Prediction of Mechanism of Action Using Ensembled Deep Neural Networks Splits

Mechanisms of Action (MoA) is a phenomenon of how drugs work at a molecular level to produce their therapeutic effects in the body. As drugs interact with specific molecules, like receptors or enzymes, they play important roles in biological responses and processes. These interactions augment the activities of these molecules, leading to changes in how cells and tissue’s function. Knowing a drug's MoA is vital in pharmacology because it allows researchers to predict how effective it will be, what side effects might occur, and how it might interact with other medications. Machine Learning (ML) provides a valuable contribution to several fields ranging from engineering, social, computing, and medical sciences. Interest has grown in the applicability of the ML in medicine and biosciences, signifying our interest in this field. If the ML models can successfully learn the MoA distribution with minimum loss and high accuracy, the obtained model can then predict a compound's MoA based on a specific cellular signature, benefiting the advanced drug discovery cause-effect process. Understanding the process of MOA is crucial in the identification of drug efficacy concentration in addition to the toxic and lethal dose along with any adverse effects.

The current paper proposes a network of EDNNS to predict the multiple targets of the MoA responses of different samples. As such, an exploratory answer to the question is reported! Can the drug's MoA based on gene expression and cell viability data be predicted using the ML paradigm? The Harvard's Laboratory for Innovation Science dataset is used to explore the answer. The dataset consists of various groups of features with more than two hundred targets of enzymes and receptors. The samples are profiled at different time points and doses. Multiple ML models are evaluated with special interest inclined towards the EDNNS model due to its superior regression performance. Our proposed approach is similar to the stacking ensemble of multiple deep NNs. For performance evaluation, the average value of the logarithmic loss function is used. The experimental evaluation shows that the ML paradigm, especially the EDNNS, provides better MOA prediction and can be used for practical applications and scenarios.

Understanding a drug's MOA is not a pre-requisite for drug approval as long as the safety and efficacy are well documented. Even though largely unclear, Metformin's MOA is proposed to control diabetes through AMP-activated protein kinase's (AMPK) inhibition [1]. Similarly, Dimebon, which was postulated to stabilize mitochondria, thereby possessing anti-Alzheimer potency, nevertheless had to be aborted during the phase 3 trial as cognition potency was due to its histamine and serotonin receptors interaction and not as an anti-Alzheimer effect [2]. A compound's MOA is attributed to the protein signaling in a pathway, the role of effector protein, and differing protein expressions depending on the site of action [3].

The latest technological research methodologies can comprehensively test any compound's hypothesized MOA at many levels using ML, pathway enrichment, connectivity mapping, causal reasoning, and the compound's previously documented site of action and signaling proteins interaction. The current study provides an overview of the different data levels, and compounds' MoA can be elucidated.

Bioactivity data like High Throughput Screening (HTS) provides resourceful data in predicting sites of MoA for orphan drugs [4] wherein a multitude of drug molecules can be evaluated against a panel of compound targets. The obtained data has its limitations as it does not consider the pharmacokinetic properties of absorption, distribution, metabolism, and excretion (ADME) and lacks data at the cellular level about the modified signaling pathway in response to drug binding on the receptors. Many other freely accessible databases like ChEMBL, PubChem, ExCAPE, and BindingDB extract the data from the published literature with different lab and experimental settings and assays. The variation in experiments may be related to a data error, adding to the complexity of the generated data. The morphological modifications like apoptosis or alteration in cytoskeletal protein upon adding any compound to the cell line can be visualized using cell imaging [5]. New assays like the Cell Painting assay [6] using fluorescent dyes depict morphological changes in organelles.

There is various segmentation software capable of analyzing and mining features from 2D cell images like CellProfiler [7], CellCognition [8], and PhenoRipper [9], in addition to automated feature mining software like Convolutional Neural Networks (CNNs) [10] that involves deep learning (DL). DL utilizes raw images for extraction and precise identification of cells or cellular sub-compartments or substructures [11, 12]. One limiting factor for DL software is the inability of all tested drug molecules to alter cellular morphology.

