Background: Deep learning’s automatic feature extraction has proven to give superior performance in many sequence classification tasks. However, deep learning models generally require a massive amount of data to train, which in the case of Hemolytic Activity Prediction of Antimicrobial Peptides creates a challenge due to the small amount of available data. Results: Three different datasets for hemolysis activity prediction of therapeutic and antimicrobial peptides are gathered and the AMPDeep pipeline is implemented for each. The result demonstrate that AMPDeep outperforms the previous works on all three datasets, including works that use physicochemical features to represent the peptides or those who solely rely on the sequence and use deep learning to learn representation for the peptides. Moreover, a combined dataset is introduced for hemolytic activity prediction to address the problem of sequence similarity in this domain. AMPDeep fine-tunes a large transformer based model on a small amount of peptides and successfully leverages the patterns learned from other protein and peptide databases to assist hemolysis activity prediction modeling. Conclusions: In this work transfer learning is leveraged to overcome the challenge of small data and a deep learning based model is successfully adopted for hemolysis activity classification of antimicrobial peptides. This model is first initialized as a protein language model which is pre-trained on masked amino acid prediction on many unlabeled protein sequences in a self-supervised manner. Having done so, the model is fine-tuned on an aggregated dataset of labeled peptides in a supervised manner to predict secretion. Through transfer learning, hyper-parameter optimization and selective fine-tuning, AMPDeep is able to achieve state-of-the-art performance on three hemolysis datasets using only the sequence of the peptides. This work assists the adoption of large sequence-based models for peptide classification and modeling tasks in a practical manner.

Antimalarial drugs are becoming less effective due to the emergence of drug resistance. Resistance has been reported for all available malaria drugs, including artemisinin, thus creating a perpetual need for alternative drug candidates. The traditional drug discovery approach of high throughput screening (HTS) of large compound libraries for identification of new drug leads is time-consuming and resource intensive. While virtual in silico screening is a solution to this problem, however, the generalization of the models is not ideal. Artificial intelligence (AI), utilizing either structure-based or ligand-based approaches, has demonstrated highly accurate performances in the field of chemical property prediction. Leveraging the existing data, AI would be a suitable alternative to blind-search HTS or fingerprint-based virtual screening. The AI model would learn patterns within the data and help to search for hit compounds efficiently. In this work, we introduce DeepMalaria, a deep-learning based process capable of predicting the anti-Plasmodium falciparum inhibitory properties of compounds using their SMILES. A graph-based model is trained on 13,446 publicly available antiplasmodial hit compounds from GlaxoSmithKline (GSK) dataset that are currently being used to find novel drug candidates for malaria. We validated this model by predicting hit compounds from a macrocyclic compound library and already approved drugs that are used for repurposing. We have chosen macrocyclic compounds as these ligand-binding structures are underexplored in malaria drug discovery. The in silico pipeline for this process also consists of additional validation of an in-house independent dataset consisting mostly of natural product compounds. Transfer learning from a large dataset was leveraged to improve the performance of the deep learning model. To validate the DeepMalaria generated hits, we used a commonly used SYBR Green I fluorescence assay based phenotypic screening. DeepMalaria was able to detect all the compounds with nanomolar activity and 87.5% of the compounds with greater than 50% inhibition. Further experiments to reveal the compounds’ mechanism of action have shown that not only does one of the hit compounds, DC-9237, inhibits all asexual stages of Plasmodium falciparum, but is a fast-acting compound which makes it a strong candidate for further optimization.

Biometric verification systems have become prevalent in the modern world with the wide usage of smartphones. These systems heavily rely on storing the sensitive biometric data on the cloud. Due to the fact that biometric data like fingerprint and iris cannot be changed, storing them on the cloud creates vulnerability and can potentially have catastrophic consequences if these data are leaked. In the recent years, in order to preserve the privacy of the users, homomorphic encryption has been used to enable computation on the encrypted data and to eliminate the need for decryption. This work presents DeepZeroID: a privacy-preserving cloud-based and multiple-party biometric verification system that uses homomorphic encryption. Via transfer learning, training on sensitive biometric data is eliminated and one pre-trained deep neural network is used as feature extractor. By developing an exhaustive search algorithm, this feature extractor is applied on the tasks of biometric verification and liveness detection. By eliminating the need for training on and decrypting the sensitive biometric data, this system preserves privacy, requires zero knowledge of the sensitive data distribution, and is highly scalable. Our experimental results show that DeepZeroID can deliver 95.47% F1 score in the verification of combined iris and fingerprint feature vectors with zero true positives and with a 100% accuracy in liveness detection.

