Dr Mahakim Newton

Protein-ligand binding affinity prediction is a key element of computer-aided drug discovery. Most of the existing deep learning methods for protein-ligand binding affinity prediction utilize single models and suffer from low accuracy and generalization capability. In this paper, we train 13 deep learning models from combinations of 5 input features. Then, we explore all possible ensembles of the trained models to find the best ensembles. Our deep learning models use cross-attention and self-attention layers to extract short and long-range interactions. Our method is named Ensemble Binding Affinity (EBA). EBA extracts information from various models using different combinations of input features, such as simple 1D sequential and structural features of the protein-ligand complexes rather than 3D complex features. EBA is implemented to accurately predict the binding affinity of a protein-ligand complex. One of our ensembles achieves the highest Pearson correlation coefficient (R) value of 0.914 and the lowest root mean square error (RMSE) value of 0.957 on the well-known benchmark test set CASF2016. Our ensembles show significant improvements of more than 15% in R-value and 19% in RMSE on both well-known benchmark CSAR-HiQ test sets over the second-best predictor named CAPLA. Furthermore, the superior performance of the ensembles across all metrics compared to existing state-of-the-art protein-ligand binding affinity prediction methods on all five benchmark test datasets demonstrates the effectiveness and robustness of our approach. Therefore, our approach to improving binding affinity prediction between proteins and ligands can contribute to improving the success rate of potential drugs and accelerate the drug development process.

Protein-ligand binding affinity plays a pivotal role in drug development, particularly in identifying potential ligands for target disease-related proteins. Accurate affinity predictions can significantly reduce both the time and cost involved in drug development. However, highly precise affinity prediction remains a research challenge. A key to improve affinity prediction is to capture interactions between proteins and ligands effectively. Existing deep-learning-based computational approaches use 3D grids, 4D tensors, molecular graphs, or proximity-based adjacency matrices, which are either resource-intensive or do not directly represent potential interactions. In this paper, we propose atomic-level distance features and attention mechanisms to capture better specific protein-ligand interactions based on donor-acceptor relations, hydrophobicity, and p-stacking atoms. We argue that distances encompass both short-range direct and long-range indirect interaction effects while attention mechanisms capture levels of interaction effects. On the very well-known CASF-2016 dataset, our proposed method, named Distance plus Attention for Affinity Prediction (DAAP), significantly outperforms existing methods by achieving Correlation Coefficient (R) 0.909, Root Mean Squared Error (RMSE) 0.987, Mean Absolute Error (MAE) 0.745, Standard Deviation (SD) 0.988, and Concordance Index (CI) 0.876. The proposed method also shows substantial improvement, around 2% to 37%, on five other benchmark datasets. The program and data are publicly available on the website https://gitlab.com/mahnewton/daap. Scientific Contribution Statement This study innovatively introducesdistance-based features to predict protein-ligand binding affinity, capitalizing onunique molecular interactions. Furthermore, the incorporation of protein sequencefeatures of specific residues enhances the model's proficiency in capturing intricatebinding patterns. The predictive capabilities are further strengthened through theuse of a deep learning architecture with attention mechanisms, and an ensembleapproach, averaging the outputs of five models, is implemented to ensure robustand reliable predictions.

Forecasting accurate water quality is very important in aquaculture, environment monitoring, and many other applications. Many internal and external factors influence water quality. Therefore, water quality parameters exhibit complex time series characteristics. Consequently, long-range accurate prediction of water quality parameters suffers from poor propagation of information from past timepoints to further future timepoints. Moreover, to synchronise the prediction model with the changes in the time series characteristics, periodic retraining of the prediction model is required and such retraining is to be done on resource-restricted computation devices. In this work, we present a low-cost training approach to improve long-range multi-step water quality prediction. We train a short-range predictor to save training effort. Then, we strive to achieve and/or improve long-range prediction using multi-step iterative ensembling during inference. Experimental results on 9 water quality datasets demonstrate that the proposed method achieves significantly lower error than the existing state-of-the-art approaches. Our approach significantly outperforms the existing approaches in several standard metrics, even in the case of future timepoints at long distances.

