Drug Design: AI tools and startups

In my previous articles I reviewed 🖊 some of the different AI tools and startups for 🔹 primary and secondary screening for the identification of drug candidates (bioactivity and toxicity prediction and physicochemical properties) and for 🔹planning and execution of chemical synthesis during drug discovery. Consequently, today I am going to talk about AI tools and startups for drug design during drug discovery, namely target-protein structure prediction and drug-protein interactions.

Lets' start from the beginning. Searching for a new drug candidate, means searching and wandering around in a vast chemical space, comprising >10^60 molecules (keep in mind that there are something like 10^22 to 10^24 stars in our Universe). In particular, the known chemical space — that includes public databases and corporate collections — probably contains something like 10^8 molecules (100 million), but the virtual chemical space might contain 10^60 compounds when considering only basic structural rules, or a more modest 10^20–10^24 molecules if combination of known fragments are considered.

Since the chemical space is far too large for an exhaustive "exploration", one is therefore left only with a partial, targeted "exploration" inside smaller virtual libraries and smaller chemical libraries. So, numerous in silico methods are used to virtual screen compounds from virtual chemical spaces along with in vitro high-throughput screening experiments of chemical libraries, in order to identify drug candidates. Then, once a hit is identified, a critical part in drug discovery is conducting "Drug Metabolism and Pharmacokinetics" studies, often referred to as ADMET (Absorption, Distribution, Metabolism, Elimination, Toxicity) studies, doing cell-based in vitro assays followed by animal studies and now also using AI tools (DeepTox algorithm, DeepNeuralNet- QSAR, Hit Dexter etc).

Subsequently, at the end of the primary and secondary screening, the drug discovery process proceeds with searching for an optimal chemical synthesis pathway to synthesise the drug candidate.

At this point, starts the drug design phase of the drug discovery process, namely the inventive process of finding new medications based on the knowledge of a biological target (design of molecules complementary in shape and charge to the biomolecular target with which they interact and therefore will bind to it).

That is, drug design (a more accurate term is ligand design) relies on the knowledge of the three-dimensional structure of the biomolecular target (protein, peptide, antibody, nucleic acid), and is known as structure-based drug design. So, 🔹prediction of the target protein (protein structure prediction) namely the deduction of the three-dimensional structure of a protein from its amino acid sequence — that is, the prediction of its secondary and tertiary structure from primary structure — and 🔹theprediction of the drug–protein interactions are all crucial during drug design.

Lets's see now the AI tools for drug design.

🎯AI in Drug Design

Homology modeling (identification of one or more known protein structures likely to resemble the structure of the query sequence) and de novo protein design have traditionally been applied to determine the 3D chemical structure of the target protein. In computational biology, de novo protein structure prediction refers to an algorithmic process by which protein tertiary structure is predicted from its amino acid primary sequence. De novo methods tend to require vast computational resources, and have thus only been carried out for relatively small proteins. Prediction of protein structure de novo for larger proteins will require better algorithms and larger computational resources such as those afforded by either powerful supercomputers (Blue Gene or MDGRAPE-3) or distributed computing projects (Folding@home, Rosetta@home, the Human Proteome Folding Project, or Nutritious Rice for the World).

👉🏻In particular, in a Critical Assessment of Protein Structure Prediction contest, the AI tool AlphaFold of Deep mind (De novo protein structure prediction by deep-learning based distance prediction) was used to predict the 3D structure of a drug target protein and performed amazingly well. In fact, AlphaFold accurately predicted 25 of 43 structures using only protein primary sequences (the code is available here), and the second-place contester correctly predicted only three out of the 43 test sequences.

And just a year ago DeepMind’s AI system AlphaFold has been recognised as a solution to the 50-year-old grand challenge of protein structure prediction, namely AlphaFold can accurately predict the structure that proteins will fold into in a matter of days, in what is being described as a Major Scientific Advance.

👉🏻Furthermore, QM or the hybrid QM/MM (quantum mechanics/molecular mechanics approach — a molecular simulation method that combines the strengths of the quantum mechanics (QM) and molecular mechanics (MM) approaches, allowing scientists the study of chemical processes in solution and in proteins — are both useful for predicting protein–ligand (drug) interactions during drug design.

👉🏻For example, for large datasets quantum chemistry-derived DFT potential energies have been calculated and used to train deep neural networks. DFT or density functional theory is a computational QM modelling method used in physics, chemistry and materials science to investigate the electronic structure or nuclear structure of atoms and molecules.

