Inference of drug off-target effects on cellular signaling using Interactome-Based Deep Learning

Background

Cellular Signalling

Biological signalling is an essential property of living organisms allowing cellular communication and coordinated responses to external environmental changes. Signalling is initiated by signalling molecules binding to receptors on the cell surface (reception), leading to a conformational change of the receptor. This results in a series of intracellular events (e.g. phosphorylation cascades), which amplifies and relays the signal through the cell (signal transduction). The transduced signal ultimately leads to a specific cellular response such as changes in gene expression and alterations in enzyme activity.

Diseases like cancer (Sever & Brugge, 2015), diabetes (Onyango, 2020) and others (Spiegel, 1996 & Zhang et al., 2019) can result from cell signalling deregulation, which is usually caused by genetic mutations, environmental factors, or a combination of both. Various therapeutic approaches are being explored to either restore cellular balance or eliminate diseased cells. One strategy to restore normal cellular behaviour involves the use of small molecules like kinase inhibitors  (Zhong et al., 2021), which aims at suppressing kinases that are vital for tumour cell survival and proliferation while being less cytotoxic to non-cancerous cells. Another approach is the use of biologics, particularly monoclonal antibodies (Zahavi & Weiner, 2020). These specialised proteins bind to specific molecules involved in aberrant signalling pathways with high specificity to either block their activity or mark them for destruction by the immune system.

Artificial neural networks

Inspired by the way our brain works, an artificial neural network (ANN) is an interconnected group of nodes (“virtual neurons”) in a layered structure that are used for tasks like recognising patterns in images or making decisions based on data. ANNs consist of one input layer, multiple hidden layers and one output layer (Fig. 1). Similar to neurons, each node takes inputs from the previous layer, performs a calculation, and passes its output to the next node. Node connections are given numerical weights (w), with positive weights indicating that one node enhances the activity of another and negative weights indicating suppression. Nodes with higher weight values exert greater influence on the other nodes in the network. By learning from examples (training data), ANNs adjust the weights of these connections and thereby improve their performance over time.

The training data for many machine learning methods results from automated high-throughput screens like the Next Generation L1000 Connectivity Map (Subramanian et al., 2017) and/or the JUMP Cell Painting project (Chandrasekaran et al., 2023) which allow to a) identify drug mechanisms of action, b) nominate therapeutics for a particular disease, and c) construct biological networks among perturbations and genes. If you are interested in a similar approach where these datasets are being used, please check out our previous preLight post.

Fig. 1 Architecture of an Artificial Neural network. Created by Benjamin Maier in 2023.

Key Findings

Overview

Building upon their previously published Large-scale knowledge-EMBedded Artificial Signaling-networks (LEMBAS) (Nilsson, Peters et al., 2022), Meimetis and colleagues developed a computational model called DT-LEMBAS. Their new model uses artificial neural networks (ANN) to simultaneously predict drug-target interactions and their cell-line specific effects on intracellular signalling pathways.

The drug-target (DT) submodule of DT-LEMBAS takes the drug concentration as input to predict the signalling effects of the drugs on their targets. This is done using prior information from known drug-target interactions as well as chemical similarities between different drugs. The second module, LEMBAS, is a recurrent ANN model designed to simulate intracellular signalling processes. It takes the signalling effects of drugs as input to estimate the activity of transcription factors (TFs).

Unlike other approaches, DT-LEMBAS incorporates a prior knowledge network of signal transduction, along with information about transcriptional regulation and the chemical structure similarity of drugs. This enhances the interpretability of both on- and off-target effects on signalling cascades and ensures that the model is mechanistically plausible. The model outperforms conventional machine learning methods in predicting TF’ activity and identifies both known and novel drug-target interactions, which the preprint authors validated in independent datasets. As a case study (not featured in this preLight article), the authors applied DT-LEMBAS to analyse the impact of the drug Lestaurtinib on downstream signalling pathways, highlighting its potential to improve therapeutic design by identifying potential off-target drug effects.

Key Finding I: Benchmarking predictability of TF activities

First, the authors trained their model using the L1000 dataset (Subramanian et al., 2017) to identify TFs that could be predicted with high accuracy and validate whether the drug module predictions generalise using a cross-validation approach. They determined that their model accurately predicts the activities of many TFs across a wide range of cell lines. Notably, TFs showing strong fits in training also excelled in validation, while those with poor fits had lower performance, possibly due to noisy data or limited impact of the TF on the cell’s transcriptomic profile. Next, the model performance was benchmarked against four basic machine learning techniques. DT-LEMBAS consistently outperformed the other approaches, as well as the randomised models, as determined by statistical tests.

Key Finding II: Controlling the number of inferred interactions

To address gaps in drug-target interaction knowledge, the model must be able to infer new interactions. One approach is zero regularisation, allowing parameter adjustments without penalties thereby allowing drug connections with all potential targets (risk of overfitting). Conversely, infinite regularisation limits the model to known interactions (risk of under-fitting). Striking a balance, the authors employ L2 regularisation (Ridge regression) to add penalties as model complexity increases. This enables the model to infer drug-target interactions that are not part of the prior knowledge to better explain transcriptional data while preventing overfitting by restricting its ability to connect drugs with all targets.

The approach assigns importance scores to interactions, where negative indicates potential inhibition and positive suggests activation. Interactions with low scores are less relevant for explaining TF activity, while high scores are deemed important. Based on quantifying the impact of interaction removal on model accuracy for each drug, the authors chose a cutoff at a 25% increase in global error. This threshold successfully retrieved about 95% (670 out of 710) of known drug-target interactions, aligning with expectations, as these interactions are supported by literature or clinical evidence and are anticipated to contribute significantly to the observed transcriptional profile.

