EBOOK

The Little Book Of Big

Changes In AI-Powered
Drug Discovery
A somewhat comprehensive rundown of all the
ways AI is transforming drug discovery and the
future of biomedical research.
Contents

Introduction Quality Data

AI Has Changed the Game
Impact on Drug Discovery About DrugBank

AI Advancements Appendix
AI and the Drug Discovery Pipeline Resources

Big Deal Breakthroughs References

Research
Commercial Cases
Insilico Medicine
Celeris Therapeutics
Cyclica

Academic Cases
DeepCE
DOCKSTRING
EGGNs
Artificial Intelligence
Has Changed The Game

AI has launched itself from the pages of science At this point if you aren’t using AI for drug discovery and repurposing, you’re
setting yourself up to be left behind. As the pace of innovation continues to
fiction and disrupted our industry. surge, hundreds of startups are centering their entire organization around AI
and seeing astounding results. Pharma companies aren’t being shy either, with
numerous big names quickly shifting to integrate AI into their practices.

Everyone, at every level of your organization, needs to be thinking about

Startups using AI1 applying AI as a means to reduce costs, increase speed, and ensure your time
and efforts are in pursuit of successful drug candidates and treatments.

Pharma companies using AI2

Drugs developed as a result of AI3

Invested in companies &

partnerships leveraging AI4

EBOOK: THE LITTLE BOOK OF BIG CHANGES IN AI-POWERED DRUG DISCOVERY 1

The Impact On Drug
Discovery Is Undeniable

There are currently more than 100 drugs in the As AI increasingly becomes standard practice, access to high-quality data
is proving to be a true differentiator. You’ve likely heard, if not already said
AI in drug discovery pipeline, and numerous “garbage in, garbage out,” when it comes to drug data. As the move to AI

drugs have been making their way to clinical continues to advance it will only become more vital to have reliable, clean, and
robust datasets. Without this, you’ll find yourself spending more and more time
trials in a matter of years — instead of decades. getting datasets ready before you can even begin your research.

Yet, data is growing at such an exponential rate that it is becoming impossible

to keep up with, maintain, structure, and normalise it at a pace and standard

Scientific journals published each year5 that enables anyone to use it to its fullest potential.

At DrugBank we work tirelessly to equip leading data

scientists with the most in-depth, highest-quality, and up-
Of systematic reviews are inaccurate
to-date drug data on the market.
within 24hrs of publication6
As a result we get to see firsthand how the industry and new technology is
transforming the future of medicine and healthcare.
Of reviews not updated in 2 years
have incorrect conclusions6

EBOOK: THE LITTLE BOOK OF BIG CHANGES IN AI-POWERED DRUG DISCOVERY 2

AI-Advancements
Over the past few decades, every facet of The exponential growth and improved coverage and availability of omics-
datasets has pushed machine-learning-powered computational methods to
the drug discovery pipeline has in one way or become a critical tool in integrating, understanding, and utilizing these datasets

another been disrupted by the move towards to their fullest potential.

AI-guided approaches.

The AI in drug discovery timeline is matter of months — without AI is 5+ years7

EBOOK: THE LITTLE BOOK OF BIG CHANGES IN AI-POWERED DRUG DISCOVERY 3

Target identification efforts have been hugely improved through the integration These advances in ML and experimental methods represent the starting point
of varying, often heterogeneous datasets such as pathways, RNA expression, for new approaches to AI-focused drug discovery. The explosion in available high-
animal models, mutations, and somatic and germline genetics. The machine quality 3D protein structures and the continued advancements in structured,
learning approaches that are making this possible continue to be increasingly machine-learning-ready drug data will open up new approaches, and demand
successful in finding targets for both new and existing diseases. entirely new optimizations to handle the intersection between the growing
protein structure and chemical compound spaces.
The application of knowledge graphs that integrate drugs, targets, diseases,
pathways, and other entities, as well as their relationships, have led to a number Additionally, the COVID-19 pandemic opened the door to many new innovative,
of promising new approaches. collaborative research projects focused on applying machine learning to better
understand the SARS-CoV-2 virus, as well as identify novel antiviral drug candidates.
One example that is still in early stages and currently seeing
an intense research focus is the use of knowledge graphs to
capture a more representative model of structural biology.

