Artificial Intelligence Platform for Molecular Targeted Therapy

The following sections are included:

• Dedication
• Foreword
• About the Author
• Preface
• Contents

We introduce AI at a fairly elementary level adopting a heuristic perspective. The presentation focuses mostly on machine learning and specifically on deep learning. The narrative is tailored to drug designers and developers seeking to accelerate the discovery cycle. Basic knowledge of linear algebra and some familiarity with Python programming is all that is required for the pharma researcher to follow the presentation in detail. A crucial distinction is delineated between standard well-established cheminformatics and pharmacoinformatics applications and AI deployed to empower the design of dynamic drug/target interfaces, the focus of the book. Thus, the chapter sets the tone for the remainder of the book which advocates the deployment of AI to tackle dynamic problems in drug design, including the incorporation, mimicking and simulation of in vivo settings. In so doing, the book introduces the arsenal of epistructural biology to provide the proper representational/conceptual framework that enables inference through the tensor flow of convolutional neural networks.

Biological processes at molecular scales remain difficult to understand and control and the efficacy of AI approaches to design targeted therapeutic interventions would be limited unless such processes are properly represented in the data under scrutiny and in the features extracted thereof. First-principle approaches to the standing problems in molecular biophysics are still facing daunting challenges. Thus, the protein folding problem, the physical underpinnings of enzyme catalysis and protein associations, and the molecular etiology of aggregation-related disease remain almost as intractable now as when they were first posed decades ago. Not surprisingly, the molecular design of therapeutic agents remains more serendipitous than rational. This state of affairs takes us now to a cross-road of several disciplines wherein a new field, epistructural biology, is emerging as a scaffolding field to AI approaches to drug design. As shown in this chapter, this is so because the protein epistructure is actually encoded in the structure, with a single structural feature, the dehydron, encompassing the multiscale complexity of the structure-solvent relationship. This is a crucial step in order to leverage AI, as the dehydron pattern may be represented in tensor form, hence becoming tractable within an AI context…

Natural proteins fold and have been selected to fold in vivo, that is in cellular environments, not in vitro. Likewise, therapeutic agents target dysfunctional or deregulated proteins in vivo, not in the test tube, where screening and affinity assays are typically performed. In regards to modeling and predictive efforts, the wanton molecular complexity of in vivo environments places them out of reach for investigation with molecular dynamics (MD) or other computational methods. The advent of artificial intelligence (AI) drastically changes this scenario, as shown in this chapter. Even in an in vitro setting, the generation of realistic protein folding pathways remains a challenge for molecular dynamics (MD) and other computational systems. This is so because for most single-domain chains, relevant timescales are inaccessible and key rare events are typically missed. Furthermore, most computational efforts have focused on in vitro folding, hence they are often misplaced as natural proteins have evolved to fold in vivo, where selective interference prevents the chain from getting kinetically trapped in misfolded states. In this chapter we empower MD by subsuming the computations within a deep learning platform that constructs an in vivo context shown to enhance the expediency of the folding process. To this effect, we build a convolutional neural network (CNN) trained with short MD runs that incorporate in vivo molecular complexity. The AI platform learns to propagate a coarse grained folding trajectory that can be decoded as an extended MD run assessing the participation of the in vivo context to expedite the folding process.

The PDB (protein data bank) constitutes an information repository on proteins with reported structure. This chapter offers a meta-analysis of the PDB where structures are examined vis-à-vis the aqueous surrounding solvent. We address the issue of enriching the field of structural bioinformatics by incorporating information on protein-water interfaces subsumed in reported protein structures. That is, we apply the methods of epistructural biology and machine learning to carry out a meta-reading of the PDB that will ultimately become helpful for drug design pursuits. As shown in Chapter 2, protein epistructure is encoded in protein structure and this chapter decodes this information…

The protein-water interface constitutes a unique marker for molecular evolution. This fact endows epistructural biology with a vantage point to explore the evolutionary axis, which is essential to understand and manipulate biological phenomena. A key result emerging from such forays is that dehydron patterns are not conserved across homologous proteins. Drug designers can take advantage of this fact to create specific targeted agents. As described previously, dehydrons may be targeted by wrapping ligands that serve as leads to therapeutic drugs. Thus, the evolutionary insights obtained from epistructural analysis have implications for the drug designer. Homologous proteins typically share the same fold and this structural similarity introduces problems for molecular targeted therapy since it leads to undesirable cross reactivities with hazardous off-target effects. As shown in this chapter, while the topology of a protein fold is conserved across homologs, the wrapping patterns are substantially different, enabling the wrapping drug to funnel its impact on clinically relevant targets. The evolutionary root of the differences in dehydron pattern across homologs is established across species and within the human species. Wrapping variations across homologs can be exploited to considerable advantage as we aim at engineering target-specific and species-specific therapeutic agents and build meaningful animal models for disease. In delineating the evolutionary forces that promote epistructural differences across orthologs, we discovered that random genetic drift plays an operational role in promoting dehydron enrichment. This structural degradation enhances the propensity for protein interactivity and becomes more pronounced in species with low population, where mildly deleterious mutations resulting from random drift have a higher probability of getting fixed in the population. The fitness consequences of this evolutionary strategy are assessed for humans, and reveal their high exposure to fitness catastrophes.