Proteomics provides vital information related to the compound's MOA through the expression of the proteins. The methodology's limitations are time-consuming, expensive, and lacking quantification of all proteins in an experiment, which can be circumvented by obtaining data from multiple experiments or studies [13, 14]. Similarly, metabolomics provides changes in metabolite enzyme activity by any drug molecule. Deeper mechanistic information of any drug molecule can thus be obtained through changes in transcription, translation, proteomics, and metabolites levels. The metabolic method's limitations require multiple experiments to obtain the complete metabolome coupled with data interpretation difficulty and wide data variability requiring experimental replication [15]. Phosphoproteomics provides additional potential pathways modulated by the drug molecules following alteration in protein phosphorylation. The limitation in phosphoproteomics is overcome through PhosphoSitePlus [16], wherein the phosphorylation sites are mapped to proteins, thereby providing biological context through disease and pathway annotations.

ML has significantly influenced drug discovery, beginning with quantitative structure-activity relationship (QSAR) modeling. This method used statistical techniques to predict biological activity based on chemical structure, despite computational and data limitations at the time. The introduction of high-throughput screening (HTS) and genomics resulted in a surge of biological data, which increased the application of ML algorithms like support vector machines (SVMs) and random forests. These algorithms enabled the rapid analysis of large datasets, facilitating the identification of potential drug targets and accelerating early drug discovery stages. Further advancements in deep learning have greatly advanced the field. For example, AtomNet has accurately predicted small molecule binding affinities to proteins, aiding in the discovery of treatments for diseases such as Ebola and multiple sclerosis. DeepMind's AlphaFold has also revolutionized protein structure prediction. Researchers have used AI to analyze biomedical data, identifying drug repurposing opportunities like baricitinib for COVID-19, which was validated through clinical trials. However, the application of ML in drug discovery faces challenges, including a scarcity of high-quality, labeled data, limited model interpretability, and difficulties in generalizing from training data to real-world scenarios. Addressing these challenges requires strategies such as enhancing data integration across domains, developing interpretable ML models, and using techniques like transfer learning and few-shot learning to tackle data scarcity. Collaborative efforts among multidisciplinary teams are also essential to ensure ML models are grounded in biological insights and practical needs. These efforts promise faster, more efficient, and effective approaches to identifying novel therapeutics and improving patient outcomes.

The article introduces an NN-based Ensemble Deep Neural Network System (EDNNS) for predicting multiple targets of Mode of Action (MoA) responses across various samples. It investigates the prediction of MoA using gene expression and cell viability data within a machine-learning framework. Comparative evaluation with MLP, Deep NN, ResNet, and Xgboost reveals that the proposed EDNNS ensemble exhibits greater robustness with lower loss compared to these individual models. The findings of this article may provide advancements in drug discovery by identifying new therapeutic targets and biomarkers, facilitating personalized medicine, and enabling safer, more effective treatments. It helps create drugs that target specific disease mechanisms, reducing side effects and increasing efficacy. Additionally, we can repurpose existing drugs for new uses, speeding up the development of therapies for conditions lacking effective treatments.

3.1 Connectivity mapping

Connectivity mapping equates modification of gene expression cell lines incubated with the drug molecule to the set of already present reference signatures associated with any disease or other known drug molecules MOA or disease.

3.2 Pathway enrichment

Pathway enrichment analysis provides the opportunity to reduce a vast array of genes or proteins, even in the absence of any biological information, to more minor processes, making them more interpretable than the gene database itself, thereby providing more context in identifying any given phenotype interest.

3.3 Causal reasoning

Any modification in mRNA expression is determined precisely through causal reasoning using prior knowledge network (PKN) molecular interaction from gene expression data [30]. Any compound's MoA is determined using such methodology through the compound-induced modulated signaling proteins. A compound's MoA is determined accurately without mixing gene level with protein activity through causal reasoning's estimated perturbed protein signaling equated to pathway enrichment. One limitation of casual reasoning is an absence of validated signaling protein output equated to experimentally determined protein activity modification.