Key Finding III: Evaluating the inference of direct interactions by an ensemble of models

In addition to retrieving known interactions, the previous analysis found 13,000 novel interactions, which might represent undiscovered direct on-target drug-target interactions, indirect effects or model artefacts (false positives). Validating these predicted interactions is challenging due to the absence of ground truth data.

Instead, the authors made use of an independent database DrugBank which contains known direct drug-target interactions which were not present in the original model’s prior knowledge. They then attempted to predict these interactions based on how frequently they were inferred by their cell-line specific models as well as by some additional statistical metrics such as recall and F1 score. When considering all interactions that appeared in at least one model, the authors observed low precision (proportion of correctly predicted positive cases) and overall accuracy (proportion of all cases correctly classified). However, when an interaction was inferred more frequently, both accuracy and precision improved, reaching perfect precision for the most frequent predictions. This suggests that interactions found in multiple models are more likely to be direct interactions, which can be used for inferring novel drug-target interactions.

Key Finding IV: Identification of transcription factors regulated by off-target effects

Using their newly established criterion to predict direct on-target drug-target interactions, the authors identified off-target effects that contribute to the transcriptional profile. They removed all non-direct interactions and checked if the model could no longer explain the transcriptional profile, suggesting the presence of additional contributing interactions induced by off-target effects.

Case Study: A subnetwork explaining off-target effects of Lestaurtinib

Next, the authors explored (off-target) interactions for the tyrosine kinase inhibitor Lestaurtinib. The model identified an average of 82 potential interactions per model, including all known interactions found in the training dataset and in the DrugBank validation dataset. The interactions were then used to construct a subnetwork to elucidate Lestaurtinib’s mechanism of action on FOXM1. The resulting signalling network (Fig. 2) suggested that Lestaurtinib might cause an indirect inhibition of FOXM1 through either CDK1 and CDK2, which is supported by the literature. More detailed downstream analysis further supported the inhibition of FOXM1 through CDK2. Overall, the case study outlines how DT-LEMBAS combined with downstream validation analysis might be used to propose potential therapeutic interventions and drug combinations.

Fig. 2 Signalling network showing how Lestaurtinib inhibits FOXM1 through CDK1 and CDK2. Figure taken from Meimetis et al. (2023), BioRxiv published under the CC-BY-NC-ND 4.0 International licence.

Conclusion and Perspective

Despite significant reductions in the cost of computing and sequencing driven by technological advancements and economies of scale, the price per new drug has experienced an exponential increase over the past few decades. Nevertheless, given the exponentially increasing rate of data collection paired with advances in computational biology making use of the big data, there is a massive potential to accelerate drug design while reducing the associated costs and resources. For instance, virtual pre-screenings of large chemical libraries can narrow down the number of compounds that need to be experimentally tested to identify potential drug candidates. Moreover, as shown e.g. in the presented preprint, computational methods can identify potential drug targets by analysing biological data to understand the roles of specific proteins, genes, or pathways in diseases.

Questions to the Authors

Q1: The OmniPath-derived prior knowledge signalling network constructed in this study offers a comprehensive understanding of “average” cell-cell interaction networks. However, it falls short in capturing patient heterogeneity and may obscure locally significant network structures. Would it be possible to integrate patient omics data in the workflow and how would this change the model training routine?

Q2: What was the reasoning for performing a L2-based regularisation (Ridge Regression) over a L1 regularisation (Lasso Regression) or a combined Elastic Net Regularisation?

Q3: What would be the next steps to expand the framework to account for metabolism and cell-cell interactions (in the tumour microenvironment) and what are the biggest challenges to overcome?

Q4: Can you speculate about how much better the model performance would be when additionally being trained with data from the JUMP Cell Painting project (Chandrasekaran et al., 2023) or using AlphaFold structure predictions?

References

Nilsson, A., Peters, J.M., Meimetis, N. et al. Artificial neural networks enable genome-scale simulations of intracellular signaling. Nat Commun 13, 3069 (2022). https://doi.org/10.1038/s41467-022-30684-y

Onyango A. N. (2020). Mechanisms of the Regulation and Dysregulation of Glucagon Secretion. Oxidative medicine and cellular longevity, 2020, 3089139. https://doi.org/10.1155/2020/3089139

Sever, R., & Brugge, J. S. (2015). Signal transduction in cancer. Cold Spring Harbor perspectives in medicine, 5(4), a006098. https://doi.org/10.1101/cshperspect.a006098

Spiegel A. M. (1996). Defects in G protein-coupled signal transduction in human disease. Annual review of physiology, 58, 143–170. https://doi.org/10.1146/annurev.ph.58.030196.001043

Zahavi, D., & Weiner, L. (2020). Monoclonal Antibodies in Cancer Therapy. Antibodies (Basel, Switzerland), 9(3), 34. https://doi.org/10.3390/antib9030034

Zhang, L., Yousefzadeh, M. J., Suh, Y., Niedernhofer, L. J., & Robbins, P. D. (2019). Signal Transduction, Ageing and Disease. Sub-cellular biochemistry, 91, 227–247. https://doi.org/10.1007/978-981-13-3681-2_9

Zhong, L., Li, Y., Xiong, L. et al. Small molecules in targeted cancer therapy: advances, challenges, and future perspectives. Sig Transduct Target Ther 6, 201 (2021). https://doi.org/10.1038/s41392-021-00572-w

Tags: drug design, machine learning, mechanism of action, neural network

Posted on: 29 November 2023

doi: https://doi.org/10.1242/prelights.36027

Read preprint (No Ratings Yet)