dimeng851/embedding repository overview

⁠🫠 Protein Sequence Embedding

⁠Welcome to Protein Sequence Embedding 🧬

This project offers three types of protein sequence embedding methods:

1️⃣ Onehot: Encode your protein sequences into one-hot representations.

2️⃣ ProtTrans: Utilize ProtTrans to embed your protein sequences.

3️⃣ MSA Transformer: Employ MSA Transformer for embedding.

⁠Inputs

• One .fasta file, put the queir sequences into one FASTA file. e.g.

>DP02585
MWERLNCAAEDFYSRLLQKFNEEKKGIRKDPFLYEADVQVQLISKGQPNPLKNILNENDIVFIVEKVPLEKEETSHIEELQSEETAISDFSTGENVGPLALPVGKARQLIGLYTMAHNPNMTHLKINLPVTALPPLWVRCDSSDPEGTCWLGAELITTNNSITGIVLYVVSCKADKNYSVNLENLKNLHKKRHHLSTVTSKGFAQYELFKSSALDDTITASQTAIALDISWSPVDEILQIPPLSSTATLNIKVESGEPRGPLNHLYRELKFLLVLADGLRTGVTEWLEPLEAKSAVELVQEFLNDLNKLDGFGDSTKKDTEVETLKHDTAAVDRSVKRLFKVRSDLDFAEQLWCKMSSSVISYQDLVKCFTLIIQSLQRGDIQPWLHSGSNSLLSKLIHQSYHGTMDTVSLSGTIPVQMLLEIGLDKLKKDYISFFIGQELASLNHLEYFIAPSVDIQEQVYRVQKLHHILEILVSCMPFIKSQHELLFSLTQICIKYYKQNPLDEQHIFQLPVRPTAVKNLYQSEKPQKWRVEIYSGQKKIKTVWQLSDSSPIDHLNFHKPDFSELTLNGSLEERIFFTNMVTCSQVHFK
>DP02606
MSRQSSVSFRSGGSRSFSTASAITPSVSRTSFTSVSRSGGGGGGGFGRVSLAGACGVGGYGSRSLYNLGGSKRISISTSGGSFRNRFGAGAGGGYGFGGGAGSGFGFGGGAGGGFGLGGGAGFGGGFGGPGFPVCPPGGIQEVTVNQSLLTPLNLQIDPSIQRVRTEEREQIKTLNNKFASFIDKVRFLEQQNKVLDTKWTLLQEQGTKTVRQNLEPLFEQYINNLRRQLDSIVGERGRLDSELRNMQDLVEDFKNKYEDEINKRTTAENEFVMLKKDVDAAYMNKVELEAKVDALMDEINFMKMFFDAELSQMQTHVSDTSVVLSMDNNRNLDLDSIIAEVKAQYEEIANRSRTEAESWYQTKYEELQQTAGRHGDDLRNTKHEISEMNRMIQRLRAEIDNVKKQCANLQNAIADAEQRGELALKDARNKLAELEEALQKAKQDMARLLREYQELMNTKLALDVEIATYRKLLEGEECRLSGEGVGPVNISVVTSSVSSGYGSGSGYGGGLGGGLGGGLGGGLAGGSSGSYYSSSSGGVGLGGGLSVGGSGFSASSGRGLGVGFGSGGGSSSSVKFVSTTSSSRKSFKS
  ...

• .a3m files (for MSA transformer only)

  Generate from HHblits⁠
  [SEQUENCE_NAME/ID].a3m, replace SEQUENCE_NAME/ID with the actural sequence ID, it should be the same as the name from .fasta file.
  e.g. DP02585.a3m and DP02606.a3m

⁠Output

⁠Format

[SEQUENCE_NAME/ID].npy files.