Amyotrophic Lateral Sclerosis (ALS) is a complex and rare neurodegenerative disorder characterized by significant genetic, molecular, and clinical heterogeneity. Despite numerous endeavors to discover the genetic factors underlying ALS, a significant number of these factors remain unknown. This knowledge gap highlights the necessity for personalized medicine approaches that can provide more comprehensive information for the purposes of diagnosis, prognosis, and treatment of ALS. This work utilizes an innovative approach by employing a machine learning-facilitated, multi-omic model to develop a more comprehensive knowledge of ALS. Through unsupervised clustering on gene expression profiles, 9,847 genes associated with ALS pathways are isolated and integrated with 7,699 genes containing rare, presumed pathogenic genomic variants, leading to a comprehensive amalgamation of 17,546 genes. Subsequently, a Variational Autoencoder is applied to distil complex biomedical information from these genes, culminating in the creation of the proposed Multi-Omics for ALS (MOALS) model, which has been designed to expose intricate genotype-phenotype interconnections within the dataset. Our meticulous investigation elucidates several pivotal ALS signaling pathways and demonstrates that MOALS is a superior model, outclassing other machine learning models based on single omic approaches such as SNV and RNA expression, enhancing accuracy by 1.7 percent and 6.2 percent, respectively. The findings of this study suggest that analyzing the relationships within biological systems can provide heuristic insights into the biological mechanisms that help to make highly accurate ALS diagnosis tools and achieve more interpretable results.

The Internet of Things (IoT) has many important applications in multiple domains that include home automation, smart cities, healthcare, agriculture, and environment. IoT comprises a wide range of sensors and actuators that communicate with each other over cloud, fog, and edge level networks. Moreover, these devices use various communication protocols and are made by different manufactures. To deal with these diversities, IoT essentially needs interoperable communication interfaces among devices. Unfortunately, existing interoperability solutions are centralized and use fog or cloud level computing resources, making IoT communications latency-prone and poorly scalable. These issues could be handled effectively, if edge level devices could be made interoperable within the edge level and without needing fog or cloud level access. This article proposes a decentralized interoperability solution that stays fully within the edge level. The solution relies on controller devices that work on the interface boundaries of the edge devices. Unlike existing solutions, the proposed solution adopts a hierarchical interoperability model to handle interoperability at network, syntactical, semantic, and organizational levels. Our solution is nonproprietary, generic over vendors and platforms, and easily extendable to new devices. We compare our proposed solution with existing interoperability solutions for edge devices and show its mobility, efficiency, and flexibility.

Protein Structure Prediction (PSP) has achieved significant progress lately. Prediction of inter-residue distances by machine learning and their exploitation during the conformational search is largely among the critical factors behind the progress. Real values than bin probabilities could more naturally represent inter-residue distances, while the latter, via spline curves more naturally helps obtain differentiable objective functions than the former. Consequently, PSP methods that exploit predicted binned distances perform better than those that exploit predicted real-valued distances. To leverage the advantage of bin probabilities in getting differentiable objective functions, in this work, we propose techniques to convert real-valued distances into distance bin probabilities. Using standard benchmark proteins, we then show that our real-to-bin converted distances help PSP methods obtain three-dimensional structures with 4%¿16% better root mean squared deviation (RMSD), template modeling score (TM-Score), and global distance test (GDT) values than existing similar PSP methods. Our proposed PSP method is named real to bin (R2B) inter-residue distance predictor, and its code is available from https://gitlab.com/mahnewton/r2b.

