Contemporary Computational Applications and Tools in Drug Discovery

Abstract

In the past decade or so there has been a dramatic increase in the number of computational applications and tools that have been developed to enable medicinal chemists to prosecute modern drug discovery programs more efficiently. The upsurge of user-friendly, well-designed computational tools that enable structure-based drug design (SBDD) and cheminformatics (CI)-based drug design has equipped the medicinal chemist with an arsenal of tools and applications that significantly augments the entire design process, thereby enhancing the speed and efficiency of the design–make–test–analyze cycle. Modern computational applications and tools transcend all areas of drug discovery, and most savvy medicinal chemists can employ them effectively in a myriad of drug discovery applications. Indeed, the sheer scope and breadth of tools available to the medicinal chemist is vast and, to our knowledge, has not been comprehensively reviewed. In this article we have catalogued many computational tools, platforms, and applications that are currently available, with four main areas highlighted: commercially available tools/platforms, open-source applications, internally developed platforms (software tools developed within a pharma or biotech organization), and artificial intelligence/machine learning-based platforms. For ease of interpretation, for these categories we provide tables organized by vendor or organization name, the name of the application, whether the tool/application is employed predominantly for SBDD or CI-based design, and a summary of the main function of the tools, with associated hyperlinks to vendor Web sites. We have tried to be as comprehensive and as inclusive as possible; however, the pace of development of new and existing tools is so rapid that there may be omissions with respect to newly developed tools and current versions of the software.

Computational applications and tools have been adopted to enable and accelerate the discovery of drugs since the early 1980s, regarded as the birth of computer-aided drug design (CADD). Since then, advances in the field have blossomed due to innovative applications of theoretical chemistry and physics, computer science, and statistics and the vast concomitant increase in computational power. On the heels of vast computer power has come the data science evolution that has enabled the derivatization of vast amounts of knowledge, thus enabling the burgeoning fields of bio- and cheminformatics. In the past decade, with the development of user-friendly software, the use of computational tools for drug discovery has shifted from a predominantly CADD specialist role to a reality now where the medicinal chemist routinely employs a spectrum of desktop tools for modern drug design purposes. Indeed, some companies have encouraged, empowered, and trained specialized medicinal chemists to utilize computational tools in a specific “designer” role to potentially streamline the drug discovery process. Whether the designer model is adopted is largely governed by the philosophy and culture of individual organizations, and the value of this approach remains an area of debate among the medicinal chemistry community. Ultimately, though, the fact remains that savvy medicinal chemists embrace state-of-the-art computational tools to effectively augment drug design. In the past decade alone, the number of computational tools that have been developed by vendors has increased significantly, with the spectrum of applications broadening appreciably in addition. Whereas 10 or 15 years ago pharma companies were more apt to develop their own bespoke tools, today the availability of commercial tools has shifted the focus for many companies toward drug design using these user-friendly, robust, well-built, and well-conceived design platforms in an attempt to improve the entire design–make–test–analyze (DMTA) cycle. The optimization of hits to leads to candidates and ultimately approved drugs is very much an iterative process where many parameters must be optimized in parallel. Balancing and optimizing multiple parameters (multiple-parameter optimization (MPO), or scoring) to produce chemical matter that is potent, selective, soluble, permeable, and metabolically stable, and has good oral bioavailability, with an adequate safety profile, is a vastly complex and challenging process which can take many years. Parallel processing is not something that humans are particularly adept at; computers, on the other hand, were built for such purposes. In addition, computers are quite useful at repetitive work where large amounts of data are handled and the likelihood of human error is high. With the advent of machine learning and, more recently, deep learning, efficient computer-based MPO is a reality and, combined with generative chemistry methods, can significantly reduce the number of DMTA cycles, the number of compounds to be synthesized, and more importantly the time and resources it takes for a drug to reach the clinic. Currently no artificial intelligence/machine learning (AI/ML)-derived compounds have made it all the way to the market; however, it is just a matter of time before this eventually transpires. This illustrates how the burgeoning field of computational drug discovery has led to a concomitant growth in innovative computational tools and applications that are now available to drug hunters.

Computational tools have applicability in many areas of drug discovery and development. Many of the applications of these tools toward specific areas of drug discovery have been reviewed several times before¹ and therefore will not be the focus of this article. Instead, for this review we have catalogued many of the available computational applications and tools that have recently emerged and evolved specifically to enable drug discovery, particularly as it relates to the DMTA cycle. Additionally, the focus of the article is centered on tools (Tables 1–5) that enable structure-based drug design (SBDD) and/or cheminformatics (CI)-based design. SBDD facilitates the design of potential drugs utilizing crystallographically derived protein structures or related homology models with and without bound ligands. Examples of such applications are virtual docking/screening and de novo design. CI-based design tools are those that facilitate the design of potential drugs through the curation, manipulation, and analysis of chemical and biological data. Examples of CI applications are chemical similarity calculations and searching, clustering, R-group analysis, and matched molecular pair analysis. Five main areas are highlighted in the tables: Table 1 highlights commercially available tools/platforms; Table 2, internally developed platforms by individual pharma companies; Table 3, open-source tools and software; Table 4, companies whose focus is predominantly on AI/ML applications to drug discovery; and Table 5, tools employed specifically for synthesis predictions using ML methods. Note also that, for AI/ML-based companies in Table 4, we have made the distinction between companies that are predominantly software and service providers and companies that use AI/ML as the prime drug discovery engine and may not necessarily offer access to their platforms/software. For ease of interpretation, these tables are organized by vendor/organization name, the name of the application, whether the tool/application is employed predominantly for SBDD or CI-based design, and a summary of the main function of the tools, with the associated hyperlinks to vendor Web sites—note that many of the descriptions are taken directly from the vendor’s Web site. The assignment of whether the computational tools are predominantly used for SBDD or CI design in some cases is subjective, as the software may have multiple functions that cross both design paradigms. For clear examples of this we have checked both boxes, but there may be instances where the bias between the two paradigms is not as clear and only one box (paradigm) was selected. We have tried to be as comprehensive and as inclusive as possible; however, the pace of development of new and existing tools is so rapid that there may be some omissions with respect to newly developed tools and the current versions of the software.

