Oncora Medical
August 1, 2022Transfer Learning
August 1, 2022Atomwise
AI to create convolutional neural networks to predict the bioactivity of molecules for drug discovery...
Overview
Atomwise is a drug discovery company headquartered in San Francisco, California. They use AI to create convolutional neural networks to predict the bioactivity of small molecules for drug discovery applications. The company works on predicting the binding of small molecules to proteins by using a statistical model that extracts insights from experimental affinity measurements and protein structures. This enables chemists to pursue hit discovery, lead optimization, and toxicity predictions. Atomwise analyzes billions of compounds to identify a specific subset for synthesis and testing.
Atomwise’s A.I Aspect
Of the roughly 20 thousand proteins encoded in the human genome, a little over 750 have FDA-approved drugs, and only about two to three thousand have a small molecule drug or biologic under development. About 4 thousand genes have evidence linked to human disease but may not yet have enough structural data for a traditional drug design. The remaining 80% of human gene targets are uncharted territory, representing the most challenging and promising future for the pharmaceutical field and general human health. Artificial intelligence (AI) has proven to help unlock drug targets in the most challenging and uncharted zones.
With their AI-based platform, AtomNet works on some of the most challenging steps in drug discovery. Their technology enables them to substantially decrease the time and costs required in the early stages of drug development.
Their industry-leading AI technology can predict billions of small molecule-protein interactions to help discover novel compounds for hit and lead optimizations.
Atomwise can identify hit compounds for a target protein and help develop a focused library of quality leads. High throughput screening and the “selective” standard libraries are things of the past. Atomwise drastically reduces physical screening efforts and helps identify leads without synthesizing or buying large compound libraries.
Lead optimization no longer needs to be the most demanding or costliest step in drug discovery. By leveraging their AI platform, Atomwise helps jump each hurdle in lead optimization, from potency to selectivity to ADME-Tox. Additionally, they set up for success in both animal and human trials by screening for toxicity in each model.
Atomwise’s Technology
AtomNet removes some of the physical barriers that previously limited the success of drug discovery. Research and development are no longer constrained by the limited number of compounds available in a library and the time traditionally required to screen these compounds. Atomwise uses its artificial intelligence software to analyze a vast chemical space, enabling them to find small subsets that are genuine candidates for successful synthesis and testing. Discovery and optimization processes that would take years in the past are now compressed with Atomwise’s technology to a matter of weeks. AtomNet has shown the industry how to apply the modeling of bioactivity and chemical interactions parting from the convolutional concepts of feature locality and hierarchical composition.
AtomNet removes some of the physical barriers that previously limited the success of drug discovery. Research and development are no longer constrained by the limited number of compounds available in a library and the time traditionally required to screen these compounds. Atomwise uses its artificial intelligence software to analyze a vast chemical space, enabling them to find small subsets that are genuine candidates for successful synthesis and testing. Discovery and optimization processes that would take years in the past are now compressed with Atomwise’s technology to a matter of weeks. AtomNet has shown the industry how to apply the modeling of bioactivity and chemical interactions parting from the convolutional concepts of feature locality and hierarchical composition.
They specifically applied it where an image is represented as a 2-dimensional grid of pixels containing channels for red, green, and blue. AtomNet represents a protein-ligand pair as a set of 3-dimensional volumetric pixels containing channels for carbon, oxygen, and nitrogen, among others. In this way, AtomNet avoids the manual process of tweaking and over-parameterizing binding features by autonomously learning the features governing molecular binding that typified traditional computational methods.
Some machine learning architectures have analogous constraints as the biochemical interactions that tend to be primarily local, so they can be used to model these biological interactions. As with edge detectors in convolutional neural networks for images, these local biochemical detectors are then hierarchically arranged into more intricate features describing the complex and nonlinear phenomenon of molecular binding.
Atomwise’s Financial Aspects
Atomwise announced that it closed at $123 million in an oversubscribed Series B financing led by B Capital Group and Sanabil Investments. The funding round includes returning investors DVVC, BV, Tencent, Y Combinator, Dolby Family Ventures, and AME Cloud Ventures, as well as new backing from two top ten global insurance companies. To date, Atomwise has raised almost $175 million; they also appointed a new board member from B Capital Group, Raj Ganguly, and a board observer, Hani Enaya, from Sanabil. The latest funding round will allow Atomwise to scale its AI technology platform and team. The company plans to expand its work with corporate partners, including Eli Lilly and Company Bayer, Hansoh Pharmaceuticals, and Bridge Biotherapeutics, major players in the biopharma field.
