In the space of only a few years, deep generative modeling has revolutionized how we think of artificial creativity, yielding autonomous systems which produce original images, music, and text. Inspired by these successes, researchers are now applying deep generative modeling techniques to the generation and optimization of molecules—in our review we found 45 papers on the subject published in the past two years. These works point to a future where such systems will be used to generate lead molecules, greatly reducing resources spent downstream synthesizing and characterizing bad leads in the lab. In this review we survey the increasingly complex landscape of models and representation schemes that have been proposed. The four classes of techniques we describe are recursive neural networks, autoencoders, generative adversarial networks, and reinforcement learning. After first discussing some of the mathematical fundamentals of each technique, we draw high level connections and comparisons with other techniques and expose the pros and cons of each. Several important high level themes emerge as a result of this work, including the shift away from the SMILES string representation of molecules towards more sophisticated representations such as graph grammars and 3D representations, the importance of reward function design, the need for better standards for benchmarking and testing, and the benefits of adversarial training and reinforcement learning over maximum likelihood based training.

Deep learning for molecular design—a review of the state of the art

Deep learning bears promise for drug discovery, including advanced image analysis, prediction of molecular structure and function, and automated generation of innovative chemical entities with bespoke properties. Despite the growing number of successful prospective applications, the underlying mathematical models often remain elusive to interpretation by the human mind. There is a demand for ‘explainable’ deep learning methods to address the need for a new narrative of the machine language of the molecular sciences. This Review summarizes the most prominent algorithmic concepts of explainable artificial intelligence, and forecasts future opportunities, potential applications as well as several remaining challenges. We also hope it encourages additional efforts towards the development and acceptance of explainable artificial intelligence techniques. Drug discovery has recently profited greatly from the use of deep learning models. However, these models can be notoriously hard to interpret. In this Review, Jimenez-Luna and colleagues summarize recent approaches to use explainable artificial intelligence techniques in drug discovery.

/pdf/drug-discovery-with-explainable-artificial-intelligence-5ed06p2guc.pdf

Drug discovery with explainable artificial intelligence

Identification of diagnostic markers for tuberculosis by proteomic fingerprinting of serum. Commentary

Current research has demonstrated that small cationic amphipathic peptides have strong potential not only as antimicrobials, but also as antibiofilm agents, immune modulators, and anti-inflammatories. Although traditionally termed antimicrobial peptides (AMPs) these additional roles have prompted a shift in terminology to use the broader term host defense peptides (HDPs) to capture the multi-functional nature of these molecules. In this review, we critically examined the role of AMPs and HDPs in infectious diseases and inflammation. It is generally accepted that HDPs are multi-faceted mediators of a wide range of biological processes, with individual activities dependent on their polypeptide sequence. In this context, we explore the concept of chemical space as it applies to HDPs and hypothesize that the various functions and activities of this class of molecule exist on independent but overlapping activity landscapes. Finally, we outline several emerging functions and roles of HDPs and highlight how an improved understanding of these processes can potentially be leveraged to more fully realize the therapeutic promise of HDPs.

/pdf/reassessing-the-host-defense-peptide-landscape-911cd7opw4.pdf

Reassessing the Host Defense Peptide Landscape

Deep learning bears promise for drug discovery, including advanced image analysis, prediction of molecular structure and function, and automated generation of innovative chemical entities with bespoke properties. Despite the growing number of successful prospective applications, the underlying mathematical models often remain elusive to interpretation by the human mind. There is a demand for 'explainable' deep learning methods to address the need for a new narrative of the machine language of the molecular sciences. This review summarizes the most prominent algorithmic concepts of explainable artificial intelligence, and dares a forecast of the future opportunities, potential applications, and remaining challenges.

Computational antimicrobial peptide design and evaluation against multidrug-resistant clinical isolates of bacteria

Tuberculosis remains a major global health challenge worldwide, causing more than a million deaths annually. To determine newer methods for detecting and combating the disease, it is necessary to characterise global host responses to infection. Several high throughput omics studies have provided a rich resource including a list of several genes differentially regulated in tuberculosis. An integrated analysis of these studies is necessary to identify a unified response to the infection. Such data integration is met with several challenges owing to platform dependency, patient heterogeneity, and variability in the extent of infection, resulting in little overlap among different datasets. Network-based approaches offer newer alternatives to integrate and compare diverse data. In this study, we describe a meta-analysis of host’s whole blood transcriptomic profiles that were integrated into a genome-scale protein–protein interaction network to generate response networks in active tuberculosis, and monitor their behaviour over treatment. We report the emergence of a highly active common core in disease, showing partial reversals upon treatment. The core comprises 380 genes in which STAT1, phospholipid scramblase 1 (PLSCR1), C1QB, OAS1, GBP2 and PSMB9 are prominent hubs. This network captures the interplay between several biological processes including pro-inflammatory responses, apoptosis, complement signalling, cytoskeletal rearrangement, and enhanced cytokine and chemokine signalling. The common core is specific to tuberculosis, and was validated on an independent dataset from an Indian cohort. A network-based approach thus enables the identification of common regulators that characterise the molecular response to infection, providing a platform-independent foundation to leverage maximum insights from available clinical data. Patients suffering from tuberculosis (TB) show a high extent of variations in their gene expression patterns. Such heterogeneity poses major road blocks to our understanding of how hosts respond to the disease. A number of studies have profiled transcriptomes of human blood samples from TB patients, but a meta-analysis indicates that very few changes are consistently seen. The problem, to a large extent, lies with the way large data is analysed. We have used a genome-wide network approach to characterise the host response and have identified a common-core in the TB response networks of different patients, indicating the presence of unified host response mechanisms. This core network provides a comprehensive view into the most significant regulators of the infection-mediated biological processes across patients from different populations, and it shows partial reversals upon treatment.