Big Deal Breakthroughs

Although now well-known in the drug-discovery world, the release of protein
structure predictions from AlphaFold will have wide-ranging implications across
the AI in drug discovery field. Beyond AlphaFold, there have also been significant
advances in cryo-electron microscopy. Combining these advances introduces a
new model for structural biology8, one which may well become a fertile ground
for drug discovery research over the coming decade.

The gaps that exist in AlphaFold predictions (namely conformational dynamics,

and disordered proteins and regions) align well with the areas where cryo-EM /
cryo-ET provide useful information. The combination of AlphaFold predictions
with the information on tomography being realized with cryo-EM technology9
will lead to near-atomic-resolution models of complexes, in their physiological
context, inside the cell.

EBOOK: THE LITTLE BOOK OF BIG CHANGES IN AI-POWERED DRUG DISCOVERY 4

 Research Worth
Talking About
AI is making a huge impact across the commercial and
academic research space. Below we explore six cases where
recent key advances in AI-powered drug discovery have
resulted in innovative tools and exciting research findings.

COMMERCIAL

Insilico Medicine is a Hong Kong, China based

company specializing in the development of
new AI technologies.

WHAT THEY’RE UP TO

Insilico has made it their mission to accelerate drug discovery and drug
development by continuously inventing and deploying new artificial intelligence
technologies. Nearly a decade old, they now have several oncology candidates in
their pipeline, and are pursuing the development of both drugs and biomarkers
in areas ranging from fibrosis, infectious diseases, immunology, and the process
of aging.

EBOOK: THE LITTLE BOOK OF BIG CHANGES IN AI-POWERED DRUG DISCOVERY 5

HOW THEY’RE USING AI WHY WE’RE SO IMPRESSED

One of the most noteworthy advancements Insilico has developed involved Insilico’s DDR1 research was able to save significant development time, completing
applying a generative pipeline to complete hit discovery, optimization, synthesis, the process in 46 days which was 15-fold faster than traditional approaches.
and validation on candidates against discoidin domain receptor 1 (DDR1), a
kinase target implicated in fibrosis and other diseases. Insilico was also able to identify the genes most important
for age prediction11, achieving Pearson correlation of 0.91
This approach uses a two-step algorithm. The first step involves learning a
for the actual age values of the muscle tissue samples.
mapping of the chemical space; the second step explores this mapping using
their proprietary deep reinforcement learning platform GENTRL (General Tensorial
Reinforcement Learning) to learn DDR1 and common kinase inhibitors. GENTRL
utilized three distinct Kohonen-based self-organizing maps (SOMs) as reward
functions for the reinforcement learning step: the trending SOM (scores compound
novelty based on patent disclosure dates), the general kinase SOM (distinguishes
kinase inhibitors), and the specific kinase SOM (isolates DDR1 inhibitors).

This approach identified four active compounds10, two active in cellular assays,
and one lead candidate that demonstrated favorable pharmacokinetics in mice.

Insilico has also advanced our understanding of aging. Applying several supervised
machine learning approaches, including neural networks, Insilico built a panel of
tissue-specific biomarkers of aging.

EBOOK: THE LITTLE BOOK OF BIG CHANGES IN AI-POWERED DRUG DISCOVERY 6

 Celeris Therapeutics is a deep learning company focused on developing therapeutics that work to degrade disease-producing proteins.
Based in the US, they also have a presence in Austria.

WHAT THEY’RE UP TO

Celeris Therapeutics focuses on undruggable pathogenic proteins that cause Applying these approaches and models, Celeris was able to identify novel West
serious conditions such as Alzheimer’s and Parkinson’s disease. Their current Nile Virus NS2B/NS3 protease inhibitors13. The viral NS2B/NS3 protease is critical
pipeline includes programs in neurology and oncology. Celeris is also using graph to the viral replication process. Using these deep learning approaches, Celeris was
neural networks to predict the properties of molecules. able to identify novel, unexplored drug candidates that demonstrate an inhibition
score statistically neighboring experimentally confirmed inhibitors, presenting new
HOW THEY’RE USING AI
candidates for treating West Nile Virus.
In their Xanthos Match Maker platform, Celeris encodes molecular structures
WHY WE’RE SO IMPRESSED
in a graph along with features such as the number of hydrogens, valence, and
aromaticity and then applies deep neural networks where information about Drugs are currently limited in their ability to treat diseases caused by
molecules and proteins are processed into an increasingly high-level form. pathogenic proteins.

