AcrTransAct

Welcome to AcrTransAct, an advanced web application powered by a transformer-based Deep Neural Network designed to predict the likelihood of Acr-mediated CRISPR-Cas inhibition. Our model consists of two main components:
1. Feature Extraction Module: This module incorporates a pre-trained Evolutionary Scale Modeling (ESM) protein transformer and the NetSurfP-3.0 secondary structure prediction system. This module efficiently extracts meaningful features from input protein sequences, capturing essential information such as secondary structure.
2. Classification Module: Comprising either a CNN or LSTM network, this module utilizes the extracted features from the previous module to classify whether a given Acr protein is likely to inhibit the input CRISPR-Cas system.

Our model has been meticulously trained on an inhibition dataset of 227 experiments sourced from two Acr databases, AcrHub and Anti-CRISPRdb, and various published works. The dataset encompasses verified interactions of Acr with CRISPR-Cas systems for subtypes I-C, I-E, and I-F.
When you input a putative Acr protein sequence, it undergoes processing through the feature extraction module. The ESM protein transformer analyzes the sequence and derives a set of essential features, and the NetSurfp-3.0 network predicts Q3, ASA, RSA, and disorder for each residue. These extracted features are then passed to the classification module, where they are processed by CNN/LSTM layers and passed to fully connected layers. The classification module generates a probability of the input Acr sequence inhibiting the input CRISPR-Cas system.

To ensure the model's reliability and effectiveness, we have meticulously divided our data into training, validation, and test sets. We perform a thorough hyperparameter search on the training-validation set to optimize model performance, and then rigorously test the top models on the unseen test set. Our results showcase remarkable accuracy and F1 scores ranging from 0.93 to 0.95, reflecting the exceptional predictive capabilities of AcrTransAct.

Model Architecture:

Our code and data are available at Github.com/USask-BINFO/AcrTransAct

Cite our paper:

                    @inproceedings{hasani2023acrtransact,
                        author = {Moein Hasani and Chantel N. Trost and Nolen Timmerman and Lingling Jin},
                        title = {AcrTransAct: Pre-trained Protein Transformer Models for the Detection of Type I Anti-CRISPR Activities},
                        booktitle = {Proceedings of The 14th ACM Conference on Bioinformatics, Computational Biology, and Health Informatics (ACM-BCB)},
                        year = {2023},
                        publisher = {ACM},
                        address = {Houston, TX, USA},
                        pages = {6},
                    }