danieltslz commited on
Commit
6b7eab1
·
verified ·
1 Parent(s): c4294de

Update data/embeddings/msa_transformer_depth1_uniprot_human/README.md

Browse files
data/embeddings/msa_transformer_depth1_uniprot_human/README.md CHANGED
@@ -3,7 +3,6 @@
3
  - Modality: Protein sequence / MSA Transformer depth-1 baseline
4
  - Algorithm: FAIR-ESM MSA Transformer layer-12 mean-pooled protein representations using one UniProt sequence per one-row MSA
5
  - Embedding source URL: https://github.com/facebookresearch/esm
6
- - Original model/code URL: https://github.com/facebookresearch/esm
7
 
8
  ## Public files
9
 
@@ -50,8 +49,6 @@ Each input was a one-row MSA containing only the reference UniProt protein seque
50
  model_checkpoint: esm_msa1b_t12_100M_UR50S
51
  input_data_type: UniProt human reference proteome FASTA
52
  raw_source: ~/data/raw_embeddings/msa_transformer_uniprot_human_v1_1/human_UP000005640_uniprot_proteome.fasta.gz
53
- processing_script: ~/scripts/process_embeddings/process_msa_transformer_uniprot_single_sequence_msa_v1_1.py
54
- environment: ~/envs/msa_transformer_embed
55
  source_download_page: https://github.com/facebookresearch/esm
56
  source_code_page: https://github.com/facebookresearch/esm
57
  source_identifier_type: UniProt accession
@@ -65,9 +62,9 @@ Each input was a one-row MSA containing only the reference UniProt protein seque
65
 
66
  ### Source-processing decisions
67
 
68
- The first UniProt REST streaming attempt produced an invalid 82-byte file containing an error message. That invalid file was deleted. The final input FASTA came from the UniProt FTP reference-proteome FASTA and was parsed with the final custom parser.
69
 
70
- The depth-1 design was intentional. A true homologous-MSA run would be biologically closer to the original purpose of MSA Transformer, but it would require generating or obtaining high-quality MSAs for roughly 19k human proteins. Daniel decided to keep this depth-1 run as the final MSA Transformer embedding for this batch and not run the true homologous-MSA version yet.
71
 
72
  Long proteins were split into non-overlapping windows of at most 1,022 amino acids. For each window, layer-12 residue representations were mean-pooled after removing special tokens. For proteins split across several windows, window embeddings were length-weighted and averaged into one final protein vector.
73
 
@@ -92,30 +89,7 @@ For this embedding:
92
 
93
  This README describes the standardized gene-level version included in the public resource. The central requirement was that every retained row must align to one gene in master_gene_table.csv and that the matrix row order must match genes.tsv exactly.
94
 
95
- This embedding is useful as a high-coverage protein-sequence baseline using the MSA Transformer architecture. In downstream analyses, compare it with ESM2 and other protein-sequence embeddings, but avoid calling it a true MSA embedding because it contains no homologous evolutionary alignment depth.
96
-
97
- ### Final QC details
98
-
99
- Status: processed_qc_passed
100
- Master rows: 20,097
101
- Source identifiers / source rows or entries: 20,652 FASTA indexed accessions; 19,472 mapped jobs attempted
102
- Final mapped genes: 19,472
103
- Master coverage: 96.8901%
104
- Embedding shape: 19,472 x 768
105
- FASTA indexed accessions: 20,652
106
- Mapped jobs attempted: 19,472
107
- Successful mapped genes: 19,472
108
- NaN count: 0
109
- Inf count: 0
110
- Zero vectors: 0
111
- Duplicated Ensembl IDs: 0
112
- Failed sequence count: 0
113
- Sequence length summary for successful genes:
114
- count: 19,472
115
- mean: 575.888044
116
- min: 12
117
- median: 428
118
- max: 34,350
119
 
120
  ### Summary of final mapping:
121
 
 
3
  - Modality: Protein sequence / MSA Transformer depth-1 baseline
4
  - Algorithm: FAIR-ESM MSA Transformer layer-12 mean-pooled protein representations using one UniProt sequence per one-row MSA
5
  - Embedding source URL: https://github.com/facebookresearch/esm
 
6
 
7
  ## Public files
8
 
 
49
  model_checkpoint: esm_msa1b_t12_100M_UR50S
50
  input_data_type: UniProt human reference proteome FASTA
51
  raw_source: ~/data/raw_embeddings/msa_transformer_uniprot_human_v1_1/human_UP000005640_uniprot_proteome.fasta.gz
 
 
52
  source_download_page: https://github.com/facebookresearch/esm
53
  source_code_page: https://github.com/facebookresearch/esm
54
  source_identifier_type: UniProt accession
 
62
 
63
  ### Source-processing decisions
64
 
65
+ The final input FASTA came from the UniProt FTP reference-proteome FASTA and was parsed with the final custom parser.
66
 
67
+ The depth-1 design was intentional. A true homologous-MSA run would be biologically closer to the original purpose of MSA Transformer, but it would require generating or obtaining high-quality MSAs for roughly 19k human proteins.
68
 
69
  Long proteins were split into non-overlapping windows of at most 1,022 amino acids. For each window, layer-12 residue representations were mean-pooled after removing special tokens. For proteins split across several windows, window embeddings were length-weighted and averaged into one final protein vector.
70
 
 
89
 
90
  This README describes the standardized gene-level version included in the public resource. The central requirement was that every retained row must align to one gene in master_gene_table.csv and that the matrix row order must match genes.tsv exactly.
91
 
92
+ This embedding is useful as a high-coverage protein-sequence baseline using the MSA Transformer architecture. In downstream analyses, compare it with ESM2 and other protein-sequence embeddings.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
93
 
94
  ### Summary of final mapping:
95