Table 1. Various Commercial Software Tools Available for Cheminformatics (CI)-Based and Structure-Based Drug Design (SBDD).

vendor	application	CI	SBDD	main function	Web site
Schrödinger	Maestro7		√	workhorse platform for all-purpose molecular modeling	https://www.schrodinger.com/maestro
	AutoQSAR8	√		automated creation and application of predictive QSAR models	https://www.schrodinger.com/autoqsar
	ConfGen9		√	accurate and efficient bioactive conformational searching	https://www.schrodinger.com/confgen
	Desmond10		√	high-performance MD simulations	https://www.schrodinger.com/desmond
	FEP+11		√	high-performance free energy calculations for drug discovery	https://www.schrodinger.com/fep
	Glide12		√	solution for ligand–receptor docking	https://www.schrodinger.com/glide
	Induced Fit13		√	fast and accurate prediction of ligand-induced changes in receptor active sites	https://www.schrodinger.com/induced-fit
	LigPrep14		√	versatile generation of accurate 3D molecular models	https://www.schrodinger.com/ligprep
	MacroModel15		√	molecular modeling platform	https://www.schrodinger.com/macromodel
	Phase16		√	easy-to-use pharmacophore modeling solution for ligand- and structure-based drug design	https://www.schrodinger.com/phase
	PyMOL17		√	high-performance molecular graphics platform for 3D visualization	https://www.schrodinger.com/pymol
	QSite18		√	high-performance QM/MM program	https://www.schrodinger.com/qsite
	Shape Screening19		√	VLS using 3D shape-based similarity	https://www.schrodinger.com/shape-screening
	Core Hopping20		√	comprehensive ligand- and receptor-based scaffold exploration tool for lead optimization	https://www.schrodinger.com/core-hopping
	e-Pharmacophores21		√	energetically optimized structure-based pharmacophores	https://www.schrodinger.com/e-pharmacophores
	Field-Based QSAR22	√	√	discover and optimize new lead compounds using quantitative predictions of binding-site chemistry	https://www.schrodinger.com/field-based-qsar
	Jaguar23		√	rapid ab initio QM electronic structure package	https://www.schrodinger.com/jaguar
	LiveDesign24	√	√	next-generation platform for collaborative drug design	https://www.schrodinger.com/livedesign/drug-discovery
	QikProp25	√		rapid ADME prediction of drug candidates	https://www.schrodinger.com/qikprop
	WaterMap26		√	solution for desolvation thermodynamics	https://www.schrodinger.com/watermap
	SiteMap27		√	fast, accurate, and practical binding site predictor	https://www.schrodinger.com/sitemap
	Canvas28	√		comprehensive cheminformatics computing environment	https://www.schrodinger.com/canvas
	EPIK29	√		fast and robust pKa predictions	https://www.schrodinger.com/epik
	CovDock30		√	work-flow for pose prediction for covalently bound ligands	https://www.schrodinger.com/covdock
Cresset	Forge31	√	√	ligand-based workbench for molecular design and SAR analysis	https://www.cresset-group.com/software/forge/
	Activity Atlas32		√	component of Forge for SAR interpretation and visualization	https://www.cresset-group.com/software/forge-activity-atlas/
	Activity Miner32		√	component of Forge to find and understand activity cliffs	https://www.cresset-group.com/software/forge-activity-miner/
	FieldTemplater32	√	√	component of Forge to generate the most accurate field pharmacophores available	https://www.cresset-group.com/software/field-templater/
	Flare33		√	protein–ligand analysis platform for advanced SBDD	https://www.cresset-group.com/software/flare/
	Spark32	√	√	scaffold hopping and R-group replacement for innovative molecular design	https://www.cresset-group.com/software/spark/
	Blaze34		√	ligand-based virtual screening	https://www.cresset-group.com/software/blaze/
	PickR32	√		advanced electrostatic diversity monomer selection tool for library design	https://www.cresset-group.com/software/pickr/
	Lead Finder35		√	high-throughput docking and scoring for VLS	https://www.cresset-group.com/software/lead-finder/
	Torx32	√		next-generation platform for smallmolecules team-working and collaboration	https://www.cresset-group.com/software/torx/
Open-Eye	cheminformatics tool kits
	FastROCS TK36	√		real-time shape similarity for VLS, lead hopping, and shape clustering	https://www.eyesopen.com/molecular-modeling-fastrocs
	OEChem TK40	√		core chemistry handling and representation	https://www.eyesopen.com/oechem-tk
	OEDepict TK	√		2D molecule rendering and depiction	https://www.eyesopen.com/oedepict-tk
	Grapheme TK	√		Advanced molecule rendering and report generation	https://www.eyesopen.com/grapheme-tk
	GraphSim TK41	√		2D molecular similarity (e.g., fingerprints)	https://www.eyesopen.com/graphsim-tk
	Lexichem TK42	√		name-to-structure and structure-to-name translator	https://www.eyesopen.com/lexichem-tk
	MolProp TK	√		molecular property calculation and filtering	https://www.eyesopen.com/molprop-tk
	Quacpac TK43	√		tautomer enumeration and charge assignment	https://www.eyesopen.com/quacpac-tk
	MedChemTK	√		matched molecular pair analysis, fragmentation, and molecular complexity metrics	https://www.eyesopen.com/oemedchem-tk
	modeling tool kits
	OEDocking TK44		√	molecular docking and scoring	https://www.eyesopen.com/oedocking-tk
	Omega TK45		√	conformer generation	https://www.eyesopen.com/omega-tk
	Shape TK46		√	3D shape description, manipulators, and interrogation	https://www.eyesopen.com/shape-tk
	Spicoli TK		√	surface generation, manipulation, and interrogation	https://www.eyesopen.com/spicoli-tk
	Szmap TK		√	understanding water interactions in binding sites	https://www.eyesopen.com/szmap-tk
	Zap TK47		√	calculate Poisson–Boltzmann electrostatic potentials	https://www.eyesopen.com/zap-tk
	Szybki TK48		√	general purpose optimization with MMFF94	https://www.eyesopen.com/szybki-tk
	Orion	√	√	cloud-based platform to develop custom computational drug discovery workflows and visualizations	https://www.eyesopen.com/orion