⁠Shape

• Onehot: (1, SEQUENCE_LENGTH, 21)
• ProtTrans: (1, SEQUENCE_LENGTH, 1024)
• MSA Transformer: (1, SEQUENCE_LENGTH, 768)

⁠Getting Started:

⁠Docker

• Pull the Docker image from DockerHub⁠

  docker pull dimeng851/embedding:v3
  

  Copy

• Edit the embeeding methods in Docker file

  Default: apply all three emedding methods: 1️⃣ onehot, 2️⃣ protTrans , and 3️⃣ MSA Transformer. If you want to generate embedding from only one or two of the embedding methods
  a. open Dockerfile
  b. delete the embedding methods you don't want from

  CMD python /embedding/main.py --embeddingType onehot,protTrans,msaTrans
  

  Copy

• run Containner

  docker run -d \
  -it \
  --name CONTAINER_NAME \
  --mount type=bind,source=PATH_TO_INPUT_FASTA_FILE,target=/embedding/data/input.fasta \
  --mount type=bind,source=PATH_TO_INPUT_A3M_FOLDER,target=/embedding/data/hmm \
  --mount type=bind,source=PATH_TO_INPUT_OUTPUT_FOLDER,target=/embedding/data/output \
  --mount type=bind,source=PATH_TO_INPUT_TORCH_CHECKPOINT,target=/root/.cache/torch/hub/checkpoints/ \
  --mount type=bind,source=PATH_TO_INPUT_HUGGINGFACE_HUB,target=/root/.cache/huggingface/hub/ \
  dimeng851/embedding:v3
  

  Copy

  Please replace the following parts:
  CONTAINER_NAME with any container name you like,
  PATH_TO_INPUT_FASTA_FILE with input fasta file path,
  PATH_TO_INPUT_A3M_FOLDER with the folder to Hblits searching results, here we require .a3m files,
  PATH_TO_INPUT_OUTPUT_FOLDER with the folder you want to put the embedding sequences. Do not change the other parts,
  PATH_TO_INPUT_TORCH_CHECKPOINT with the parent folder to two MSA Transformer models (esm_msa1b_t12_100M_UR50S-contact-regression.pt & esm_msa1b_t12_100M_UR50S.pt) if you have the models downloaded. Otherwise, a folder to save the pretranined models.
  PATH_TO_INPUT_HUGGINGFACE_HUB with the parent folder to ProtTrans (models--Rostlab--prot_t5_xl_uniref50) if you have the models downloaded. Otherwise, a folder to save the pretranined model.

  Here is an example,

  docker run -d \
  -it \
  --name embed_con \
  --mount type=bind,source=/home/dimeng/caid3/test.fasta,target=/embedding/data/input.fasta \
  --mount type=bind,source=/home/dimeng/project/linker_caid/a3m,target=/embedding/data/hmm \
  --mount type=bind,source=/home/dimeng/caid3/output/embedding,target=/embedding/data/output \
  --mount type=bind,source=/home/dimeng/.cache/torch/hub/checkpoints/,target=/root/.cache/torch/hub/checkpoints/ \
  --mount type=bind,source=/home/dimeng/.cache/huggingface/hub/,target=/root/.cache/huggingface/hub/ \
  dimeng851/embedding:v3
  

  Copy

• Check the embedded results from the output folder you provided

• Here are some information about the Docker version this project used

  Client:
     Cloud integration: v1.0.35+desktop.10
     Version:           25.0.3
     API version:       1.44
     Go version:        go1.21.6
     Git commit:        4debf41
     Built:             Tue Feb  6 21:13:26 2024
     OS/Arch:           darwin/amd64
     Context:           desktop-linux
  
  Server: Docker Desktop 4.27.2 (137060)
     Engine:
      Version:          25.0.3
      API version:      1.44 (minimum version 1.24)
      Go version:       go1.21.6
      Git commit:       f417435
      Built:            Tue Feb  6 21:14:25 2024
      OS/Arch:          linux/amd64
      Experimental:     false
     containerd:
      Version:          1.6.28
      GitCommit:        ae07eda36dd25f8a1b98dfbf587313b99c0190bb
     runc:
      Version:          1.1.12
      GitCommit:        v1.1.12-0-g51d5e94
     docker-init:
      Version:          0.19.0
      GitCommit:        de40ad0
  

  Copy