To make a molecular graph more performant, Celeris uses additional ML By establishing reliable, AI-backed methods to leverage
techniques to improve molecular fragment linking (linking two fragments binding
the body’s natural cell-based mechanisms to degrade
in nearby subpockets together has become an important technique in fragment-
these proteins, they were able to identify novel potential
based drug discovery to optimize the binding potency of fragment hits).
treatment options.
In late 2021, Celeris published work12 describing the use of Variational
Autoencoders (VAEs) to augment existing data with a bond-angle-torsion
coordinate system, trained on the ZINC dataset, that demonstrated an
improvement of 9.3 percent (79 to 88.3) over the previous model (DeLinker).

EBOOK: THE LITTLE BOOK OF BIG CHANGES IN AI-POWERED DRUG DISCOVERY 7

 Cyclica is a drug discovery company headquartered in Toronto, Canada, with teams also located in the US and UK. They work to harness
AI and machine learning and utilize a custom-built interactome library to model potential protein interactions.

WHAT THEY’RE UP TO

Cyclica takes an interdisciplinary, collaborative approach to identify molecules

that address protein malfunction. The company leverages polypharmacology, a
method of concurrently evaluating interactions, to discover new drugs.

HOW THEY’RE USING AI

Cyclica’s machine learning platform, MatchMaker, combines features derived

from protein targets and small molecules to distinguish binding from non-
binding protein-ligand pairs. MatchMaker, trained on ~1.5M human bioactivities
(including DrugBank), innovates on existing drug-target interaction models by
augmenting the protein representations with structural data.

This approach14 involves mapping drug-target interaction pairs onto protein

binding sites of 3D protein structures. Uncertainties in these predicted and
experimental mappings are handled in the model using a deep neural network
method called Filtered Transfer Learning (FTL). FTL defines multiple tiers of data
confidence as separate tasks in a transfer learning setting. Fine-tuning of the
Figure: Data Partitioning by Confidence for Filtered Transfer Learning
DNN is achieved in a hierarchical process by iteratively removing data points with
lower label confidence and retraining.
WHY WE’RE SO IMPRESSED
In 2021, Cyclica applied many of these techniques to quickly identify
repurposed drug candidates for COVID-19. Cyclica developed a database
Cyclica’s platform MatchMaker, with its unique approach
called PolypharmDB15, a deep learning-based resource based on the DrugBank and generalization capacities, has demonstrated an ability
knowledge base. This resource was then utilized by researchers at Ryerson to improve its speed, efficiency, and success rates when
University to predict novel drug candidates16. predicting potential molecules of interest.

EBOOK: THE LITTLE BOOK OF BIG CHANGES IN AI-POWERED DRUG DISCOVERY 8

ACADEMIC

DeepCE In a collaborative project between Ohio State University, City University of New York, and Cornell University, researchers Thai-Hoang
Pham, Yue Qiu, Jucheng Zeng, Lei Xie, and Ping Zhang developed DeepCE, a mechanism-driven neural network-based method.

WHAT IT DOES

DeepCE expands on phenotype-based compound screening by modelling To validate the model’s effectiveness, the authors utilized DrugBank as a source for
chemical substructure-gene and gene-gene associations, predicting the clinically relevant drug-target and disease relationships. The results indicated that
differential gene expression profile perturbed by de novo chemicals. Essentially, integrating gene expression profiles generated with DeepCE can solve problems
DeepCE uses deep learning to predict how drugs will influence the amounts of related to unreliable data in the standard (L1000) dataset, leading to better
RNA, and therefore the amounts of various proteins, produced by a cell, which in performances on downstream prediction tasks. This specific application of DeepCE
turn provides insights into how the drug may modulate the disease. represents the first work of phenotype-based drug repurposing for COVID-19.