BioSolve-IT	SeeSAR49	√	√	visual compound prioritization and evolution; structure-based design work supporting MPO; visualize binding affinity at interfaces	https://www.biosolveit.de/SeeSAR/
	infiniSee50	√		finds molecules chemically similar to given molecule or template; navigates through “almost infinite space” and employs FTrees technology	https://www.biosolveit.de/infiniSee/
	FlexX51		√	structure-based docking; binding mode prediction	https://www.biosolveit.de/FlexX/
	HYDE52	√	√	structure-based scoring; compound classification	https://www.biosolveit.de/HYDE/
	FlexS53	√	√	ligand-based alignment in 3D; prep for 3D QSAR; find scaffold/compound mimics; virtual screening	https://www.biosolveit.de/FlexS/
	FTrees54	√		ligand-based similarity; fuzzy matching to detect novel molecular scaffolds	https://www.biosolveit.de/FTrees/
	ReCore54	√		fragment-based; replace central elements in known bioactive molecules to create new custom 3D scaffold	https://www.biosolveit.de/ReCore/
	FTrees-FS54	√		combinatorial fragment space extension module for FTrees; builds up compounds from fragments while still comparing and searching for similarity	https://www.biosolveit.de/products/#FTrees
	CoLibri54	√		toolkit for chemical space exploration; generate chemical space; focus on synthetically accessible entities	https://www.biosolveit.de/CoLibri/
	REAL Space Navigator	√		collaboration with Enamine, “world’s largest, ultrafast” searchable chemical space; 13 billion compounds	https://www.biosolveit.de/REALSpaceNavigator/
	GalaXi	√		explore chemical space commercially available on demand from Wuxi	https://www.biosolveit.de/2019/10/10/launch-of-galaxi/
	KnowledgeSpace	√		virtual chemical space containing 100+ literature reactions to produce likely synthetically accessible compounds	https://www.biosolveit.de/CoLibri/spaces.html
Dotmatics	Browser	√		searching and browsing across all corporate scientific databases	https://www.dotmatics.com/products/browser
	D40	√		“Dotmatics for Microsoft Office” connector to integrate with Microsoft Office suite	https://www.dotmatics.com/products
	Blueprint	√		Web-based platform for visualization and scientific data analysis for small-molecule design	https://www.dotmatics.com/products/blueprint
	Vortex	√		chemically intuitive data visualization and analysis’ expansion on the “spreadsheet”; can do statistical analysis and create visualizations	https://www.dotmatics.com/products/vortex/
Discngine	Assay	√		design and manage cellular, molecular, and high-content screening campaigns	https://www.discngine.com/assay
	3Decision		√	store, analyze, and share protein–ligand structures, sequences, and associated data	https://3decision.discngine.com/
	Chemistry Collection	√		generate and visualize scaffold networks, perform molecular fragmentation, or use Pharmacophore Graph to design matched molecule pairs analysis	https://www.discngine.com/chemistry-collection
	Network Collection	√		represent, manage, and analyze complex network structures using advanced graph theory	https://www.discngine.com/network-collection
Optibrium	Stardrop	√		small-molecule design, optimization, and data analysis; can do predictive modeling and QSAR	https://www.optibrium.com/stardrop/stardrop-features.php
	Asteris	√		iPad app for drug discovery	https://www.asteris-app.com/
	Ocura	√		load and browse molecular structures; data viewer	https://www.optibrium.com/ocura/
TIBCO	Spotfire	√		AI-powered search-driven experience; built-in data wrangling and advanced analytics	https://www.tibco.com/products/tibco-spotfire
	TIBCO Data Science	√		share and deploy analytics models across organizations for increased collaboration	https://www.tibco.com/products/data-science
Dassault Systemes	Biovia Pipeline Pilot	√		data pipelining tool for creating simple as well as advanced workflows for various data science initiatives such as cheminformatics, bioinformatics, ML, etc.	https://www.3ds.com/products-services/biovia/products/data-science/pipeline-pilot/
	Discovery Studio		√	data pipelining tool for advanced 3D design such as molecular mechanics, free energy calculations, biotherapeutics developability, and more in a common environment	https://www.3ds.com/products-services/biovia/products/molecular-modeling-simulation/biovia-discovery-studio/
CCDC	mercury		√	knowledge-based geometry and interaction preferences	https://www.ccdc.cam.ac.uk/mercury/
	GOLD		√	protein–ligand docking	https://www.ccdc.cam.ac.uk/solutions/csd-discovery/components/gold/
	CSD-CrossMiner	√		discovery-oriented data mining of protein and small-molecule crystal structures	https://www.ccdc.cam.ac.uk/solutions/csd-discovery/components/csd-crossminer
	SuperStar	√		knowledge-based pharmacophore prediction	https://www.ccdc.cam.ac.uk/solutions/csd-discovery/components/superstar/
	CSD Python API	√		Python-based programmatic access to the tools	https://www.ccdc.cam.ac.uk/solutions/csd-materials/components/csd-python-api/
	Mogul		√	knowledge-based library of molecular geometry derived from the Cambridge Structural Database (CSD)	https://www.ccdc.cam.ac.uk/solutions/csd-system/components/mogul/
CCG	Molecular Operating Environment (MOE)55		√	integrated computer-aided molecular design platform; includes 3D visualization, structure-based design, SAR explorer, data modeling, virtual screening, simulations, QSAR, etc. for small molecules as well as antibody–drug conjugates and biologics	https://www.chemcomp.com/Products.htm
	MOESaic55	√		allows small-molecule SAR analysis such as MMPs, R-group, etc.	https://www.chemcomp.com/Products.htm#MOEsaic-SAR_Explorer
	PSILO55		√	data repository and visualization application to store curated RCSB protein structures, pocket similarity, and other features for data analysis	https://www.chemcomp.com/Products.htm#PSILO-Structure_Database
Medchemica	MCPairs		√	AI/ML-based SAR analysis application to generate match pairs and new ideas	https://www.medchemica.com/products/
