In this contribution, we will review recent developments in the field of artificial intelligence in drug discovery, and discuss their likely impact on project outcomes in the near future. In particular, this will cover the aspect of model validation and the impact of 'AI' on the transition of compounds to the in vivo stages, which is their only raison d'être in the first place.

Poor solubility is arguably the biggest hurdle affecting drug development and the use of solubility enhancing technologies is commonplace throughout the industry.  The ability to predict a compound’s suitability for such technologies could be critical in drug discovery.  Nanoform’s Starmap2.0 uses sparse-data AI to augment experimental results from its CESS nanoparticle engineering process to predict the potential success of nanoforming drug molecules and enable the development of patient centric medications.

Artificial Intelligence (AI) is becoming widely adopted for small molecule drug discovery, yet the methods for leveraging AI for peptide drug discovery are lagging far behind. These challenges are primarily driven by limited availability of peptide datasets, high dimensionality of the design space as well as poorly developed methods for encoding peptides for deep learning algorithms. In this presentation, we will discuss the advances that were made in developing feature encodings for peptides to enable development of high-performing models and efficient exploration of peptide design space. We will demonstrate how AI can be utilized to prioritize peptide variants for testing through in silico prediction of performance. Moreover, methods to explain the models and visualize feature importance will be presented. Lastly, results from wet-lab validation experiments will be presented in the areas of immunogenicity and antimicrobial peptide development.  

A new approach that combines a heavy in silico front-end with the experimental back-end has emerged. The approach is based on full inclusion of quickly expanding structural and activity data incorporated in structural and AI in silico models. Examples of the application of this approach to finding new antiviral candidates, antiprotozoal candidates, neurological disorders, and aiming at new cancer targets, are presented.

Accurate prediction of absorption, distribution, metabolism, and excretion (ADME) properties can facilitate the identification of promising drug candidates. At Janssen we have developed and implemented generic Target Product Profile (gTPP) models, which predict 15 early ADME properties, at various thresholds and different species where applicable.  Logging of predictions allows monitoring of usage and assessment of impact on decision-making in chemistry.  Model development, performance and benefit will be discussed.

Higher throughput optical biosensors have dramatically reduced the time to run label free binding assays, allowing the more routine analysis of increasing numbers of compounds.  But the innovation in analysis of data has not kept pace with instrument throughput creating a significant burden in time and expertise.  In this presentation we will present a new path that allows the efficient and robust analysis of large data sets from optical biosensors

We have developed a suite of automated tools called MegaSyn (representing 3 components: a new hill-climb algorithm which makes use of SMILES-based RNN generative models, analog generation software, and retrosynthetic analysis coupled with fragment analysis to score molecules for their synthetic feasibility) for designing new molecules. We will describe how it might be used to optimize molecules or prioritize promising lead compounds using test case examples for drug discovery.

Machine learning has been applied with great success to a variety of drug-design problems including binding-affinity prediction, virtual screening, and QSAR, but it is less often used for lead optimization. DeepFrag is a deep convolutional neural network that fills this gap. Given the structure of a protein/ligand complex, DeepFrag (1) predicts the structural properties of a suitable optimizing chemical moiety and (2) searches a database of known moieties to identify those with properties that most closely match the prediction. In this talk, I will discuss recent DeepFrag development and applications.

Aria Pharmaceutical’s unique AI-enabled drug discovery approach can save ~3 years from project initiation to in vivo results while generating a 30x hit rate at those milestones, compared to traditional methods. Aria’s AI platform builds novel pathogenesis models to identify molecules with unique first-in-class MOAs. Using their platform, Aria has developed a pipeline of treatments in 18+ programs across a range of therapeutic areas, including SLE, CKD, NASH, and IPF.

X-Chem’s DNA-Encoded Chemical Library (DEL) platform generates a large amount of binding data for each target it is applied to. In this presentation, we will discuss the application of predictive machine learning to DNA-Encoded Chemical Library technology towards the identification of inhibitors of ERα, one of two main types of receptors activated by estrogen. We will also introduce other applications of machine learning to DEL data currently being developed at X-Chem.

The University of California Drug Discovery Consortium (UCDDC) strives to leverage the biomedical research and commercialization strengths of the University of California (UC) system to accelerate the development of life-saving therapies and to translate discoveries into knowledge driven commercial enterprises that stimulate California’s economy. The UCDDC Executive Committee works with each UC-campus to advance drug development by providing expertise in drug discovery and by building relationships between industry and academics.

The ATOM Consortium aims to accelerate and democratize drug discovery by partnering an open-source molecular design platform with active learning to experimentally validate candidate molecules. We apply this platform to design brain penetrant small molecule inhibitors. Property prediction models are built for blood-brain barrier permeability and efflux. These models are integrated into our generative molecular design software to propose novel molecular entities that achieve our desired design criteria.

The Critical Assessment of Computational Hit-finding Experiment (CACHE), is a new competition to guide the development of virtual screening methods to predict bioactive small molecules. Compounds predicted by participants for pre-selected protein targets will be purchased and tested experimentally at the CACHE experimental hub. All data will be publicly released. Participants will have a choice to remain anonymous. Computational methods will be explained but algorithms may remain proprietary. A new challenge/target will start every 3 months. 

Accelerating the synthesis of newly designed molecules means investing only in synthetic sequences that are more likely to work. Structural diversifications, different substitution patterns, unique heterocyclic systems and arrays of functionalities can lead to unexpected reactivity. Integrating Machine Learning retrosynthetic tools with quantum mechanics calculations for prospective analyses significantly improve success rate in our MedChem syntheses.

Multi-parameter optimization (MPO) is a major challenge confronting the drug discovery process, making it difficult to identify promising synthetically accessible molecules meeting the TPP. To overcome this, we have developed an AI-driven pipeline (Makya and Spaya) for generative de novo drug design to enhance chemical space exploration while tackling the MPO challenge under synthetic constraints. We will discuss the implementation and impact of this pipeline to real-life drug discovery projects.

We describe a novel method that combines experimental biophysical data (HDX-MS) with weighted ensemble simulations (WES) to accurately predict ternary complex structures at atomic resolution. We show that the WES+HDX approach generates accurate structures (RMSD below 2.0 Å to x-ray) and can reproduce solution-state dynamic behavior of the ternary complex. We also show how we are extending this approach to predict degradation propensity of different heterobifunctional and glue molecules.

The potential of providing realistic protein-protein interaction models, and of their assembly with a third (smaller) molecule, such as PROTACs or molecular glue, is, no doubt, a very hot topic. Current molecular modeling approaches have too much bias (and correlation) with the linker degrees of freedom when modeling a PROTAC. At the same time, the lack of large amounts of data severely hinders the application of machine learning techniques. In this talk we will showcase the development of our mixed models, taking advantage of molecular modeling, a consensus bioinformatics approach, and machine learning techniques using data augmentation from modeling. Such a combined approach is capable of enhancing the degrader selection, providing accurate structural interaction models, and screening hundreds of ternary complexes formations.

Targeted Protein Degradation (TPD) is complex and only partially understood. In contrast to orthosteric or allosteric target inhibition, TPD approaches have emerged as novel therapeutic modalities that promote distinct pharmacological profiles and enable modulation of previously undruggable targets. To systematically assess which potential drug targets might be most amenable to TPD therapeutic development, we present a deep learning approach based on a differentiable heterogeneous biomedical knowledge graph built to support data-driven decision-making for TPD portfolio planning and prioritization.