HOW THEY’RE USING AI Going one step further, Deep CE was applied to the full DrugBank dataset,
combined with gene expression data from patients with SARS-CoV-2, to predict
DeepCE uses a neural network-based model for gene expression profile
highly relevant drug candidates.
prediction consisting of several components. A graph convolutional network
is used to learn a vector representation for each chemical compound from WHY WE’RE SO IMPRESSED
its graph structure. A feed-forward neural network is used to learn vector
representations for cell line and chemical dose size. DeepCE offers improved performance compared to
existing methods and has the advantage of providing data
These vector representations are then put into the interaction component (two
augmentation, which makes it possible to tackle areas with
multihead attention modules, concatenated into a normalization layer followed
by feed-forward layer and another normalization layer) to learn high-level feature
minimal or unreliable data.
associations, including chemical substructure-gene and gene-gene feature
associations. Finally, the prediction component (two-layer feed-forward neural
network with a rectified linear unit activation function) takes the interaction
component’s outputs as inputs to simultaneously predict the gene expression
values for all L100017 genes.

EBOOK: THE LITTLE BOOK OF BIG CHANGES IN AI-POWERED DRUG DISCOVERY 9

DOCKSTRING
A research team from the Engineering, Chemistry, and Statistical Laboratory departments at the University of Cambridge, UK
created DOCKSTRING, a software and data bundle for meaningful and robust comparison of machine learning models.

WHAT IT DOES

One challenge in drug discovery is being able to utilize the full spectrum of The code is an open-source Python package, and the dataset is the first to
knowledge available. Often, approaches that would be beneficial require the include docking poses. It is also the most extensive dataset that offers a
researchers to have a deep level of understanding of the underlying biology. One complete matrix of docking scores for all ligand-target-pairs. This feature enables
example of this is molecular docking. It requires extensive domain knowledge experiments in transfer learning and multi-objective optimization.
to set up experiments and train machine learning correctly. DOCKSTRING was
WHY WE’RE SO IMPRESSED
created to help address this challenge.
DOCKSTRING, like other recent tools, is aiming to lower the barrier to entry for
HOW THEY’RE USING AI
drug discovery startups.
As machine learning methods for drug discovery continue to be developed,
benchmarks are required to compare performance against experimental data, It goes beyond structure-based modelling and brings more
giving an indication of what performance can be expected in the real world. complex techniques for predicting binding affinity into
more ligand design pipelines.
While other benchmarking methods exist, DOCKSTRING offers standardized and
accessible benchmarking capabilities based on molecular docking. The three-
component DOCKSTRING bundle includes code, datasets, and benchmarking
tasks which allow ML practitioners without biological expertise to obtain
meaningful docking scores.

EBOOK: THE LITTLE BOOK OF BIG CHANGES IN AI-POWERED DRUG DISCOVERY 10

EGNNs The UvA-Bosch Delta Lab at the University of Amsterdam focuses on the fundamentals of deep learning. At the 38th International
Conference on Machine Learning in 2021 their team of researchers introduced Equivariant Graph Neural Networks (EGNNs)18.

WHAT IT DOES WHY WE’RE SO IMPRESSED

Satorras, Hoogeboom, and Welling introduced the EGNN architecture for graphs Graph Neural Networks (GNNs) can accelerate the drug
that is translation, rotation, reflection, and permutation equivariant.
discovery process by providing an ability to analyze
Trained and tested against the QM919 20 dataset (a standard in ML for chemical molecules and their properties at a previously unattainable
property prediction tasks), Equivariant Graph Neural Networks (EGNNs) produces level, and EGNNs in particular represent a step forward in
highly competitive results in all property prediction tasks while remaining
terms of simplicity and efficiency.
simple, not requiring the use of higher-order molecular representations,
molecular angles, or spherical harmonics.

HOW THEY’RE USING AI

Computational de novo design of new drugs and optimization of known

compounds requires rigorous and unbiased exploration of chemical compound
space. Recent advances in Graph Neural Networks (GNNs) have made significant
improvements in this space in terms of accuracy and computational efficiency.

The EGNN-based model can predict all features from the QM9 dataset including
equilibrium geometries, frontier orbital eigenvalues, dipole moments, harmonic
frequencies, polarizabilities, and thermochemical energetics corresponding to
atomization energies, enthalpies, and entropies at ambient temperature.