ChemAlive	ConstruQt	√		high-throughput quantum chemistry for molecular design; allows automated library-scale deployment of quantum chemical calculations	https://www.chemalive.com/construqt/
CDD Vault	CDD Vault	√		informatics platform; molecular/data viewing and browsing	https://www.collaborativedrug.com/benefits/
BioSymetrics	Augusta	√		biomedical-specific ML framework; drug discovery through small-molecule activity prediction	https://www.biosymetrics.com/products/moa-prediction/
Eidogen Sertanty	Target Informatics Platform (TIP)	√		interrogate the druggable genome from a structural perspective; bridge the gap between bio- and cheminformatics	https://www.eidogen.com/tip.php
	TIP Calculation Engine (STRUCTFAST, SiteSeeker, SiteSorter, SLiC)		√	structure determination; binding site annotation; similarity assessment	https://www.eidogen.com/tae.php
	Eidogen Visualization Environment (EVE)		√	visualize and compare small-molecule binding sites of targets of interest	https://www.eidogen.com/eve.php
	Oncology Knowledgebase (OKB)	√		database of structure–activity data across targets of oncological interest	https://www.eidogen.com/oncologykb.php
	Kinase Knowledgebase (KKB)	√		database of kinase structure–activity and chemical synthesis data	https://www.eidogen.com/kinasekb.php
	ChIP	√		algorithm-driven enumeration engine for automated generation of medicinally relevant, novel molecules with proven synthetic access	https://www.eidogen.com/chip.php
VeraChem	VM2		√	software package for molecular-binding free energy calculations	http://www.verachem.com/products/vm2
	VConf	√		2D-to-3D small-molecule conversion	http://www.verachem.com/products/vconf
	VCharge		√	accurate partial atomic charges of drug-like molecules	http://www.verachem.com/products/vcharge
	VFilter	√	√	analyze ensembles of molecular conformations and remove repeats	http://www.verachem.com/products/vfilter
	Vrms	√	√	provides symmetry-corrected root-mean-square deviation between molecular conformers	https://www.verachem.com/products/vrms/
	VDisplay		√	3D molecular viewer	https://www.verachem.com/products/vdisplay/
NextMove	Arthor56	√		fast, state-of-the-art substructure and chemical similarity search capabilities for ultra-large databases of 100s of millions of compounds, using SMARTS optimization, Just-In-Time compilation, and/or GPUs	https://www.nextmovesoftware.com/arthor.html
	Matsy57	√		set of tools for creating and analyzing matched molecular series (the general form of MMPS); In particular, can be used to suggest what compound to make next in a med-chem program	https://www.nextmovesoftware.com/matsy.html
	MPSearch57	√		rapidly searches a database to find matched pairs related to a query molecule This type of search is used to explore previous medicinal chemistry strategies	https://www.nextmovesoftware.com/mpsearch.html
	Patsy58	√		speed up SMARTS pattern matching by creating optimized SMARTS patterns of source code; speed gains are particularly large when multiple SMARTS patterns are matched against a single structure	https://www.nextmovesoftware.com/patsy.html
	SmallWorld59	√		index of chemical space based on more than 230 billion substructures; can be used to measure similarity based on graph-edit distance, find the maximum common subgraph of two or more molecules, analyze HTS results, and more	https://www.nextmovesoftware.com/smallworld.html
	CaffeineFix60	√		rapidly match chemical names or terms against a dictionary of grammar (e.g., a grammar for IUPAC names); use in text mining; can be used to provide autocomplete functionality and spell-correction	https://www.nextmovesoftware.com/caffeinefix.html
	LeadMine61	√		extracts chemical names and terms from text; incorporates CaffeineFix technology to find terms that match appropriate dictionaries or grammars; has enhanced functionality to handle patent literature	https://www.nextmovesoftware.com/leadmine.html
	Casandra62	√		server for delivering real-time safety warning of experimental hazards straight to pharmaceutical ELNs	https://www.nextmovesoftware.com/casandra.html
	HazELNut63	√		suite of tools used to extract, normalize, and analyze information in ELNs; can be used to implement a search interface, find/eliminate duplicates, find similar reactions, etc.	https://www.nextmovesoftware.com/hazelnut.html
	NameRxn64	√		used to classify and name reactions; particularly useful in the context of ELN analysis but also as a plug-in to chemical drawing software; builds on NextMove Patsy technology	https://www.nextmovesoftware.com/namerxn.html
	Pistachio65	√		reaction dataset browser providing loading, querying, and analytics of chemical reactions; with over 9 million chemical reactions extracted from U.S. and EPO patents, it demonstrates an AI interface to faceted (structure) search	https://www.nextmovesoftware.com/pistachio.html
Alvascience	alvaMolecule	√		visual analytics platform allowing users to standardize and organize chemistry project data	https://www.alvascience.com/alvamolecule/
	alvaModel	√		build and deploy QSAR/QSPR regression models; consists of two pieces of software: alvaModel and alvaRunner	https://www.alvascience.com/alvamodel/
	alvaDesc	√		allows calculation of over 5000 0D/1D/2D/3D molecular descriptors	https://www.alvascience.com/alvadesc/
	alvaBuilder	√		molecule design platform utilizing user-selected property criteria	https://www.alvascience.com/alvabuilder/
Datagrok	Datagrok for Cheminformatics	√		chemistry data visualization and analytics; provides a structure/substructure search as well as mechanism to build ML models	https://datagrok.ai/cheminformatics
Molsoft	ICM Chemist Pro	√		allows scientists to draw and edit chemicals, create and view chemical spreadsheets, and perform chemical searching, chemical clustering, and many other routine cheminformatics functionalities	https://www.molsoft.com/icm-chemist-pro.html
	ICM Pro		√	desktop software environment for high-quality protein structure analysis, modeling, and docking	https://www.molsoft.com/icm_pro.html