👉🏻Aurora also employs QM, thermodynamics, and an advanced continuous water model for solvation effects to calculate ligand's binding affinities with a protein. By including the entropy and aqueous electrostatic contributions in to the calculations directly, Aurora algorithms produce much more accurate and robust values of binding free energies.

👉🏻Moreover, in order to improve scoring-docking-screening powers of protein-ligand docking functions simultaneously (molecular docking is a method which predicts the preferred orientation of one molecule to a second when bound to each other to form a stable complex), a new scoring function was introduced, the Δvina RF, which employs 20 descriptors in addition to the AutoDock Vina score. AutoDock is a 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. Current distributions of AutoDock consist of two generations of software: AutoDock 4 and AutoDock Vina. This new (Delta Vina) scoring function incorporates explicit water molecules and ligand stability that has been developed using the Delta-Vina XGBoost (ΔvinaXGB) algorithm. It has achieved superior performance in all power tests of CASF-2016 benchmark and two rescoring tests (LocalOpt and FlexDock).

Finally, other AI-based computational tools for drug discovery are:

🔹DeepChem is a python-based AI tool for various drug discovery task predictions such as: predicting the solubility of small drug molecules, predicting binding affinity for small molecules to protein targets, predicting physical properties of simple materials, analysing protein structures and extracting useful descriptors and count the number of cells in a microscopy image. DeepChem contains a variety of very helpful modules and tools, but has limitations in its ability to robustly train models from a wide selection of hyperparameters, and published performance evaluation is limited to a small number of public data sets with less diverse pharmaceutical relevance.🔹REINVENT is a molecular de novo design tool using recurrent neural networks and reinforcement learning, applied on drug discovery projects while navigating the chemical space.🔹Open Drug Discovery Toolkit (ODDT) is modular and comprehensive toolkit for use in cheminformatics and molecular modeling. ODDT is written in Python, and make extensive use of Numpy/Scipy.🔹ORGANIC an efficient molecular generation tool to create molecules with desired properties.🔹PotentialNet a ligand-binding affinity prediction tool based on a graph convolutional neural network.🔹QML a Python toolkit for quantum ML.🔹SIEVE-Score (Similarity of Interaction Energy VEctor Score) is another AI tool of structure-based virtual screening via interaction energy based learning.🔹NNScore 2.0 is also a neural network-based scoring function for protein–ligand interactions.

🎯AI startups involved in drug design

X-37🤖 (San Francisco Bay Area, West Coast, Western US 2018) uses AI to perform structure-based drug design and discovery (using the AtomNet platform from Atomise) to discover novel drug candidates by screening vast libraries of chemical compounds against high-value pharmaceutical targets. X-37 has raised a total of $14.5M💵 in funding over 2 rounds. Their latest funding was raised on Nov 14, 2019 from a Series A round.

Cloud Pharmaceuticals🤖 (North Carolina, US 2014) uses AI to search a virtual chemical space and predict binding affinity. Their research platform is the Quantum Molecular Design (QMD) and the key component of the QMD is the AI engine with the application of quantum mechanics and molecular mechanics to accurately predict ligand binding affinities in solvents and in protein environments. Cloud Pharmaceuticals has raised a total of $1.8M💵 in funding over 4 rounds. Their latest funding was raised on Feb 9, 2016 from a Venture - Series Unknown round. And has a post-money valuation in the range of $1M to $10M💵 as of Jul 25, 2014.

Acellera🤖 (London, England, UK 2006) uses AI to predict protein-ligand binding, exclude toxic or reactive molecules and improve ADMET profiles. Acellera has raised a total of €50K💵 in funding over 1 round. This was a Grant round raised on Jun 1, 2015. Acellera is funded by EASME - EU Executive Agency for SMEs.

Aqemia🤖 (Paris, Ile-de-France, France 2019) is an in silico drug discovery start-up that combines AI and structure-based affinity algorithms to discover rapidly more innovative therapeutic molecules. Thanks to unique Quantum-inspired Statistical Mechanics algorithms they predict drug-target affinity among other properties, in a 10 000x faster way than the competition. Aqemia has raised a total of €1.6M💵 in funding over 2 rounds. Their latest funding was raised on Oct 17, 2019 from a Pre-Seed round.

FAR Biotech 🤖 (Huston Texas, US 2016) applies quantum mechanics and machine learning to identify drug-like molecules throughout the chemical space of about 1,5 trillion chemical structures (including new chemical entities, known compounds, and repurposing drugs).

