and the number of peptide sequences . The implementation details can be found in the sample.py file .

For example:

python sample.py \
  --out-path example/output/peptide.fasta \
  --ckpt-path checkpoint \
  --receptor-seq-path example/receptor_seq_emb.pkl \
  --receptor-structure-path example/receptor_structure_emb.pkl \
  --length 15 \
  --temperature 1.0 \
  --num 10

This will display the following help message:

usage: sample.py [-h] [--out-path OUT_PATH] [--ckpt-path CKPT_PATH] [--receptor-seq-path RECEPTOR_SEQ_PATH] [--receptor-structure-path RECEPTOR_STRUCTURE_PATH] [--length LENGTH]
                 [--temperature TEMPERATURE] [--num NUM]

options:
  -h, --help            show this help message and exit
  --out-path OUT_PATH   Path to the output FASTA file where the generated protein sequence will be saved.
  --ckpt-path CKPT_PATH
                        Path to the model weights file.
  --receptor-seq-path RECEPTOR_SEQ_PATH
                        Path to the target sequence characterization file.
  --receptor-structure-path RECEPTOR_STRUCTURE_PATH
                        Path to the target structure characterization file.
  --length LENGTH       Length of the peptide sequence to be generated.
  --temperature TEMPERATURE
                        Temperature between 0.2 and 1.0.
  --num NUM             Num of the peptide sequence to be generated.

Third, The generated sequences, stored in FASTA format, can be further processed using ESMFold to predict their corresponding 3D structures. The resulting PDB files can then be utilized for downstream analyses, such as binding affinity evaluation.

For example:

python get_pdb.py \
  --out-path example/output/pdb_dir \
  --fasta-path example/output/peptide.fasta

This will display the following help message:

usage: get_pdb.py [-h] [--out-path OUT_PATH] [--fasta-path FASTA_PATH]

optional arguments:
  -h, --help            show this help message and exit
  --out-path OUT_PATH   Path to the output PDB dir.
  --fasta-path FASTA_PATH
                        Path to the input FASTA file.

3.Train

The model training-related weights can be downloaded from this link

First, we perform adapter tuning to enable the model to learn how to generate plausible peptide sequences conditioned on the given receptor. The implementation details can be found in the train.py file under the FoldPep directory. For example:

python train.py \
  --num-steps 30001 \
  --batch-size 32 \
  --checkpoint-steps 5000 \
  --learning-rate 1e-3 \
  --min-learning-rate 5e-5 \
  --cosine-cycle-steps 30000 \
  --model ../FoldPep_progen_ckpt \
  --struct ../datasets/receptor_structure \
  --seq ../datasets/receptor_seq \
  --train ../datasets/train.csv \
  --output res_model

This will display the following help message:

usage: train.py [-h] [--num-steps N_STEPS] [--batch-size BATCH_SIZE] [--checkpoint-steps CHECKPOINT_STEPS] [--learning-rate LEARNING_RATE] [--min-learning-rate MIN_LEARNING_RATE]
                [--cosine-cycle-steps COSINE_CYCLE_STEPS] [--model MODEL_PATH] [--struct STRUCTURE_PATH] [--seq SEQ_PATH] [--train TRAIN_CSV_PATH] [--output OUTPUT_PATH]

optional arguments:
  -h, --help            show this help message and exit
  --num-steps N_STEPS   Total number of training steps.
  --batch-size BATCH_SIZE
                        Batch size for training.
  --checkpoint-steps CHECKPOINT_STEPS
                        Number of steps between saving checkpoints.
  --learning-rate LEARNING_RATE
                        Initial learning rate.
  --min-learning-rate MIN_LEARNING_RATE
                        Minimum learning rate for cosine annealing.
  --cosine-cycle-steps COSINE_CYCLE_STEPS
                        Number of steps for one cosine annealing cycle.
  --model MODEL_PATH    Path to the pretrained model checkpoint.
  --struct STRUCTURE_PATH
                        Path to receptor structure data.
  --seq SEQ_PATH        Path to receptor sequence data.
  --train TRAIN_CSV_PATH
                        Path to the training dataset (CSV file).
  --output OUTPUT_PATH  Directory to save trained models and outputs.

Second, after adapter tuning, we further apply a dynamic feedback–based reinforcement learning strategy to refine the model. This enables the model to generate peptide sequences that not only satisfy receptor-specific constraints but are also structurally plausible. The implementation details can be found in the RL_train.py file under the FoldBack directory. For example:

python RL_train.py \
  --num-steps 801 \
  --batch-size 6 \
  --checkpoint-steps 200 \
  --learning-rate 5e-4 \
  --min-learning-rate 1e-5 \
  --cosine-cycle-steps 800 \
  --model ../FoldPep/res_model/FoldPep_15000 \
  --struct ../datasets/receptor_structure \
  --seq ../datasets/receptor_seq \
  --train ../datasets/train.csv \
  --temperature 1.0 \
  --alpha 0.5 \
  --beta 2.0 \
  --gamma 0.8 \
  --output res_model

This will display the following help message:

usage: RL_train.py [-h] [--num-steps N_STEPS] [--batch-size BATCH_SIZE] [--checkpoint-steps CHECKPOINT_STEPS] [--learning-rate LEARNING_RATE]
                   [--min-learning-rate MIN_LEARNING_RATE] [--cosine-cycle-steps COSINE_CYCLE_STEPS] [--model MODEL_PATH] [--struct STRUCTURE_PATH] [--seq SEQ_PATH]
                   [--train TRAIN_CSV_PATH] [--temperature TEMPERATURE] [--alpha ALPHA] [--beta BETA] [--gamma GAMMA] [--output OUTPUT_PATH]

optional arguments:
  -h, --help            show this help message and exit
  --num-steps N_STEPS   Total number of tuning steps.
  --batch-size BATCH_SIZE
                        Batch size for training.
  --checkpoint-steps CHECKPOINT_STEPS
                        Number of steps between saving checkpoints.
  --learning-rate LEARNING_RATE
                        Initial learning rate.
  --min-learning-rate MIN_LEARNING_RATE
                        Minimum learning rate for cosine annealing.
  --cosine-cycle-steps COSINE_CYCLE_STEPS
                        Number of steps for one cosine annealing cycle.
  --model MODEL_PATH    Path to the pretrained model checkpoint.
  --struct STRUCTURE_PATH
                        Path to receptor structure data.
  --seq SEQ_PATH        Path to receptor sequence data.
  --train TRAIN_CSV_PATH
                        Path to the training dataset (CSV file).
  --temperature TEMPERATURE
                        Temperature between 0.2 and 1.0.
  --alpha ALPHA         Scaling factor 1
  --beta BETA           Scaling factor 2
  --gamma GAMMA         RL Scaling factor
  --output OUTPUT_PATH  Directory to save trained models and outputs.

Notes

• When using only the RL loss (i.e., setting α = 0.0), the model may over-optimize for high pLDDT scores while neglecting receptor-specific constraints.This can lead to mode collapse, where the model generates a large number of highly similar or even identical sequences, reducing diversity and practical usefulness.
• Accumulating logits during training may lead to excessively large gradients, which can negatively affect training stability.In addition, incorporating ESMFold inference within the training loop introduces significant computational overhead in terms of both memory consumption and runtime, further reducing training efficiency.
• Therefore, the current approach still has room for improvement. Future work may explore more efficient optimization strategies, such as PPO with clipping or Direct Preference Optimization (DPO), to stabilize training and improve efficiency, while enabling the generation of structurally plausible peptide sequences.