EBOOK: THE LITTLE BOOK OF BIG CHANGES IN AI-POWERED DRUG DISCOVERY 11

Quality Data, Defined

If there’s one thing we know for sure, AI is at its Does it have quality coverage?
best, and your research will be too, when you Coverage means knowing that our data sufficiently captures all relevant
medical information.
have the highest quality drug data that spans a
vast range and depth of detail. Is it consistent?
All data must be input in a consistent manner. At DrugBank we have
In order to ensure that our data is of the highest quality we uphold strict
strict curation specifications that all of our data must meet before it is
criteria. Before we are satisfied with the quality of our datasets we ask ourselves
incorporated in our datasets. By standardizing this multi-step peer review
a series of questions.
process we ensure consistency and accuracy.

Does it tie back to common data entities?

As we add more data to our datasets we create and strengthen many new
connections between data points. Each additional connection improves the
interoperability of our datasets.

Is it hierarchical and flexible?

Quality data enables you to zoom out or drill down to the appropriate
level to adapt the data to the problem being solved. Our data needs to
encompass the full variety and complexity of information available, so we
incorporate as many levels of detail as possible.

EBOOK: THE LITTLE BOOK OF BIG CHANGES IN AI-POWERED DRUG DISCOVERY 12

 How structured is it? Whether you’re midway through years of research or just
Structured data is easier to search, find, use, and reason with. We work to starting out, it is always worthwhile to assess the quality
create highly structured, detailed data so that our users have total control in and source of your data to ensure that you’re operating as
how they manipulate and explore it. efficiently as possible, and getting results you can trust.

Is it evidence-based and can you follow

the data-lineage?
Quality data is evidence-based. By ensuring that all our data is based on
evidence and its lineage can be traced we are able to review and adjust our
data as the underlying evidence changes overtime.

Does it have the appropriate meta-data?

We make sure that we can trace all data curation actions so that we can
internally audit and question our data. At any time, we can review who
added or updated data and when they did it.

EBOOK: THE LITTLE BOOK OF BIG CHANGES IN AI-POWERED DRUG DISCOVERY 13

Hi, We’re DrugBank!
DrugBank was founded at the University of Alberta and is the world’s first
Here’s a handful of our most requested datasets:
intelligent and comprehensive drug knowledge platform.

With the help of artificial intelligence, our team of medical and scientific
Adverse Effects Metabolism
experts gather, author, verify, and organize all of the latest, most relevant
biomedical information into one machine-learning ready platform. This platform
Indications Targets
is accessible through data downloads or software integrations and is constantly
updated to include the latest findings.
Protein Relationships Chemical Structure
We’re working to augment human intelligence so that
the world’s medical information can be used to its fullest Drug Categories Pharmacology
potential and ensure that everyone has access to the best
possible medical outcomes.

Our datasets are ideal for all kinds of machine learning, drug discovery Contact us to learn more about DrugBank, our
applications. As a result we get to work alongside many leading researchers comprehensive drug database, & potential applications.
and institutions.

[email protected]

drugbank.com/datasets

EBOOK: THE LITTLE BOOK OF BIG CHANGES IN AI-POWERED DRUG DISCOVERY 14

Appendix

Resources
PAPERS GENERATIVE ALGORITHMS ADME AND TOXICITY PREDICTION
Bengio et al, 2021. Flow Network based Generative Models for Non- Siramshetty et al, 2021. Validating ADME QSAR Models Using
SURVEY PAPERS Iterative Diverse Candidate Generation. [Code] Marketed Drugs.