Kuano 🤖 (London, UK 2020) is developing novel AI and quantum solutions for designing molecules. Their Platform is informed by data on structure of the target enzyme/catalytic site, combining quantum simulation and quantum-inspired AI/chemistry, and enables them to tackle intractable targets and development challenges such as specificity and drug resistance.

Menten AI 🤖 (Palo Alto, California, US 2018) develops a software platform for protein design powered by ML and quantum computing. Menten AI was the first company that created a peptide using a quantum computer.  Menten AI has raised a total of $4M💵 in funding over 2 rounds. Their latest funding was raised on Jun 30, 2020 from a Seed round.

Pharmacelera 🤖 (Barcelona, Catalonia, Spain 2015) is applying quantum theory to boost drug design via their two primary software packages: PharmScreen (for an accurate ligand-based virtual screening) and PharmQSAR (3D quantitative structure-activity relationship/QSAR tool that enables a combination of multiple fields of interaction in order to perform comparative molecular field analysis and comparative similarity indices analysis studies). Pharmacelera has raised a total of €2M💵 in funding over 3 rounds. Their latest funding was raised on Mar 15, 2020 from a Grant round.

PharmCADD 🤖 (Busan, Pusan-jikhalsi, South Korea 2019) its main technology platform is called “Pharmulator” that operates with deep neural network algorithms, molecular dynamic simulation and quantum mechanics computation, in order to reconstruct 3D protein structure from amino acid sequence within just several seconds. PharmCADD raised $16M💵 in Series B Funding.  

Polaris Quantum Biotech 🤖 (Durham, North Carolina, US 2020) also known as Polarisqb is using (collaboration with Fujitsu) a combination of artificial intelligence with quantum computing, to speed-up the drug discovery process from 5 years to 4 months. Recently the company received $250K 💵 in equity.

ProteinQure 🤖 (Toronto, Ontario, Canada 2017) is combining quantum computing, reinforcement learning and atomistic simulations to design novel protein drugs. Using their proprietary algorithms and external supercomputing resources, ProteinQure can design small peptides (including cyclic) therapeutics, and explore protein structures without known crystal structures. ProteinQure has raised a total of $4.6M💵 in funding over 3 rounds. Their latest funding was raised on Jul 1, 2019 from a Seed round.

Riverlane 🤖 (Cambridge, Cambridgeshire, UK 2017) builds ground-breaking software to unleash the power of quantum computers with early adopters in materials design and drug discovery. Riverlane has raised a total of $24.1M💵 in funding over 3 rounds. Their latest funding was raised on Mar 23, 2021 from a Seed round.

Silicon Therapeutics 🤖 (Boston, Massachusetts, US 2016) brings a physics-driven approach to drug design that is tightly coupled to molecular simulations, quantum physics, statistical thermodynamics and molecular dynamics for the improvement of conventional drug discovery. The company was acquired by Roivant Sciences (a biopharmaceutical company focused on completing the development of promising late-stage drug candidates) on Feb 26, 2021 for $450M💵.

XtalPi 🤖 (Cambridge, Massachusetts, US 2014) is applying quantum mechanics to augment research modeling. Its Intelligent Digital Drug Discovery and Development (ID4) platform, incorporating quantum mechanics, AI and high-performance cloud computing algorithms, allows predicting with high precision physiochemical and pharmaceutical properties of small-molecule drug candidates, as well as their crystal structures. XtalPi has raised a total of $386.4M💵 in funding over 9 rounds. Their latest funding was raised on Sep 28, 2020 from a Series C round and has a post-money valuation in the range of $1B to $10B💵 as of Sep 28, 2020.

ApexQubit 🤖 (Berkeley, California, US 2018) is using combination of reinforcement learning, generative models and quantum computing to search for the most promising undiscovered small molecules and peptides. ApexQubit has raised a total of $430.5K💵 in funding over 6 rounds. Their latest funding was raised on Aug 27, 2020 from a Non-equity Assistance round. ApexQubit has acquired RNP Therapeutics (a biotechnology company targeting the components of tumor microenvironment directly through non-coding RNA) on Dec 1, 2020.

Thank you for reading 👓💙

And if you liked this post why not share it?

@MetaphysicalCells

#science #health #pharma #drugdiscovery #drugdevelopment #AI #biotechAI

💻💊🔬💉📊🤖

👉🏻References

AI Drug Development Startups

Advancing Drug Discovery via Artificial Intelligence

AMPL: A Data-Driven Modeling Pipeline for Drug Discovery

Artificial intelligence in drug discovery (part 3)