Protein structure prediction (PSP) is a vital challenge in bioinformatics, structural biology and drug discovery. Protein secondary structure (SS) prediction is critical since three-dimensional (3D) structures are primarily made up of secondary structures. With the advancement of deep learning approaches, SS classification accuracy has been significantly improved. Many existing methods use an ensemble of complex neural networks to improve SS prediction. Because of the high dimensionality of the hyperparameter space, deep neural networks with complex architectures are typically challenging to train effectively. Also, predicting secondary structures in the boundary regions between different types of SS is challenging. This study presents Multi-S3P, which employs bidirectional Long-Short-Term-Memory (BILSTM) and Convolutional Neural Networks (CNN) with a self-attention mechanism to improve the secondary structure prediction using an effective training strategy to capture the unique characteristics of each type of secondary structure and combine them more effectively. The ensemble of CNN and BILSTM can learn both contextual information and long-range interactions between the residues. In addition, using a self-attention mechanism allows the model to focus on the most important features for improving performance. We used the SPOT-1D dataset for the training and validation of our model using a set of four input features derived from amino acid sequences. Further, the model was tested on four popular independent test datasets and compared with various state-of-the-art predictors. The presented results show that Multi-S3P outperformed the other methods in terms of Q3, Q8 accuracy and other performance metrics, achieving the highest Q3 accuracy of 87.57% and a Q8 accuracy of 77.56% on the TEST2016 test set. More importantly, Multi-S3P demonstrates high performance in SS boundary regions. Our experiment also demonstrates that the combination of different input features and a multi-network-based training strategy significantly improved the performance.

More training instances could lead to better classification accuracy. However, accuracy could also degrade if more training instances mean further noises and outliers. Additional training instances arguably need additional computational resources in future data mining operations. Instance selection algorithms identify subsets of training instances that could desirably increase accuracy or at least do not decrease accuracy significantly. There exist many instance selection algorithms, but no single algorithm, in general, dominates the others. Moreover, existing instance selection algorithms do not allow direct controlling of the instance selection rate. In this paper, we present a simple and generic cluster-oriented instance selection algorithm for classification problems. Our proposed algorithm runs an unsupervised K Means Clustering algorithm on the training instances and with a given selection rate, selects instances from the centers and the borders of the clusters. On 24 benchmark classification problems, when very similar percentages of instances are selected by various instance selection algorithms, K Nearest Neighbours classifiers achieve more than 2%¿3% better accuracy when using instances selected by our proposed method than when using those selected by other state-of-the-art generic instance selection algorithms.

Predicted inter-residue distances are a key behind recent success in high quality protein structure prediction (PSP). However, prediction of both short and long distance values together is challenging. Consequently, predicted short distances are mostly used by existing PSP methods. In this paper, we use a stacked meta-ensemble method to combine deep learning models trained for different ranges of real-valued distances. On five benchmark sets of proteins, our proposed inter-residue distance prediction method improves mean Local Distance Different Test (LDDT) scores at least by 5% over existing such methods. Moreover, using a real-valued distance based conformational search algorithm, we also show that predicted long distances help obtain significantly better protein conformations than when only predicted short distances are used. Our method is named meta-ensemble for distance prediction (MDP) and its program is available from https://gitlab.com/mahnewton/mdp.

Real-life water quality monitoring applications such as aquaculture domains and water resource management need long range multi-step prediction for disaster control. However, prediction accuracy usually degrades gradually as the prediction target timepoint is further away from the current timepoint. To address this, recent water quality forecasting methods mostly rely on complex deep learning models. In this paper, we propose a simple time-variant iterative ensembling method that strives to significantly improve the performance of a given arbitrary long range multi-step time series predictor for water quality data with minimal increase in computational cost. With the given predictor, our proposed method iteratively uses ensembles of predicted values for preceding steps to improve the prediction accuracy for the succeeding steps. The iterative ensembling operation is performed on the trained model and only at the inference stage, and so does not need any further computing-intensive training for the performance improvement. We experimentally show that the proposed method is effective with 7 predictors and 9 water quality datasets of various types, and it outperforms the state-of-the-art results in those datasets by around 2%¿29% in mean absolute error (MAE), root mean squared error (RMSE) and mean absolute percentage error (MAPE) metrics. Similar improvement has also been found in two other metrics such as normalized Nash¿Sutcliffe model efficiency coefficient (NNSE) metric and Taylor diagram plot. Overall, the proposed iterative ensembling is a promising approach for multi-step long range water quality prediction for high-frequency water quality monitoring systems.

