DPST : De Novo Peptide Sequencing with Amino-Acid-Aware Transformers - Pytorch Implementation

Dependency

• python 3.9.4
• torch 1.9.1+cu111
• cuda 11.2.1
• NCLL
• gcc 9.3.0
• pytorch_lightning 1.5.7
• pytorch3d 0.6.0
• cosine_annealing_warmup

Dataset

• DeepDenovo dataset
• SMSNet dataset

We follow the same data preprocessing pipline as PointNovo

Train DPST

• Please run main.py, while set the parameters properly. e.g. python3 main.py --local_attn --global_attn --value_attn --first 4 --ape --output f0e --fold 0 --batch 16 --gpus 4 --workers 2 --prefix ENTER_PATH_TO_YOUR_DATASET
• It takes 30m - 50m for constructing knapsack_file when running for a first time.

python main.py --help

--fold                                 #Select training fold
--epoch                                #Max training epochs
--gpus                                 #Number of GPUs to use
--acce                                 #ddp or dp
--val_interval                         #Frequency for evaluation
--profiler                             #Please select a pytorch_lightning profiler
--lr                                   #learning rate
--verbose_step                         #Frequency of updating trainning log
--weight_decay                         #Weight decay
--hidden                               #Number of hidden units
--heads                                #Number of Attention heads
--factor                               #FFD expansion factor
--layers                               #Number of layers
--batch                                #Batch Size
--workers                              #Number of dataLoader workers  
--checkpoints                          #Path to resume from checkpoints
--local_attn                           #Use local branch
--global_attn                          #Use global branch
--value_attn                           #Use confidence value aggregation
--output                               #Save director
--save_memory                          #Save gpu memory vs. speeding update calculation
--ape                                  #Use abosulte postional embedding or relative positonal embedding
--kq                                   #Knn for encoder self-attention
--kv                                   #Number of spectra used for confidence value aggregation
--down                                 #Downsampling hidden units for global branch
--first                                #Number of local branches used
--prefix                               #Prefix path to load data

Code for Related Work

• DeepDenovo: https://github.com/nh2tran/DeepNovo
• SMSNet: https://github.com/cmb-chula/SMSNet
• PointNovo: https://github.com/volpato30/DeepNovoV2
• Transformer: https://github.com/jadore801120/attention-is-all-you-need-pytorch
• Point Transformer: https://github.com/lucidrains/point-transformer-pytorch
• DGCNN: https://github.com/WangYueFt/dgcnn

Todo

• Upload pretrained checkpoints
• Upload testing script for PSM in diffirent organizations

Notice

Code is only tested on x86_64 x86_64 x86_64 GNU/Linux

License

The dataset loading is based on PointNovo's implementation: https://github.com/volpato30/DeepNovoV2

The model architecture is based transformer: https://github.com/jadore801120/attention-is-all-you-need-pytorch and https://github.com/lucidrains/vit-pytorch