This chapter unravels the catalytic role of dehydrons in biological chemistry. It prompts a significant revision of the field to encompass the chemical functionality of frustrated interfacial water. Through the induced functionalization of vicinal water, the dehydron is shown to be an enabler and a promoter of enzymatic activity…

This chapter is devoted to the control of specificity for the next-generation targeted therapeutic agents. It is shown that epistructure-based molecular design enables a funneling of the drug inhibitory power towards targets of clinical relevance while suppressing off-target cross reactivity. Wrapping design is thus established as a paradigm in molecular engineering that broadens the technological base of the pharmaceutical industry.

As most clinicians would acknowledge, targeted drug-based therapies have met only moderate success in cancer treatment. This is partly because of the evolutionary substrate of the disease that typically selects for resistant phenotypes, and partly because of the drug-induced suppression of the adaptive immune response. In this chapter we shall address the second issue and develop immuno-synergistic targeted agents. The discussion focuses on epistructural tuning of extant targeted therapeutic agents to make them compatible with immunotherapy.

The immune checkpoint blockers have proven most successful to treat solid tumors harboring a DNA mismatch repair (MMR) deficiency or other conditions causative of high antigenic activity related to hypermutability. Using the methods of epistructural biology aided by AI, we show that it is possible to create a drug-promoted cancer phenotype akin to the MMR deficiency in solid tumors, and thereby engineer a drug-induced susceptibility to checkpoint blockers through drug-promoted metabolic stress on DNA synthesis. Issues of compatibility between checkpoint immunotherapy and targeted therapy are first sorted out in order to enable and empower the checkpoint blocker. The potential of such drug-checkpoint-blocker combinations as universal treatments for solid tumors is shown to merit clinical evaluation…

Molecular dynamics (MD) remains the most valuable tool for the predictive understanding of the behavior of biological mater at the atomistic level. However, its true potential remains largely untested because relevant timescales are often beyond reach, limited portions of conformation space can get sampled, and infrequent events of direct relevance to biophysical processes are missed. A culprit for these shortcomings is the huge informational burden that is required to iterate the integration steps in order to generate an MD trajectory. To address the problem, we apply deep learning to a) encode the dynamics into a simplified embodiment that retains only essential topological features of the vector field that steers MD integration, and b) to propagate the simplified trajectory beyond the timescales accessible to atomistic MD. We simplify the flow via an equivalence relation that identifies conformations within basins of attraction in potential energy and encodes the dynamics onto a modulo-basin “quotient space” where fast motions are averaged out. Encoding MD-generated information in quotient space enables coverage of physically meaningful timescales while unraveling the underlying dynamic hierarchy. Deep learning is exploited to propagate the coarse grained state and to reconstruct it back at the atomistic level at a later time. As shown, the quotient-encoding-propagating-decoding scheme generates within few GPU hours protein folding pathways with experimentally verified outcomes. By contrast, MD computations covering comparable timespans would take hundreds of days on today’s fastest special-purpose supercomputers. Thus, quotient space constitutes a generative model that provides hierarchical understanding of MD simulation while enabling access to realistic timescales and the capture of physically meaningful rare events.

The absence of reported structure for a therapeutic target and the lack of information on its regulation pose major hurdles for rational drug design. This chapter delineates the exploitation of AI to steer the engineering of molecular targeted therapy in such challenging situations. The chapter focuses specifically on protein associations, where drug-based disruption of dysfunctional complexes poses a major challenge to targeted therapy. The problem becomes daunting when the structure and regulated modulation of the complex are unknown. To address the challenge, we leverage an artificial intelligence platform that learns from epistructural information and infers regulation-susceptible regions that also generate interfacial tension between protein and water, thereby promoting contextually tuned protein associations. The input consists of sequence-derived 1D-features. The network is configured with evolutionarily coupled residues and taught to search for phosphorylation-modulated binding epitopes. The discovery platform is benchmarked against a PDB-derived testing set and validated against experimental data on a therapeutic disruptor designed according to the inferred epitope for a large deregulated complex known to be recruited in heart failure. Thus, dysfunctional “molecular brakes” of cardiac contractility get released through a therapeutic intervention guided by artificial intelligence.

In this chapter we implement a deep learning (DL) platform that teaches drugs to target proteins. This is accomplished by introducing chemical scaffold modifications steered by a priori predictions of drug-induced binding-competent conformations of the intended target. In this way, we deal with the fact that target proteins are usually not fixed targets fitting the lock-key paradigm: they structurally adapt to ligands in ways that need to be predicted to empower drug discovery. In contrast with protein folding predictors that do not include disorder propensities as input, the neural network presented solves a “drug-induced folding problem” by incorporating as input sequence-derived signals for structural disorder. The DL system thus predicts conformations in floppy regions of the target protein that rely on associations with a conformation-selecting purposely designed drug to acquire and maintain their structural integrity. This is tantamount to generate within an AI-empowered drug discovery platform the a-priori drug-induced folding ensemble without committing to a specific ligand and then steer the modification of a lead compound to select a specific drug-induced conformation from within the ensemble.

The following sections are included:

• Epilogue: AI constructs its own physics
• Appendix 1: Code for dehydron identification
• Index