Table 5. Leading Cutting-Edge AI/ML-Based Computer-Aided Synthetic Prediction (CASP) Organizations Providing Innovation in Drug Discovery.

vendor	application	main function	Web site
ChemPass	ChemPass	rule-based AI for forward reaction-based design	https://chempassltd.com/
DeepMatter	ICSynth	ML-based method to generate chemistry rules	https://www.deepmatter.io/products/icsynth/
Sigma-Aldrich/Merck KGaA	SYNTHIA (previously known as Chematica)	algorithms using a database of manually generated rules	https://www.sigmaaldrich.com/US/en/services/software-and-digital-platforms/synthia-retrosynthesis-software?gclid=EAIaIQobChMI94G7_Zqn8gIVE73ICh1yrQjqEAAYASAAEgJZxfD_BwE
IBM	IBM RXN	natural language-based model on reaction data sets	https://www.research.ibm.com/blog/rxn-cleaning-chemical-datasets
CAS	ChemPlanner (previously known as ARChem)	method-based or rule-based as well as ML algorithms	https://www.cas.org/solutions/cas-scifinder-discovery-platform/cas-scifinder/retrosynthesis-planning
Elsevier	Reaxys	algorithms using a database of manually generated rules	https://www.elsevier.com/solutions/reaxys
AstraZeneca	AiZynthFinder	open-source MonteCarlo tree-based search using ANN	https://github.com/MolecularAI/aizynthfinder
Eli Lilly Co.	LillyMol	ML-based method using model trained on reaction transformations	https://github.com/EliLillyCo/LillyMol
MIT Consortia	ASKCOS	a multifaceted AI/ML-based retrosynthetic analysis tool developed at the Massachusetts Institute of Technology	https://askcos.mit.edu/help/modules
Iktos	Spaya.ai	AI-fueled tool to discover and prioritize synthetic routes	https://iktos.ai/spaya/
University of Notre Dame	C-CAS	provides synthetic chemistry prediction using quantitative, data-driven approaches	https://ccas.nd.edu/research/#thrust-3

Table 2. Various Pharma Proprietary Software Tools Available for Cheminformatics (CI)- and Structure-Based Drug Design (SBDD).