Atomwise announced that it closed at $123 million in an oversubscribed Series B financing led by B Capital Group and Sanabil Investments. The funding round includes returning investors DVVC, BV, Tencent, Y Combinator, Dolby Family Ventures, and AME Cloud Ventures, as well as new backing from two top ten global insurance companies. To date, Atomwise has raised almost $175 million; they also appointed a new board member from B Capital Group, Raj Ganguly, and a board observer, Hani Enaya, from Sanabil. The latest funding round will allow Atomwise to scale its AI technology platform and team. The company plans to expand its work with corporate partners, including Eli Lilly and Company Bayer, Hansoh Pharmaceuticals, and Bridge Biotherapeutics, major players in the biopharma field.
They will also partner with emerging biotechnology companies such as StemoniX and SEngine Precision Medicine. Atomwise will continue to use AtomNet for drug discovery to grow its portfolio of joint ventures with leading researchers, like the ongoing projects with X-37 and Atropos Therapeutics.
Atomwise will continue to use AtomNet for drug discovery to grow its portfolio of joint ventures with leading researchers, like the ongoing projects with X-37 and Atropos Therapeutics.
AI for COVID-19 Therapy Discovery
Deep learning has proven to be essential in the discovery of SARS-COV-2 therapies, which enables providers to have a faster and more accurate response to the pandemic. Artificial intelligence (AI) has been involved in different studies of drug discovery and vaccine development that is impactful for SARS-COV-2 therapy discovery.
Machine learning works by creating models that can manage different data and automatic feature extraction from raw data. The automatic feature extraction ability of deep learning delivers more effective and accurate results. Moreover, deep learning models can generate molecules and epitope prediction, thus minimizing the chances of failure.
Vaccine research against COVID-19 studied the Spike protein, which had proven to be the main qualifier for virtual vaccine discovery because it is required for viral access. Antibodies against the receptor-binding domain of Spike can block the attachment and the fusion of viral proteins.
Current Methods and Challenges for Deep Learning in Drug Discovery
The rise of deep learning in the last decades has transformed many industries, including speech recognition, computer vision, language processing, and machine translation. Chemistry has a long history of data-driven methods. AI systems have been used to predict chemical reaction rates, compound properties, and bio-activities at least since the early 1990s. Deep learning methods have improved the golden standard on molecular prediction tasks. An essential factor is the range of suitable and flexible architectures to supersede handcrafted features.
The current challenges include data that is often sparse, noisy, biased, and inconsistent, which makes training and selecting the best methods a difficult process. However, there is potential for success, including developing strategies for collaboration across disciplines, generating more reliable data, and promoting a knowledge-sharing culture.
Miro1 Protein Removal - Promising Results for Treatment of Parkinson’s Disease
Miro is a protein localized on the outer mitochondrial membrane, and its role is to support mitochondria to microtubules. During the process of mitophagy, it is removed from the surface of mitochondria. Mitochondrial protein Miro1 has been proposed as a way of detecting Parkinson’s Disease (PD) and as a drug target for treating the disease. AI technology has been involved in searching for small molecules that are predicted to bind to Miro1. In future clinical trials, PD patients could be analyzed for the presence of Miro1 phenotype using their cultured fibroblasts in a dish. The fibroblasts could be used to experiment sensitivity to drugs and identify the proper treatment.
Miro is a protein localized on the outer mitochondrial membrane, and its role is to support mitochondria to microtubules. During the process of mitophagy, it is removed from the surface of mitochondria. Mitochondrial protein Miro1 has been proposed as a way of detecting Parkinson’s Disease (PD) and as a drug target for treating the disease. AI technology has been involved in searching for small molecules that are predicted to bind to Miro1. In future clinical trials, PD patients could be analyzed for the presence of Miro1 phenotype using their cultured fibroblasts in a dish. The fibroblasts could be used to experiment sensitivity to drugs and identify the proper treatment.
Discovery of a Novel Inhibitor of a Critical Brain Enzyme Using Artificial Intelligence
More than 8,000 diseases are classified as rare and together affect nearly 350 million people worldwide. This healthcare burden creates the need to develop methods to improve drug discovery and development. Canavan disease is a rare disorder affecting several hundred people in the United States. This disease is caused by mutations in the aspA gene that codes for the enzyme aspartoacylase. The faulty functioning of this enzyme has been proposed to lead to decreased fatty acid biosynthesis and neuronal demyelination.
For this study, researchers used a deep neural network called AtomNet to identify a set of five novel scaffolds with low-micromolar potency against aspartate N-acetyltransferase, a promising target for the treatment of patients suffering from Canavan disease.
These compounds could be used as the starting point for future treatment programs. The system supported early drug discovery even when target data was scarce or unavailable. The results promise that novel ML algorithms could be applied for the discovery of treatments for even the rarest diseases.