Berenger and Tsuda, 2021. Molecular generation by Fast Assembly of Göller et al, 2020. Bayer’s in silico ADMET platform: a journey of
Walters and Barzilay, 2021. Critical assessment of AI in drug discovery. (Deep)SMILES fragments. [Code] machine learning over the past two decades.
Coley, 2020. Defining and Exploring Chemical Spaces. Gao et al, 2021. Amortized Tree Generation for Bottom-up Synthesis Ryu et al, 2020. DeepHIT: a deep learning framework for prediction of
Chuang et al, 2020. Learning Molecular Representations for Planning and Synthesizable Molecular Design. [Code] hERG-induced cardiotoxicity. [Code]
Medicinal Chemistry. Takeuchi et al, 2021. R-group replacement database for
Walters and Barzilay, 2020. Applications of Deep Learning in Molecule medicinal chemistry. SYNTHETIC ACCESSABILITY AND
Generation and Molecular Property Prediction. Imrie et al, 2020. Deep Generative Models for 3D Linker Design. [Code] RETROSYNTHETIC PLANNING
Cai et al, 2020. Transfer Learning for Drug Discovery. Jin et al, 2020. Hierarchical Generation of Molecular Graphs using Fortunato et al, 2020. Data augmentation and pretraining for
Structural Motifs. [Code] template-based retrosynthetic prediction in computer-aided
synthesis planning.
REPRESENTATION AND TRANSFER LEARNING Polishchuk, 2020. CReM: chemically reasonable mutations framework
for structure generation. [Code] Koch et al, 2020.Reinforcement Learning for Bioretrosynthesis.
Ahmad et al, 2021. ChemBERTa-2: Towards Chemical Foundation
Somnath et al, 2020. Learning Graph Models for
Models. [Code]
Retrosynthesis Prediction.
Satorras et al, 2021. E(n) Equivariant Graph Neural Networks. [Code] HIT FINDING AND POTENCY PREDICITON
Townshend et al, 2021. ATOM3D: Tasks On Molecules in Bender et al, 2021. A practical guide to large-scale docking. VISUALIZATION AND INTERPRETABILITY
Three Dimensions. García-Ortegón et al, 2021. DOCKSTRING: easy molecular docking Humer et al, 2021. ChemInformatics Model Explorer (CIME):
Chuang and Keiser, 2020. Attention-Based Learning on yields better benchmarks for ligand design. [Code] [Data] Exploratory analysis of chemical model explanations. [Code]
Molecular Ensembles. Graff et al, 2021. Accelerating high-throughput virtual screening Matveieva and Polishchuk, 2021. Benchmarks for interpretation of
Li and Fourches, 2020. Inductive transfer learning for molecular through molecular pool-based active learning. [Code] QSAR models. [Code]
activity prediction: Next-Gen QSAR Models with MolPMoFiT. [Code] Gentile et al, 2020. Deep Docking: A Deep Learning Platform for
Augmentation of Structure Based Drug Discovery. [Code]

Cáceres et al, 2020. Adding Stochastic Negative Examples into Machine

Learning Improves Molecular Bioactivity Prediction.

Lin et al, 2019. Ultra-large library docking for discovering

new chemotypes.