/pdf/meta-analysis-of-host-response-networks-identifies-a-common-52dz4p7vjl.pdf

Meta-analysis of host response networks identifies a common core in tuberculosis.

Lipocalins constitute a superfamily of extracellular proteins that are found in all three kingdoms of life. Although very divergent in their sequences and functions, they show remarkable similarity in 3-D structures. Lipocalins bind and transport small hydrophobic molecules. Earlier sequence-based phylogenetic studies of lipocalins highlighted that they have a long evolutionary history. However the molecular and structural basis of their functional diversity is not completely understood. The main objective of the present study is to understand functional diversity of the lipocalins using a structure-based phylogenetic approach. The present study with 39 protein domains from the lipocalin superfamily suggests that the clusters of lipocalins obtained by structure-based phylogeny correspond well with the functional diversity. The detailed analysis on each of the clusters and sub-clusters reveals that the 39 lipocalin domains cluster based on their mode of ligand binding though the clustering was performed on the basis of gross domain structure. The outliers in the phylogenetic tree are often from single member families. Also structure-based phylogenetic approach has provided pointers to assign putative function for the domains of unknown function in lipocalin family. The approach employed in the present study can be used in the future for the functional identification of new lipocalin proteins and may be extended to other protein families where members show poor sequence similarity but high structural similarity.

/pdf/structure-based-phylogenetic-analysis-of-the-lipocalin-1ujiqtm32e.pdf

Structure-Based Phylogenetic Analysis of the Lipocalin Superfamily

In biological systems, diseases are caused by small perturbations in a complex network of interactions between proteins. Perturbations typically affect only a small number of proteins, which go on to disturb a larger part of the network. To counteract this, a stress-response is launched, resulting in a complex pattern of variations in the cell. Identifying the key players involved in either spreading the perturbation or responding to it can give us important insights. We develop an algorithm, EpiTracer, which identifies the key proteins, or epicenters, from which a large number of changes in the protein-protein interaction (PPI) network ripple out. We propose a new centrality measure, ripple centrality, which measures how effectively a change at a particular node can ripple across the network by identifying highest activity paths specific to the condition of interest, obtained by mapping gene expression profiles to the PPI network. We demonstrate the algorithm using an overexpression study and a knockdown study. In the overexpression study, the gene that was overexpressed (PARK2) was highlighted as the most important epicenter specific to the perturbation. The other top-ranked epicenters were involved in either supporting the activity of PARK2, or counteracting it. Also, 5 of the identified epicenters showed no significant differential expression, showing that our method can find information which simple differential expression analysis cannot. In the second dataset (SP1 knockdown), alternative regulators of SP1 targets were highlighted as epicenters. Also, the gene that was knocked down (SP1) was picked up as an epicenter specific to the control condition. Sensitivity analysis showed that the genes identified as epicenters remain largely unaffected by small changes. We develop an algorithm, EpiTracer, to find epicenters in condition-specific biological networks, given the PPI network and gene expression levels. EpiTracer includes programs which can extract the immediate influence zone of epicenters and provide a summary of dysregulated genes, facilitating quick biological analysis. We demonstrate its efficacy on two datasets with differing characteristics, highlighting its general applicability. We also show that EpiTracer is not sensitive to minor changes in the network. The source code for EpiTracer is provided at Github (
 https://github.com/narmada26/EpiTracer
 ).

/pdf/epitracer-an-algorithm-for-identifying-epicenters-in-dd3ov275ji.pdf

EpiTracer - an algorithm for identifying epicenters in condition-specific biological networks.

Sorafenib (SFB), a multi-kinase inhibitor, is the only approved drug for treating hepatocellular carcinoma (HCC). However, SFB shows low efficacy in many cases. HCC related mortality therefore remains to be high worldwide. SFB, a multi-kinase inhibitor is also known to modulate the redox homeostasis in cancer cells. To understand the effect of SFB on the redox status, a quantitative understanding of the system is necessary. Kinetic modeling of the relevant pathways is a useful approach for obtaining a quantitative understanding of the pathway dynamics and to rank the individual factors based on the extent of influence they wield on the pathway. Here, we report a comprehensive model of the glutathione reaction network (GSH(net)), consisting of four modules and includes SFB-induced redox stress. We compared GSH(net) simulations for HCC of six different etiologies with healthy liver, and correctly identified the expected variations in cancer. Next, we studied alterations induced in the system upon SFB treatment and observed differential H2O2 dynamics in all the conditions. Using metabolic control analysis, we identified glutathione S-transferase (GST) as the enzyme with the highest selective control coefficient, making it an attractive co-target for potentiating the action of SFB across all six etiologies. As a proof-of-concept, we selected ethacrynic acid (EA), a known inhibitor of GST, and verified ex vivo that EA synergistically potentiates the cytotoxic effect of SFB. Being an FDA approved drug, EA is a promising candidate for repurposing as a combination therapy with SFB for HCC treatment.

https://febs.onlinelibrary.wiley.com/doi/pdf/10.1111/febs.14641

Madhulika Mishra

Papers

Computational antimicrobial peptide design and evaluation against multidrug-resistant clinical isolates of bacteria

Meta-analysis of host response networks identifies a common core in tuberculosis.

Structure-Based Phylogenetic Analysis of the Lipocalin Superfamily

EpiTracer - an algorithm for identifying epicenters in condition-specific biological networks.

Identification of a co-target for enhancing efficacy of sorafenib in HCC through a quantitative modeling approach.