train Mgnify
then finetune on uniprot SPROT (well curated)
then predict small molecules
skin microbiome interesting, in general outer/ inner surfaces
# Install dependencies
sh install.sh
# Get data
osf -p pefs7 clonefor i in 1 0.5 0.25 0.125 0.0625 0.03125
do
python preprocess.py --seq uniref50.fasta.gz --out uniref50.${i}.dayhoff.txt -p ${i} --skip-header --maxlen 2000 --excluded-aa XBZJ
done
# for p=1 -- uniref50.clean.dayhoff.txt
sort -R -o uniref50.clean.dayhoff.shuffle.txt uniref50.clean.dayhoff.txt
# wc -l yields 39,222,662 sequences
head -n 35000000 uniref50.clean.dayhoff.shuffle.txt > uniref50.clean.dayhoff.train.lm.txt
head -n 37000000 uniref50.clean.dayhoff.shuffle.txt | tail -n 2000000 > uniref50.clean.dayhoff.dev.lm.txt
tail -n 2222662 uniref50.clean.dayhoff.shuffle.txt > uniref50.clean.dayhoff.test.lm.txt
# 35000000 + 2000000 + 2222662 - 39222662 = 0
head -n1000000 uniref50.clean.dayhoff.shuffle.txt > lm_redux/train.lm.txt
head -n1100000 uniref50.clean.dayhoff.shuffle.txt > lm_redux/tmp
tail -n100000 lm_redux/tmp > lm_redux/dev.lm.txt
head -n1200000 uniref50.clean.dayhoff.shuffle.txt > lm_redux/tmp
tail -n100000 lm_redux/tmp > lm_redux/test.lm.txt
rm lm_redux/tmpfrom tokenizers import CharBPETokenizer
from picotext.utils import train_tokenizer
files = ['uniref50.0p0625.dayhoff.txt']
tokenizer = train_tokenizer(
CharBPETokenizer(bert_normalizer=False),
files,
vocab_size=30000, min_frequency=5, special_tokens=['<unk>', '<pad>'])
tokenizer.save('.', 'uniref50.0p0625.dayhoff.vocab30k.freq5')# TODO: https://docs.floydhub.com/guides/jobs/installing_dependencies/
# Allow training for 12 h ie 43200 seconds
DATAVERSION=11
floyd run --gpu --data phiweger/datasets/lm_redux/${DATAVERSION}:data --mode job --env pytorch-1.4 --message "lm redux" --max-runtime 43200 --follow "\
pip install --upgrade pip && \
pip install screed tqdm tokenizers==0.7.0 && \
git clone https://github.com/phiweger/picotext && \
pip install git+https://github.com/phiweger/picotext && \
python /floyd/home/picotext/picotext/models/RNN_LM/main_lm.py --config /floyd/home/picotext/picotext/models/RNN_LM/config.json"Use other pretrained sequences, e.g. "Unirep" as coded for JAX here -- supposed to be much faster and Tensorflow independent.
conda create -y -n unirep python=3.7 ipython
conda activate unirep
# https://github.com/google/jax#installation
pip install --upgrade pip
pip install --upgrade jax jaxlib # CPU-only version
# https://github.com/ElArkk/jax-unirep
pip install tqdm optuna sklearn
pip install git+https://github.com/ElArkk/jax-unirep.gitfrom jax_unirep import get_reps
sequence = "ASDFGHJKL"
# h_avg is the canonical "reps"
h_avg, h_final, c_final = get_reps(sequence)
h_avg.shape
# (1, 1900)load data:
- json
- script: fasta > json
API:
toxenize(seq, fn=[dayhoff, BPE, ...])
# Will tokenize seq in this orderSuccess metrics:
- accuracy in the prediction of the next token
- classification accuracy of helices and other known folds
https://academic.oup.com/peds/article/16/11/799/1508626
We could collect this as secondary structure annotation from the PDB. Browse annotation. Maybe mainly alpha vs. mainly beta fold annotation as a good start to classification. Or something specific lika a B-box zinc-binding domain
AMPs:
- APD3: the antimicrobial peptide database, 3185 proteins,
wget http://aps.unmc.edu/AP/APD_AMPs_fasta_DongChuan.fa - DBAASP, > 15k
- CAMPR3, > 8k
- dbAMP
Antimicrobial peptides (AMPs), naturally encoded from genes and generally contained 10–100 amino acids, are crucial components of the innate immune system and can protect the host from various pathogenic bacteria, as well as viruses.
At present, the identified structural data is far from meeting the needs of researchers. In addition, due to the rapid development of computer technology, the newly identified sequences are exploding every year. Face this situation, structural prediction may be an appropriate way to bridge this gap26,27,28. Currently predicted structures or methods have been added to certain databases, such as: DBAASP29, AVPdb19, ParaPep20, SATPdb23. The 3D structure is generally predicted by homology modeling, which is mainly base on high homology between sequences30. Homology-modeled structures can be optimized by molecular dynamics and used to visualize the interaction between AMP and biomembrane31,32,33. In this paper, we used MOE2016 (https://www.chemcomp.com/) to establish an initial molecular structure by homology modeling, and Amber14 (http://ambermd.org/) to optimize the molecular structure by molecular dynamics, as previously reported34. In the second version of DRAMP, 82 predicted structures were added and these structures can be downloaded in pdb format (Fig. 3).
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent