Master thesis code

To make my thesis project, I followed this approach:

##########################################################
0: DataPreprocessing
##########################################################
0)
Inspired by the NetSolP approch, I went to their website and went to the download-page.
From here I downloaded the collected data (netsolp-1.0.ALL.tar.gz)and extracted the content.
Since the NESG dataset have been split, I choose to go to the original source and download the original dataset.
The Psi-Biology dataset I've used the version delivered by NetSolP
1)
I then ran the notebook "0_DataCollectionAndPreprocessing".
This notebook loads the NESG and Psi-Bio datasets and cleans up the data. \ This means that it concatenates the datasets, removes duplicates and sequences with contradictorary labels.
I also removed sequences longer than 1000 AA, since I will be using ESM with a limit of 1024 AA.


##########################################################
1: Structures
##########################################################
To avoid predicting too many structures, the AlphaFold database came in handy.
It currently contains structure predictions of all UniProt proteins.
0)
All accession numbers from the AlphaFold database were downloaded from:
http://ftp.ebi.ac.uk/pub/databases/alphafold/accession_ids.txt
1)
The ID's were extracted by running the notebook "1a_format_accession_ID_file".
This produced the Accessions.txt file.
2)
All uniprot fastafiles were downloaded by running:
$ for ID in $(cat Accessions.txt); do wget https://www.uniprot.org/uniprot/$ID.fasta; done
3)
The downloaded fastafiles were concatenated into suitable sizes of databases, and databases where made by running:
$ makeblastdb -in X -parse_seqids -blastdb_version 5 -title "alphafold_DB" -dbtype prot -max_file_sz '3GB' -out X_DB
X being the name of the partition. Due to the large database size, it was not possible to run this in one single go, creating one single database
4)
To find if the the protein data in use already had AlphaFold structures, a protein blast was performed against each of the newly made databases:
$ blastp -query ../bulk.fasta -db X_DB/X_DB -outfmt 6 -qcov_hsp_perc 80 -num_threads 4 -out ../blast_output/blastp_X.txt
5)
The blast output was analyzed using the notebook "1b_blast_analysis".


##########################################################
2: Embeddings
##########################################################
0)
I used the "2a_ESMEmbeddings" notebook to make a bulk fasta file and esm embeddings
1)
I used the "2b_ProteinSolverEmbeddings" notebook to make ProteinSolver embeddings
2)
I used the approach of "2c_PCA" notebook to lower the dimensionality of the embeddings
Note: due to lacking processing power of my local machine, the PCA of the embeddings were performed on the computerome cluster at DTU
3)
Since I did not have a 100% match filter on the structures (80% query coverage, 90% ID and 80% subject coverage), some of the embeddings don't have the same dimension due to gaps in the alignment.
To overcome this, I have to add zeroes where the alignment indicates a gap.
I did so by using the notebook "2d_AddGapZeros"


##########################################################
3: PreTraining
##########################################################
0)
Clustering was perfromed using MMseqs2 with an identity threshold of 50% and alignment coverage of 80%.
this was done by the following commands:
$ mmseqs createdb bulk.fasta DB
$ mmseqs cluster DB DB_clu tmp --min-seq-id 50
$ mmseqs createtsv DB DB DB_clu DB_clu_50_id.tsv
1)
In order to explore the model performance, different learning rates and the use of sub-datasets, the scripts in "3a_hyperparameters" were run on the computerome cluster at DTU.
2)
The final model was pre-trained on all the available data as shown in the notebook "3b_MultitaskTraining".


##########################################################
4: FineTuning
##########################################################
0)
To preprocess the fine-tuning data in a similar manner to the pre-training data, the notebooks 4a-4d were run.
1)
I used the notebook "4e_FuneTuning" to fine-tune the model on the fine-tuning dataset.


##########################################################
5: Testing
##########################################################
0)
To see if the model was able to learn solely from the fine-tuning dataset, the notebook "5a_FineTuning-JainOnly" was used.
1)
I used the notebook "5b_Comparison" to test other aggregation predictors on the fine-tuning dataset, in order to compare performances.
2)
I used the notebook "5c_RFC" to test the performances of Random Forest Classifiers on the fine-tuning data \