Protein Function Prediction From Amino Acids Sequence

Proteins are large and complex molecules that play a vital role in almost every function of our body. From contracting muscles to converting food into energy or protecting us from viruses and bacterias.

Understanding the activity of discovered proteins can help to understand the mechanisms of diseases and guide lab experiments reducing tests, time and costs. At the moment, the most reliable procedure to discover protein's function is through manual experiments. However, this approach can't keep the peace with the technologies used to sequencing proteins. There is a need for reliable and fast automated annotators.

In this project, I'm going to develop a model that given a new protein in its amino acids sequence form,it can predict its functions.

• A more detailed project description is available Here
• A web application running the deployed model is accessible Here

Getting Started

This project has been developed using python 3.6 and tensorflow 2.0 (beta). In order to have the project running on your machine you have to:

1. Setup virtual environment and install the dependencies running:

python install -r requirements.txt

2. Download the UniProtdataset in XML format from Here and put it inside extract folder.
3. Run the script extract/parsingXML.py which generates proteins_goes and proteins_seqs files.

Project Details

• Run python main.py --mode=createdata to generate the TFRecords for training and testing.
• Run python main.py --mode=train to train the model defined in model/model.py
• Run python main.py --mode=evaluate to test a model stored in evaluate/models/

Extract Folder

Here the script to parse the UniProt dataset in XML format into two files called proteins_goes and proteins_seqs.

Prepare Folder

Here the jupyter notebook to analyze the dataset and create the hyperparams to filter the examples in the dataset. While createDataset.py generates the TFRecords needed to train the model in tensorflow.

Evaluate Folder

Here the scripts used to evaluate the model. custom_metrics.py defines the metrics used by evaluator.py to assess the model's performance. While the jupyter notebook results_analysis is used to generate the graphs displayed in Results section.

Results

The model that I come up with it can reliably predict the 1918 most common functions. More precisely, the model achieved a score (protein-centric metric) of F1=0.87, precision=0.84 and recall=0.92 on 140.000 unseen amino acid sequences. In simple words, the method developed predicts 92% of all functions that the protein actually have. At the same time, only 16% of the predicted functions are wrong meaning that the protein doesn't have that function.

Moreover, the method seems to be robust at the variation of prediction threshold.

With the best performance achieved with a threshold of 0.3. Similar it is the performance when the method is evaluated for each GO class independently.

Finally, to assess how well each GO term is predicted, F1 max, recall, and precision are calculated for each label. Some labels are available only in a few examples, so they have been plotted in relation to the number of times that they appear.