3.4 Clustering aggregation

These algorithms encompasses essential tools for exploration, serving the vital purpose of identifying and organizing groups of related or interacting samples [31]. These methods operate based on the principles of similarity or distance, utilizing metrics like k-means clustering or data density measures such as DBSCAN. They play a pivotal role in analyzing complex datasets containing gene expression, chemical compositions, and image-based information, which are often unstructured and high-dimensional [32]. By uncovering patterns and relationships inherent in these intricate datasets, clustering techniques help to unravel the complexity of biological processes. Ultimately, they provide valuable insights that significantly contribute to the advancement of scientific understanding in various fields.

3.5 Group Factor Analysis

Group Factor Analysis (GFA) is an integrative analysis determining the correlation of various data like chemical descriptors and genetic processes [33]. Multi-omics Factor Analysis (MOFA) and MOFA+ are examples of GFA linked to clinically relevant findings in which the integrated clusters comprising 24 anti-tumor compounds' biological data demonstrated to possess MEK inhibitor activity on hematopoietic cells [34].

3.6 Modeling and prediction

Supervised ML trains and identifies patterns in the presence of identified labels used to optimize a function by connecting to features like gene expression with an endpoint like a compound's activity at specific points on the label. Supervised ML demonstrates the compound's MoA by predicting primary and off-target drug interactions [35]. DL methods are typical artificial neural networks (ANN) with a multitude of discrete layers along with specialized training modules, thereby mimicking the human brain's complex neural network. 

Each methodology has its benefit and limitations like network and pathway depend on preliminary information related to the curation quality. In addition to being time-constrained, ML has its share of other limitations in interpreting data, and the data obtained is restricted to a small part of the high-level MoA space. Similarly, different data types obtain a varied version of the MoA biology and thus enable a more comprehensive understanding of compound MoA. ML contributes to MOA elucidation through prediction, which upon further interpretation, can be utilized by researchers.

The dataset for MoA prediction is obtained from and available on Kaggle and is provided as such by the Connectivity Map (CM), a project by MIT, Harvard, LISH, and the NIH-LINCS [36]. The dataset is utilized for academic purposes and experimental evaluation of the ML models only. The dataset consists of the following features:

(1) Gene expression

(2) Cell viability

(3) Multiple targets of MoA

The problem represented by MoA data is a multilabel classification problem. Thus, data has multiple targets; however, these are not multiple classes. First, the data analysis is presented, followed by MoA prediction using ensemble stacking.

Table 1 presents a sample of the MoA data. The training features set represents the gene expression data and cell viability and is represented by "g" and "c" symbols. In the dataset, the samples cp_type is treated with a cp_vehicle compound or with ctrl_vehicle (control perturbation). The ctrl_vehicle has no MoA. cp_time and cp_dose represent 24, 48, and 72 hours’ duration, and the dose is either high or low.

From a data wrangling point of view, no missing values were found. For visual analysis of the data, Figure 2 shows the plot for the gene expression features, only representing g-0 and g-2. Figure 3 shows the cell viability features, C-0 and C-5, as representative samples.

2.png

Figure 2. Plot of gene expression features g-0 and g-2

3.png

Figure 3. Plot of cell viability features c-0 and c-5

Table 1. Sample of data from the training set

The current paper proposes NN based EDNNS approach in predicting multiple targets of the MoA responses of different samples. As such, the current study explored the answer to the question of the prediction of MoA based on gene expression and cell viability data using the ML paradigm. The current study evaluated MLP, Deep NN, ResNet, Xgboost and found that the proposed ensemble EDNNS is more robust than the MLP, Deep NN, ResNet, Xgboost having less loss. The comparatively less loss thus means a more robust and accurate model for prediction. This work can benefit the advanced drug discovery cause-effect by providing valuable insights and exciting directions for future research.