Generative adversarial networks have been able to generate striking results in various domains. This generation capability can be general while the networks gain deep understanding regarding the data distribution. In many domains, this data distribution consists of anomalies and normal data, with the anomalies commonly occurring relatively less, creating datasets that are imbalanced. The capabilities that generative adversarial networks offer can be leveraged to examine these anomalies and help alleviate the challenge that imbalanced datasets propose via creating synthetic anomalies. This anomaly generation can be specifically beneficial in domains that have costly data creation processes as well as inherently imbalanced datasets. One of the domains that fits this description is the host-based intrusion detection domain. In this work, ADFA-LD dataset is chosen as the dataset of interest containing system calls of small foot-print next generation attacks. The data is first converted into images, and then a Cycle-GAN is used to create images of anomalous data from images of normal data. The generated data is combined with the original dataset and is used to train a model to detect anomalies. By doing so, it is shown that the classification results are improved, with the AUC rising from 0.55 to 0.71, and the anomaly detection rate rising from 17.07% to 80.49%. The results are also compared to SMOTE, showing the potential presented by generative adversarial networks in anomaly generation.

The advancements in the Internet has enabled connecting more devices into this technology every day. The emergence of the Internet of Things has aggregated this growth. Lack of security in an IoT world makes these devices hot targets for cyber criminals to perform their malicious actions. One of these actions is the Botnet attack, which is one of the main destructive threats that has been evolving since 2003 into different forms. This attack is a serious threat to the security and privacy of information. Its scalability, structure, strength, and strategy are also under successive development, and that it has survived for decades. A bot is defined as a software application that executes a number of automated tasks (simple but structurally repetitive) over the Internet. Several bots make a botnet that infects a number of devices and communicates with their controller called the botmaster to get their instructions. A botnet executes tasks with a rate that would be impossible to be done by a human being. Nowadays, the activities of bots are concealed in between the normal web flows and occupy more than half of all web traffic. The largest use of bots is in web spidering (web crawler), in which an automated script fetches, analyzes, and files information from web servers. They also contribute to other attacks, such as distributed denial of service (DDoS), SPAM, identity theft, phishing, and espionage. A number of botnet detection techniques have been proposed, such as honeynet-based and Intrusion Detection System (IDS)-based. These techniques are not effective anymore due to the constant update of the bots and their evasion mechanisms. Recently, botnet detection techniques based upon machine/deep learning have been proposed that are more capable in comparison to their previously mentioned counterparts. In this work, we propose a deep learning-based engine for botnet detection to be utilized in the IoT and the wearable devices. In this system, the normal and botnet network traffic data are transformed into image before being given into a deep convolutional neural network, named DenseNet with and without considering transfer learning. The system is implemented using Python programming language and the CTU-13 Dataset is used for evaluation in one study. According to our simulation results, using transfer learning can improve the accuracy from 33.41% up to 99.98%. In addition, two other classifiers of Support Vector Machine (SVM) and logistic regression have been used. They showed an accuracy of 83.15% and 78.56%, respectively. In another study, we evaluate our system by an in-house live normal dataset and a solely botnet dataset. Similarly, the system performed very well in data classification in these studies. To examine the capability of our system for real-time applications, we measure the system training and testing times. According to our examination, it takes 0.004868 milliseconds to process each packet from the network traffic data during testing.

The highway network is an economically necessary form of transportation that has a significant impact on the quality of the life of the citizens who use it. Cost overruns in highway projects have been a universal occurrence that jeopardize the development, maintenance, and expansion of this vital infrastructure. Incorrect cost estimations can drive decision makers to pass ineffective policies that have played a large role in the cost overruns of transportation construction projects. The existing prediction models in the literature are limited in one or multiple areas of modeling approach, inputs, and model development robustness. In this research, a model was developed to accurately predict the total construction cost of highway projects by utilizing machine learning algorithms. This study developed a modeling pipeline to automate much of the cost forecasting process, reducing the amount of manual work and dependence on skilled data scientists. This study used the Florida Department of Transportation’s (FDOT’s) critical highway construction cost items between 2001 and 2017 to test the model. The highways of Florida were selected for testing due to the states’ population growth, high immigrant population, logistics, and hurricane frequency. This study used a pool of five categories of independent variables (69 variables total), including the construction market, energy market, socioeconomics, US economy, and temporal variables, which were compiled from relevant sources and existing literature. The results revealed that our linear model exhibits superiority in generalization and prediction of cost items over nonlinear models and is capable of accurately forecasting highway construction costs. Our suggested approach in this study also provides more accurate forecasts for the detailed cost estimation by considering the monthly historical information for the average 92.6% of the six highway construction types mentioned with a 92.51% prediction accuracy. By employing our developed model, local governments, network operators, contractors, and logistics sectors would be capable of a more exact prediction of highway construction costs.