EBOOK: THE LITTLE BOOK OF BIG CHANGES IN AI-POWERED DRUG DISCOVERY I

DATA SETS References
• ADME@NCATS 1. Smith, S. 230 Startups Using Artificial Intelligence in Drug Discovery. 14. Cyclica, Comparison of MatchMaker to DeepConv-DTI reveals
2021 Jun 10. superior performance and compute efficiency for predicting drug-
• AMED Cardiotoxicity Database
target interactions. 2021.
2. Smith, S. 43 Pharma Companies Using Artificial Intelligence in Drug
• BindingDB
Discovery. 2021 May 25; Accessed January 27, 2022. 15. Redka DS, MacKinnon SS, Landon M, et al. PolypharmDB, a
• ChEMBL Deep Learning-Based Resource, Quickly Identifies Repurposed
3. Smith, S. 116 Drugs in the Artificial Intelligence in Drug Discovery
Drug Candidates for COVID-19. ChemRxiv; (2020). DOI: 10.26434/
• DrugBank Online Pipeline. 2021 May 14; Accessed January 27, 2022.
chemrxiv.12071271.v1.
• DrugMatrix 4. Zhang D, Mishra S, Brynjolfsson E, Etchemendy J, Ganguli D,
16. Sugiyama MG, Cui H, Redka DS, Karimzadeh M, Rujas E, Maan
Grosz B, Lyons T, Manyika J, Niebles JC, Sellitto M, Shoham Y,
• Enamine Real database H, Hayat S, Cheung K, Misra R, McPhee JB, Viirre RD, Haller
Clark J, Perrault R. The AI Index 2021 Annual Report. AI Index
A, Botelho RJ, Karshafian R, Sabatinos SA, Fairn GD, Madani
• hERG Central Steering Committee: Human-Centered AI Institute, Stanford
Tonekaboni SA, Windemuth A, Julien JP, Shahani V, MacKinnon
University. 2021 March.
• MoleculeNet SS, Wang B, Antonescu CN. Multiscale interactome analysis
5. Davis B. How many scientific papers are published each year? 2021 coupled with off-target drug predictions reveals drug repurposing
• MONA: DB of Mass spec + other readouts
Jun 1; Accessed January 27, 2022. candidates for human coronavirus disease. Sci Rep. 2021 Dec
• NPASS database of natural products 2;11(1):23315. doi: 10.1038/s41598-021-02432-7. PMID: 34857794;
6. Shojania KG, Sampson M, Ansari MT, Ji J, Doucette S, Moher
PMCID: PMC8640055.
• PubChem D. How quickly do systematic reviews go out of date? A survival
analysis. Ann Intern Med. 2007 Aug 21;147(4):224-33. doi: 17. Subramanian A, Narayan R, Corsello SM, Peck DD, Natoli TE, Lu
• The Open Reaction Database
10.7326/0003-4819-147-4-200708210-00179. Epub 2007 Jul 16. X, Gould J, Davis JF, Tubelli AA, Asiedu JK, Lahr DL, Hirschman JE,
• Therapeutic Data Commons PMID: 17638714. Liu Z, Donahue M, Julian B, Khan M, Wadden D, Smith IC, Lam
D, Liberzon A, Toder C, Bagul M, Orzechowski M, Enache OM,
• Zinc 7. Tsao N. AI in Drug Discovery Cuts Timeline From 5+ Years to
Piccioni F, Johnson SA, Lyons NJ, Berger AH, Shamji AF, Brooks
Months. 2021 Jun 10; Accessed January 27, 2022.
AN, Vrcic A, Flynn C, Rosains J, Takeda DY, Hu R, Davison D, Lamb
8. Subramaniam S, Kleywegt GJ. A paradigm shift in structural J, Ardlie K, Hogstrom L, Greenside P, Gray NS, Clemons PA, Silver
For more up-to-date resources, please biology. Nat Methods 19, 20–23. 2022 Jan 11. S, Wu X, Zhao WN, Read-Button W, Wu X, Haggarty SJ, Ronco LV,
Boehm JS, Schreiber SL, Doench JG, Bittker JA, Root DE, Wong
visit the Awesome Small Molecule 9. de Oliveira TM, van Beek L, Shilliday F, Debreczeni JÉ, Phillips
B, Golub TR. A Next Generation Connectivity Map: L1000 Platform
C. Cryo-EM: The Resolution Revolution and Drug Discovery. SLAS
and the First 1,000,000 Profiles. Cell. 2017 Nov 30;171(6):1437-
Machine Learning GitHub repository. Discov. 2021 Jan;26(1):17-31. doi: 10.1177/2472555220960401.
1452.e17. doi: 10.1016/j.cell.2017.10.049. PMID: 29195078; PMCID:
Epub 2020 Oct 5. PMID: 33016175.
PMC5990023.
10. Zhavoronkov A, Ivanenkov YA, Aliper A, et al. Deep learning
18. Satorras VG., Hoogeboom E, Welling M. E(n) Equivariant Graph
enables rapid identification of potent DDR1 kinase inhibitors. Nat
Neural Networks. 2021.
Biotechnol 37, 1038–1040. 2019 Sep 2.
19. L. Ruddigkeit, R. van Deursen, L. C. Blum, J.-L. Reymond,
11. Mamoshina P, Volosnikova M, Ozerov IV, Putin E, Skibina E,
Enumeration of 166 billion organic small molecules in the
Cortese F, Zhavoronkov A. Machine Learning on Human Muscle
chemical universe database GDB-17, J. Chem. Inf. Model. 52,
Transcriptomic Data for Biomarker Discovery and Tissue-Specific
2864–2875, 2012.
Drug Target Identification. Frontiers in Genetics. 2018 Jul 12..
20. R. Ramakrishnan, P. O. Dral, M. Rupp, O. A. von Lilienfeld,
12. Flack M, Webber N, Trummer C. Decoupled coordinates for
Quantum chemistry structures and properties of 134 kilo molecules,
machine learning-based molecular fragment linking. 2021 Nov 1.
Scientific Data 1, 140022, 2014. [bibtex]
13. Celeris, Idendification of Novel West Nile Virus NS2B/NS3 Protease
Inhibitors. 2021.

EBOOK: THE LITTLE BOOK OF BIG CHANGES IN AI-POWERED DRUG DISCOVERY II

[email protected] | drugbank.com