Cell-PLoc 2.0[1]
There is a flowchart to show the prediction process of the predictors in Cell-PLoc 2.0, where ensemble classifier 1 is for processing the GO descriptor samples, while ensemble classifier 2 is for the FunD (functional domain) and PseEvo (pseudo sequential evolution) descriptor samples.
Team:USTC-Software/Model
Model
Biological Models
Jpred4[2]
Both secondary structure and affinity are predicted by the JNet neural network, and the neural network used in Jpred4 is a retrained version, JNet 2.3.1. The JNet 2.0 neural network-based predictor has been retrained to make JNet 2.3.1 by 7-fold cross-validation using one representative for each of the 1358 SCOPe/ASTRAL v.2.04 superfamily domain sequences. Multiple alignments for each sequence were built by PSI-BLAST through searching UniRef90 v.2014_07. In addition to retraining, the HMM building step in JNet was updated to HMMer 3 and some improvements were made to the code to simplify management and future algorithmic developments. The initial JNet model[3] is shown below.
Isoelectric Point Calculator 2.0[4]
Overview of the IPC 2.0 architecture. The input (amino acid sequence in the plain format or multiple sequences in the FASTA format) is processed by individual machine learning models. Separate models depending the prediction task are used. Isoelectric point prediction for peptides is based on separable convolution model.
DeepTMHMM[5]
Protein structure prediction using deep learning methods have seen several advancements within the last years. In this project, we investigate deep learning for prediction of the membrane topology of transmembrane proteins. Transmembrane proteins are relevant for drug development since they make up more than 50% of all human drug targets . We present TMHMM3, a deep learning model based on Long Short-Term Memory (LSTM) units and Conditional Random Fields (CRF). We explore several variations of this model such as using CRFs based on different state machines, using Marginal Probabilities for type prediction and using Connectionist Temporal Classification (CTC) as an alternative to CRFs. Finally, we compare the performance of these models with existing tools for transmembrane topology prediction. We show that TMHMM3 performs better than many of the most popular tools (such as Phobius and TMHMM2) and achieves similar performance to the current state-of-the-art (TOPCONS2). Because TMHMM3 do not rely on protein profiles it is significantly faster at making predictions than models that do use profiles, such as TOPCONS2.
Computer Models
Service Structure
1. Nginx[6] proxy and load balance
To build up a web application of protein prediction service, we choose the structure of backend as below: the Nginx servers receive messages from multiple clients, and filter the URL according to configured rules: if URL start with '/api/, Http packages are redirected to Gunicorn[7] process through the prepared Unix socket; if not, Nginx will answer the request according to file hierarchy. Gunicorn serve as a manager of threads of WSGI server – Flask[8], distribute requests to balance the load. When client post a request of specific protein for specific calculation, one of the idle Flask server comes to handle.
2. Protein digest and task submission
A post request is required for task submission, and when sequence received, we validate it and calculate it’s md5 digest to identify the protein. Then we push the digest-sequence pair to our Redis database and return the digest immediately to client.
To avoid the server thread stuck at specific calculation process, we adopt celery[9], the task queue manager. With celery, Flask thread can be devoted only to web service, and deliver the prediction part to celery workers. Each prediction tasks are running in separate threads, and push the digest – result key into Redis[10] database.
3. Result distribution
Each task will be allocated a specific database. During the submit of prediction task, a "{status: running}" will be set with digest as key in result database, when task was finished, the digest key’s value is automatically set to JSON output by celery workers. And if task raise exception, it can also put related information in the result JSON. In this way, we detach the responsibility of status and error detection from Flask thread (Flask just check if given hash is present in the database and then return the value if yes), and make it flexible for new components to add in.
Clients will request their results from “/api//” with GET method. For the sake of Redis cache, it would be possible for clients to get results immediately if the result of specific protein and specific module has been calculated. But it’s so lucky for most cases, so if the first request get a status of running, client will silent for a fixed length of time, and repeat the request for several times. And if finally the status remains running, let user decide if sending request again.
Hot-plug Extensible Module
The professional version of CAT aims to provide a platform for different deep-learning modules. So we design a hot-plug extensible structure for more modules to be directly inserted into our service. Our background structure was born with hot-plug features. To be specific, each of scripts from contributors are packed into different python modules. And each module should export a class with the same name of module, which includes the method called get_result_json that takes in required keyword arguments and returns JSON serialized results. Model contributors may also export a dict of error messages corresponding to their Exceptions.
In our service part, we aggregate modules into one folder, and import them according to a python config file. Different Redis databases will be distributed among those models, and will be paired with the name of module in dict for quick access. Then we need to asynchronize the task with only one celery “template”. It takes model function as first argument to call and preserve arguments behind to feed the model function. Firstly, function called with try catch, and then if no error, result will be dumped into Redis. But if some exception raised, the exception will be matched against the exception – message dict, and dump corresponding message to Redis.
In ideal cases, our platform will work as following: When receiving the POST request from /api/, we match the model segment in URL and after verifying the existence of model, directly call the corresponding celery async function with posted JSON as keyword arguments. Then it’s responsible for celery worker to validate the posted information. Because validation always occur at very beginning, the worker may go through a fast-fail process and insert error message to Redis before the query GET arrived, so clients may not suffer a long-time wait before their frustrating error arrives.
To cooperate with our backend, the professional version of C.A.T. frontend may contain list of foldable cards. Those cards representing different modules will submit their own requests to backend and represent the result in place. What’s more, according to the argument list provided by contributors of models, we can automatically organize the needed widget in the card, like input box, Boolean switch, drop-down selector. In this way, we realize the suppose of hot-plug use of our C.A.T. professional platform.
References
[1]"Cell-PLoc 2.0: an improved package of web-servers for predicting subcellular localization of proteins in various organisms" written by Kuo-Chen Chou, Hong-Bin Shen,published by Natural Science, Vol.2 No.10, 2010
[2]Alexey Drozdetskiy, Christian Cole, James Procter, Geoffrey J. Barton, JPred4: a protein secondary structure prediction server, Nucleic Acids Research, Volume 43, Issue W1, 1 July 2015, Pages W389–W394, https://doi.org/10.1093/nar/gkv332
[3]Application of multiple sequence alignment profiles to improve protein secondary structure prediction James A. Cuff,Geoffrey J. Barton,First published: 14 June 2000 https://doi.org/10.1002/1097-0134(20000815)40:3<502::AID-PROT170>3.0.CO;2-Q
[4]Lukasz Pawel Kozlowski, IPC 2.0: prediction of isoelectric point and pKa dissociation constants, Nucleic Acids Research, Volume 49, Issue W1, 2 July 2021, Pages W285–W292, https://doi.org/10.1093/nar/gkab295
[5]Jeppe Hallgren, Konstantinos D. Tsirigos, José Juan Almagro Armenteros, Paolo Marcatili, Henrik Nielsen, Anders Krogh and Ole Winther (2020). TMHMM3. https://biolib.com/jeppe/DeepTMHMM-mn92
[10]https://redis.io
Contact Us
Mail:
USTC_Software2021@163.com
University of Science and Technology of China, No.96, JinZhai Road Baohe District, Hefei, Anhui, 230026, P.R.China