vendor	application	CI	SBDD	main function	Web site
AbbVie	AIDEAS	√	√	AIDEAS (AbbVie’s Integrated Design Explorer and Analytics Solution) is a one-stop shop for all cheminformatics tools such as phys-chem property calculation, ADME ML calculations, library enumeration, docking, R-group/MMP analysis, patent data extraction, clustering, etc.	https://www.bio-itworldexpo.com/software (Oral presentation)
Vertex	ASAP		√	ASAP (Affinity, Selectivity, Activity, Properties and PK) is built using OpenEye toolkits and ChemAxon, for searching and exploring chemical and biological data	https://chemaxon.com/presentation/asap-emphasizing-multidimensional-drug-discovery
	MedChem2	√		a modeling and design tool that integrates a number of cheminformatics tools such as molecular superposition, protein–ligand docking, property calculation, library design etc.	https://link.springer.com/article/10.1007/s10822-016-9994-0 (ref 1a)
Pfizer	PCAT	√		a tool for clustering, organizing, and visualizing molecules with their associated properties and biological activities	https://pubs.acs.org/doi/pdf/10.1021/ci9002443?src=recsys (ref 66)
	PGVL			virtual library design-based platform	https://pubmed.ncbi.nlm.nih.gov/23020747/ (ref 67)
	CCT			?	https://link.springer.com/article/10.1007/s10822-016-9997-x (ref 68)
GlaxoSmithKline	BRADSHAW	√		BRADSHAW (Biological Response Analysis and Design System using an Heterogenous, Automated Workflow) is GSK’s experimental automated design environment	https://link.springer.com/article/10.1007/s10822-019-00234-8
Novartis	FOCUS	√		platform to produce, visualize, and share information on various aspects of a drug discovery project, such as cheminformatics, data analysis, structural information, and design; built on MolSofts ICM	https://pubs.acs.org/doi/abs/10.1021/ci500598e (ref 69)

Table 3. Various Open-Source Software Tools Available for Cheminformatics (CI)- and Structure-Based Drug Design (SBDD).

vendor	application	CI	SBDD	main function	Web site
OpenMolecules	DataWarrior		√	open-source cheminformatics application; includes several data analysis and visualization methods such as clustering, phys-chem property calculations, large-scale similarity calculations, dimensionality reduction, etc.	http://www.openmolecules.org/datawarrior/ (ref 70)
MayaChem	MayaChem Tools	√	√	open-source package for property computation, library enumeration, R-group decomposition, Python scripts for PyMOL plug-ins, etc.	http://www.mayachemtools.org/ (ref 71)
KNIME	KNIME	√		open-source software for data science innovation; create workflows in GUI-like pipeline pilot	https://www.knime.com/
AutoDock	AutoDock Vina		√	open-source suite of automated docking tools; designed to predict how small molecules, such as substrates or drug candidates, bind to a receptor of known 3D structure	http://vina.scripps.edu/ (ref 72)
Open Drug Discovery Toolkit	ODDT	√		open-source modular toolkit written in Python for cheminformatics and chemical modeling	https://jcheminf.biomedcentral.com/articles/10.1186/s13321-015-0078-2 (ref 73)
Source Forge	Python Prescription (PyRx)	√	√	virtual screening software for computational drug discovery; screen libraries against potential targets; includes docking wizard (uses AutoDock Vina) and built on Python	https://pyrx.sourceforge.io/

Table 4. Leading Cutting-Edge AI/ML-Based Organizations Providing Innovation in Drug Discovery.

vendor	application	CI	SBDD	main function	Web site
Exscientiaa	Centaur Chemist	√	√	uses AI to learn best-practices from drug discovery data and experienced drug hunters	https://www.exscientia.ai/
Insilico Medicine	Chemistry42	√	√	AI/DL generative platform for de novo drug design	https://insilico.com/chemistry42
Atomwisea	Services Options	√	√	utilizes deep-learning AI technology for structure-based small-molecule drug discovery.	https://www.atomwise.com/ (ref (74))
Cyclica	Ligand Design, Ligand Express	√	√	apply a combination of computational biophysics and ML to design novel drugs	https://www.cyclicarx.com/ (ref 75)
XtalPi	RENOVA	√	√	generate novel drug-like molecules by combining QM and AI; predict crystal formation of small molecules	https://www.xtalpi.com/en/
Recursion	ReChem, RePredict, ReAnalyze	√	√	develop novel chemical compounds by utilizing AI to conduct experimental biology, at scale, by testing thousands of compounds on hundreds of cellular disease models in parallel	https://www.recursion.com/
IBMa	Generative Models	√	√	develop new AI methods for de novo compound generation as well as methods for retrosynthetic analysis to accelerate drug discovery	https://research.ibm.com/science/generative-models/
1910 Geneticsa	ELVIS, ROSALYND	√	√	discover new drugs using computation, physics, and ML; develop a robust in silico platform utilizing AI/ML methods	https://1910genetics.com/
Insitroa		√	√	Insitro is a drug discovery and development startup that utilizes ML and biology to transform drug discovery	https://insitro.com/
OneThree Biotecha		√	√	combining systems biology with AI to uncover new insights and build next-generation drug discovery	https://onethree.bio/ (ref (76))
Iktos	Makya, Spaya	√	√	AI/DL platforms for generative de novo drug design and retrosynthetic analysis	https://iktos.ai/
DeepMind	AlphaFold		√	uses AI to accelerate drug discovery; discovered one of the most successful protein structure prediction algorithms, Alphafold 2.0	https://deepmind.com/
ACELLERA	PlayMolecule	√	√	one-click molecular discovery platform consisting of innovative MD- and DL-based tools that transcend the drug discovery continuum design paradigm	https://www.acellera.com/products/playmolecule/
Glamorous AI (now X-Chem)	ROSALINDAI	√	√	SaaS platform for AI/ML-based drug discovery; largest repository of state-of-the-art AI/ML-based models employed for applications including de novo generative design	https://www.glamorous.ai/rosalindai
BenevolentAIa	The Benevolent Platform	√	√	knowledge graph-based platform for drug discovery	https://www.benevolent.com/