Protein contact maps capture coevolutionary interactions between amino acid residue pairs that are spatially within certain proximity threshold. Predicted contact maps are used in many protein related problems that include drug design, protein design, protein function prediction, and protein structure prediction. Contact map prediction has achieved significant progress lately but still further challenges remain with prediction of contacts between residues that are separated in the amino acid residue sequence by large numbers of other residues. In this paper, with experimental results on 5 standard benchmark datasets that include membrane proteins, we show that contact map prediction could be significantly enhanced by using ensembles of various state-of-the-art short distance predictors and then by converting predicted distances into contact probabilities. Our program along with its data is available from https://gitlab.com/mahnewton/ecp.

Toxicity prediction using quantitative structure-activity relationship has achieved significant progress in recent years. However, most existing machine learning methods in toxicity prediction utilize only one type of feature representation and one type of neural network, which essentially restricts their performance. Moreover, methods that use more than one type of feature representation struggle with the aggregation of information captured within the features since they use predetermined aggregation formulas. In this paper, we propose a deep learning framework for quantitative toxicity prediction using five individual base deep learning models and their own base feature representations. We then propose to adopt a meta ensemble approach using another separate deep learning model to perform aggregation of the outputs of the individual base deep learning models. We train our deep learning models in a weighted multitask fashion combining four quantitative toxicity data sets of LD50, IGC50, LC50, and LC50-DM and minimizing the root-mean-square errors. Compared to the current state-of-the-art toxicity prediction method TopTox on LD50, IGC50, and LC50-DM, that is, three out of four data sets, our method, respectively, obtains 5.46, 16.67, and 6.34% better root-mean-square errors, 6.41, 11.80, and 12.16% better mean absolute errors, and 5.21, 7.36, and 2.54% better coefficients of determination. We named our method QuantitativeTox, and our implementation is available from the GitHub repository https://github.com/Abdulk084/QuantitativeTox.

Mixed Blocking Permutation Flowshop Scheduling Problem (MBPFSP) with the objective of makespan minimisation is NP-Hard. It has important industrial applications that include the cider production industry. MBPFSP has made some progress in recent years. However, within practical time limits, existing incomplete algorithms still either find low quality solutions or struggle with large problems. One key reason behind this is the typical way of using generic heuristics or metaheuristics that usually lack problem specific structural knowledge. In MBPFSP, a machine could be blocked with the currently finished job until the subsequent machine is available to process the same job. These blocking constraints affect the makespan. So MBPFSP search should naturally take explicit steps to take the blocking constraints into account. Unfortunately, existing research on MBPFSP just uses only the makespan to compare generated solutions, but the search otherwise is not aware of the blocking constraints. Moreover, existing such methods use either an exhaustive or a random neighbourhood generation strategy. In this work, we aim to advance MBPFSP search by better exploiting the problem specific structural knowledge. We use the constraint and the objective functions to obtain such problem specific knowledge and we exploit such knowledge both in a constructive search method and in a local search method. In this paper, we also present an acceleration method to efficiently evaluate insertion-based neighbourhoods of MBPFSP. Our experimental results on three standard testbeds demonstrate that our proposed algorithms significantly improve over existing best-performing algorithms.

Aircraft sequencing problem (ASP) is an NP-Hard problem. It involves allocation of aircraft to runways for landing and takeoff, minimising total tardiness. ASP has made significant progress in recent years. However, within practical time limits, existing incomplete algorithms still either find low quality solutions or struggle with large problems. One key reason behind this is the typical way of using generic heuristics or metaheuristics that usually lack problem specific structural knowledge. As a result, existing such methods use either an exhaustive or a random neighbourhood generation strategy. So their search guidance comes only from the evaluation function that is used mainly after the neighbourhood generation. In this work, we aim to advance ASP search by better exploiting the problem specific structural knowledge. We use the constraint and the objective functions to obtain such problem specific knowledge and we exploit such knowledge both in a constructive search method and in a local search method. Our motivation comes from the constraint optimisation paradigm in artificial intelligence, where instead of random decisions, constraint-guided more informed optimisation decisions are of particular interest. We run our experiments on a range of standard benchmark problem instances that include instances from real airports and instances crafted using real airport parameters, and contain scenarios involving multiple runways and both landing and takeoff operations. We show that our proposed algorithms significantly outperform existing state-of-the-art aircraft sequencing algorithms.