^aDrug discovery companies which may not offer access to AI/ML tools and applications.

As we consider the drug discovery continuum from Hit ID → Hit to Lead → Lead Optimization (see Scheme 1), computational tools are essential for of all these stages of drug discovery. For Hit ID, for example, there are many approaches to virtual screening that use both protein-based and ligand-based computational approaches. There are pros and cons for both of these approaches; however, there have been significant innovations and enhancements in recent years, with many vendors offering tailored searching/docking algorithms that often identify both complementary (similar) and orthogonal (dissimilar) hits.

Scheme 1. Some Key Areas along the Drug Discovery Continuum Where Computational Applications and Tools Are Employed.

At the early stages of a project, where screening is required for Hit ID, it is essential to understand the scope, breadth, and extent of the chemical space that one needs to explore to identify tractable chemical matter. Corporate screening collections are finite and have numbers that do not adequately sample the vastness of drug-like chemical space. Even with the introduction of DNA encoded library (DEL) technologies, which has dramatically expanded synthetically accessible chemical space into the billions, both curating and exploring ultra-vast chemical space can only be achieved using innovative computational methods. There are many ways to approach building virtual libraries, either using fully enumerated structures with large but limited numbers (10⁶) or with topological/pharmacophoric feature-encoded fragmented molecular representations, such as Feature Trees (FTrees⁵⁴), that enable the virtual expansion of chemical space to unprecedented levels (10²⁶).² There are undoubtedly pros and cons associated with these approaches, but they have been effective at Hit ID, employed either individually or in concert. The exploration of chemical space is enabled using computational methods, with vendors offering a variety of tools to enable searching, clustering, similarity assessments, and mapping (dimensionality reduction), to mention a few. Note that virtual screening can be employed at any stage of the drug discovery continuum, often to augment existing hit series and to identify new orthogonal tractable chemical series with the goal of increasing the probability of success. Indeed, if a project is enabled with a robust protein crystal structure, virtual docking is arguably the main vehicle for the optimization of ligand binding. Without structure, but equipped with a known binder, ligand-based virtual screening or lead/scaffold hopping can be used to identify new series to pursue. Fortunately, there are a growing number of computational tools from vendors that enable both ligand-based virtual screening and lead/scaffold hopping employing very innovative approaches that have been used successfully, particularly in Hit ID, but also in the Hit to Lead phase of the drug discovery continuum.

The Hit to Lead and Lead Optimization stages of the discovery continuum are complex and challenging paradigms, where many different properties of lead series must be optimized concurrently and iteratively to achieve the desired efficacious in vivo (pharmacokinetics/pharmacodynamics, PK/PD) profile in a relevant disease model. This may be achieved efficiently by adopting a MPO approach, a process that is becoming mainstream in contemporary medicinal chemistry optimization campaigns.³ Therefore, in concert with the optimization of binding and functional activity, concurrent physicochemical property-based optimization is essential for the success of any medicinal chemistry campaign. To achieve this, a mix of both in vitro and in silico parameters must be employed to derive multi-parametric scoring functions tailored to an individual project. A growing number of computational design platforms are enabled to perform MPO using a range of tailored methods; in addition, vendors now offer this capability employing a variety of mathematical approaches.

Oftentimes, particularly in Lead Optimization, there is sufficient data to build robust quantitative structure–activity relationship (QSAR) models which may significantly enable a project team by augmenting the optimization of a lead to the desired outcome of candidate selection. Additionally, as most pharmaceutical companies have vast repositories of absorption, distribution, metabolism, excretion, and toxicity (ADMET) data, quantitative structure–property relationship (QSPR) models using a variety of deep-learning (DL)/machine-learning (ML) methods are now par for the course, with the predicted ADMET properties often calculated after registration of new chemical entities prior to capturing initial primary in vitro ADMET assay data. Models can then be augmented and improved continuously through machine-based learning methods employing the newly added data, with accuracy often improved for a given similar series. Many pharmaceutical and biotech companies are enabled with IT infrastructure that facilitates many of the computational methods needed to sufficiently augment and improve the drug discovery process. The extent and utility of these methods depends on the individual companies, with a mix of both internally and externally developed applications and platforms adopted. In the past couple of years, the power of computational drug discovery has been demonstrated, and the concomitant interest in the medicinal chemistry community has been heightened by the development of generative de novo design methods (Table 2). Using DL models, these methods can generate novel chemical entities which are optimized, in silico, to possess targeted affinity, PK, and even safety profiles. Moreover, AI-based platforms can learn which classes of chemical entities bind selective protein targets and generate associated novel chemical matter capable of binding and functionally modulating the protein target of interest, ultimately leading to the optimization of leads to clinical candidates very rapidly.⁴ Of course, the success of generative platforms is directly proportional to the extent and accuracy of the data used in the training sets for both the generation and optimization phases. Typically, therefore, this approach works most effectively with well-trodden targets such as kinases, where a wealth of robust data is at hand to enable more accurate predictions and the most promising in silico de novo designs.⁵ Drugging less tractable therapeutic targets, however, by targeting more challenging targets such as PROTACS and protein–protein interactions (PPIs) using this approach is significantly more challenging, as is the case for traditional medicinal chemistry approaches. However, the rate of change in the development of these generative platforms is so rapid that it will be very interesting to see how far these evolve in the coming years and whether they are able to generate novel chemical matter that can drug these challenging, high-value targets.