Toxicity prediction of chemical compounds is a grand challenge. Lately, it achieved significant progress in accuracy but using a huge set of features, implementing a complex blackbox technique such as a deep neural network, and exploiting enormous computational resources. In this paper, we strongly argue for the models and methods that are simple in machine learning characteristics, efficient in computing resource usage, and powerful to achieve very high accuracy levels. To demonstrate this, we develop a single task-based chemical toxicity prediction framework using only 2D features that are less compute intensive. We effectively use a decision tree to obtain an optimum number of features from a collection of thousands of them. We use a shallow neural network and jointly optimize it with decision tree taking both network parameters and input features into account. Our model needs only a minute on a single CPU for its training while existing methods using deep neural networks need about 10 min on NVidia Tesla K40 GPU. However, we obtain similar or better performance on several toxicity benchmark tasks. We also develop a cumulative feature ranking method which enables us to identify features that can help chemists perform prescreening of toxic compounds effectively.

Empty or limited storage capacities between machines introduce various types of blocking constraint in the industries with flowshop environment. While large applications demand flowshop scheduling with a mix of different types of blocking, research in this area mainly focuses on using only one kind of blocking in a given problem instance. In this paper, using makespan as a criterion, we study permutation flowshops with zero capacity buffers operating under mixed blocking conditions. We present a very effective scatter search (SS) algorithm for this. At the initialisation phase of SS, we use a modified version of the well-known Nawaz, Enscore and Ham (NEH) heuristic. For the improvement method in SS, we use an Iterated Local Search (ILS) algorithm that adopts a greedy job selection and a powerful NEH-based perturbation procedure. Moreover, in the reference set update phase of SS, with small probabilities, we accept worse solutions so as to increase the search diversity. On standard benchmark problems of varying sizes, our algorithm very significantly outperforms well-known existing algorithms in terms of both the solution quality and the computing time. Moreover, our algorithm has found new upper bounds for 314 out of 360 benchmark problem instances.

All-interval series is a standard benchmark problem for constraint satisfaction search. An all-interval series of size n is a permutation of integers [0, n) such that the differences between adjacent integers are a permutation of [1, n). Generating each such all-interval series of size n is an interesting challenge for constraint community. The problem is very difficult in terms of the size of the search space. Different approaches have been used to date to generate all the solutions of AIS but the search space that must be explored still remains huge. In this paper, we present a constraint-directed backtracking-based tree search algorithm that performs efficient lazy checking rather than immediate constraint propagation. Moreover, we prove several key properties of all-interval series that help prune the search space significantly. The reduced search space essentially results into fewer backtracking. We also present scalable parallel versions of our algorithm that can exploit the advantage of having multi-core processors and even multiple computer systems. Our new algorithm generates all the solutions of size up to 27 while a satisfiability-based state-of-the-art approach generates all solutions up to size 24.

We describe a constraint-based automated planner named Transition Constraints for Parallel Planning (TCPP). TCPP constructs its constraint model from a redefined version of the domain transition graphs (DTG) of a given planning problem. TCPP encodes state transitions in the redefined DTGs by using table constraints with cells containing don't cares or wild cards. TCPP uses Minion the constraint solver to solve the constraint model and returns a parallel plan. We empirically compare TCPP with the other state-of-theart constraint-based parallel planner PaP2. PaP2 encodes action successions in the finite state automata (FSA) as table constraints with cells containing sets of values. PaP2 uses SICStus Prolog as its constraint solver. We also improve PaP2 by using don't cares and mutex constraints. Our experiments on a number of standard classical planning benchmark domains demonstrate TCPP's efficiency over the original PaP2 running on SICStus Prolog and our reconstructed and enhanced versions of PaP2 running on Minion.