In addition to discussing generative models, as well as a plethora of software tools developed to generate and test a variety of hypotheses, it is important to consider methods related to prediction of chemical synthesis.⁶ Over the past few years, several new tools have emerged (Table 3) that allow medicinal chemistry teams to quickly assess the synthesis of difficult-to-synthesize or novel chemotypes. These approaches range from ML-based methods trained on large reaction datasets (e.g., USPTO dataset) to hand-coded rules. Some of these methods are commercial, and a small handful of them are open-source (e.g., AiZynthFinder) or part of a consortium (e.g., ASKCOS). The overall outlook is favorable for these tools, and this will continue due to the continuous addition of new reaction datasets. With these tools now emerging online, it is important that each pharma company ensure that clean and reliable electronic lab notebook (ELN) data, utilizing both positive and negative data (failed reactions), is employed to train models. In addition, combining internal together with publicly available data may improve the overall quality of the models. A variety of commercial software providers are already moving in this direction, i.e., allowing the retraining of models with both internal and external datasets.

One important area of debate currently is whether to implement Web-based or desktop-based software, a deployment issue that leads to multiple concerns such as optimal machine configuration, security, etc. Routine software upgrades and updates can be a significant challenge for IT organizations to implement for each scientist, while understanding the implications of these upgrades on other software. In the past few years, most cheminformatics and molecular modeling applications have been moving to Web-based platforms. In addition, hosting applications on the “cloud” not only makes them easy for the user to access but also simplifies application deployment for the IT organizations. Over the past decade, with computer resources becoming relatively inexpensive, newer methods are being continuously developed and adopted that realize the potential of computationally intensive calculations. This raises an important issue about running some of the intensive calculations on high-performance computing (HPC) facilities available on-premises versus hosted on external infrastructures (e.g., rented within academic institutions or cloud providers). The decision to build an in-house HPC versus other options may provide a huge barrier to software acquisition from vendors and its deployment. Based on the usage and the compute nature of the software, pharma organizations should carefully plan the right architecture for software deployment, as it directly affects the users as well the IT organizations responsible for managing such software. Another aspect that is important to consider is software-as-a-service (SaaS), platform-as-a-service (PaaS), or on-premises installation of software or technology. In the past several years, many software providers have offered their solutions as either SaaS or PaaS. Pharma organizations, when making the decision to license or utilize a solution, must review this prior to capitalizing on the technology.

Both the number and the degree of computational applications and tools available to the medicinal chemist is expanding rapidly, with the rate expected to continue in the coming years. Computational tools are integral to the drug discovery continuum, and they have largely transformed the way medicinal chemistry is practiced, particularly in the past decade. In the tables we have highlighted well over 150 tools or applications from many different vendors, with over 50 focused predominantly on SBDD and over 80 focused on CI-based applications, with at least 10 having utility for both. Many of the vendors mentioned have relationships with most of the top pharmaceutical companies, and thus the drug discovery industry shares many of the same tools. It is unlikely, however, that any one organization has access to all the tools mentioned in the tables; therefore, we hope this will give a perspective on the landscape of computational tools as it stands presently. We know this list will likely be outdated by the time this article is published, but we feel that many important computational applications have been captured and that these tools will continue to help the medicinal chemist successfully navigate the drug discovery continuum for years to come.

Glossary

Abbreviations

CADD

computer-aided drug design

DMTA

design–make–test–analyze

AI/ML

artificial intelligence/machine learning

SBDD

structure-based drug design

CI

cheminformatics

PK/PD

pharmacokinetics/pharmacodynamics

MPO

multiple-parameter optimization

QSAR

quantitative structure–activity relationship

QSPR

quantitative structure–property relationship

ADMET

absorption, distribution, metabolism, excretion, and toxicity

PPI

protein–protein interaction

USPTO

United States Patent and Trademarks Office

ASKCOS

automated system for knowledge-based continuous organic synthesis

ELN

electronic laboratory notebook

HPC

high-performance computing

SaaS

software as a service

PaaS

platform as a service

VLS

virtual ligand screening

QM/MM

quantum mechanics/molecular mechanics

Author Present Address

^# Novartis Institutes for BioMedical Research, 181 Massachusetts Ave., Cambridge, MA 02139, USA

The authors declare the following competing financial interest(s): Both P.C. and R.G. are/were employees of AbbVie at the time of this study. The design, study content, and financial support for the study were provided by AbbVie. AbbVie participated in the interpretation of data, review, and approval of the publication.