The Rise of PharmaTech: Can Studying Self Driving Cars Steer Us Toward the Future of Drug Discovery?

Over the past several years, we’ve seen several industries undergo major disruption, perhaps none more so than in transportation. Companies at the forefront of this transition have positioned themselves as technology-first organizations with clear strategies to capture value from AI. What can life science companies learn from disruption in others industries ahead in the shift to Industry 4.0?

Autonomous Vehicle Development as a Case Study 

AV companies had turned to Software 2.0 - deep learning algorithms learning from data, or training, to perform a particular skill – as the tool to help realize their goal of bringing AVs to market. Companies scrambled to collect their real-world-driving data sets. Deploying fleets of camera and lidar-enabled vehicles to crawl city streets and build what would become the source code for autonomous vehicle development.  

After several years of development, they then began to understand the scope of their challenge; to collect enough real-world data to train an AI that could drive at a lower failure rate than humans do, it would take a fleet of 100 cars, driving 24/7/365, for 225 years. An intractable task, and certainly unacceptable for eager executives and investors.  

Enter simulation, or the use of software and computers to imitate the real world, physics and all. Because simulation is only limited by the amount of compute thrown at the problem, in-silico driving situations could be scaled up immensely, effectively helping to bring critical mass to the data challenge. Today, some of the world’s most powerful supercomputers are operated by AV companies for simulations + AI training to tackle the enormous problem of self-driving.  

Automotive companies are also leveraging machine learning and analytics to impact other parts of the enterprise. From supply chain, to manufacturing, to sales and marketing, to customer experience - business processes are being rethought through the lense of using data and algorithms to drive efficiencies.

So, what do self-driving cars have to do with drug discovery and life-sciences? Maybe more than you think.  

Simulating and Evaluating Candidates (You Thought Solving AV Was Hard?) 

We’ve seen the stats; drug discovery programs fail 90% of the time and the average cost of getting a new drug to market is an astounding $1.3 billion. The end-to-end drug discovery process is incredibly complex and resource intensive, requiring specialized biology, chemistry, and analysis of clinical patient data to succeed.  Pharma has always been an information-rich industry.  Similar to the disruption seen in automotive, a new breed of PharmaTech companies are rethinking how drugs are discovered using data, computing, and AI (and talented bilingual medicine/computer scientists) to reduce the number of failed drug programs and improve the likelihood of success from the start of a program.  

There are more than 10^60 potential drug-like molecules in the known chemical space. Like we've seen in AV development, scientists and researchers in drug discovery also have turned to leveraging simulation to approach this near-infinite problem and have been for 40 years. Modern techniques and computational power are now enabling teams to screen billions of molecules per day and predict chemical properties at experimental accuracy. These accurate in-silico predictions are done at a fraction of the cost of synthesizing and testing molecules in the lab. CADD experts had only dreamed of this a few decades ago.  

The combination of these modern physics-based approaches with AI techniques is becoming mainstream; applications of generative models and reinforcement learning techniques are “imagining” de-novo molecules with desired properties. Graph neural network approaches can predict quantum-level energies at thousand-fold less computational cost. Networks are "learning" to predict absolute free energy calculations from MD simulations, the list goes on. 

The Age of Biology, Enabled by AI 

Many mysteries remain in our understanding of the human body and how diseases work. To change that, researchers and pharma organizations are amassing petabytes of biomedical data sets and using machine learning to study disease at scale.  

Drug programs that start with a genetically supported target are twice as likely to succeed in the clinic. As sequencing programs across the globe ramp from hundreds of thousands to millions of genomes, analysis is quickly becoming the bottleneck of genomic research. Accelerated computing and AI are enhancing variant calling, allowing researchers to get more from their data through obtaining a better understanding of the disease pathway and ultimately, a druggable protein target along the pathway. 

Speaking of proteins… The scientific community was at minimum giddy, if not picking their collective jaws off the floor on the news of the AlphaFold breakthrough and the possibilities of predicting protein structures directly from sequences of amino acids. The implications of this AI-enabled science holds immense promise for structure-based drug design and therapy development.  

Phenotypic screening has been reimagined over the past several years with the advent of deep learning. Taking a target agnostic approach, image analysis is done on cell-based assays to quantify and classify the response. The combination of robotic-powered labs conducting biological data collection at scale, bilingual data scientists working on algorithm development, and powerful supercomputers training and providing insights on which therapies may lead to the best outcomes is a glimpse into the AI-enabled pharmaceutical company of the future – except they exist today. 

Synthesizing Knowledge from Language

There are troves of unstructured data like scientific journal articles, physician notes, and real-world evidence that have the potential to unlock insights in healthcare. There are hundreds of thousands of medical research articles published each year – an unimaginable amount of data to extract knowledge from. Organizations need the superpowers of modern NLP (Natural Language Processing) to keep up with the text data deluge. Use-cases ranging from protein target identification, drug repurposing, and gene interpretation to text summarization, clinical data de-identification, and patient selection are now possible because of domain-specific language models like BioMegatron. Making unstructured biomedical datasets computable can deliver tremendous value for healthcare and research organizations.  

AI and simulation are proving to be the essential tools for modern research and innovation, both of which are being made possible by accelerated computing. Studying disrupted industries such as transportation tells an interesting story of what opportunities lie ahead. Pharma companies that successfully transform into PharmaTech companies through the application of data strategy, simulation, talent, and AI will deliver much needed medicines faster and create enormous value.

Innovators Share Success Stories and Breakthrough Methods at GTC21 

Drug Discovery technologists and innovators will be coming together at NVIDIA’s GTC21 to share the latest and greatest AI developments that are altering every phase of drug discovery. The free, weeklong event will feature over 150 healthcare sessions. 

The conference will kick off on Monday, 4/12 at 8:30am PDT with a must-see keynote from NVIDIA CEO Jensen Huang announcing the latest technologies and partnerships advancing AI, HPC, and graphics. On Tuesday, 4/13 at 8am PDT, NVIDIA VP of Healthcare, Kimberly Powell, will deliver a special address on the latest advances in AI and how they are being applied. 

Here are some drug discovery sessions that I would highly recommend adding to your session scheduler after registering: 

• Machine Learning: A New Approach to Drug Discovery [S33061] Daphne Koller, Founder and CEO at Insitro 
• Dive Deep into Automated Drug Discovery: How Scalable Generative Platform Discovers New Lead-Like Molecules in Days [S32004] Alex Zhavoronkov, CEO at Insilico Medicine  
• Zero to COVID-19 Treatments in Under Four Weeks with Deep-Learning-Driven Drug Screens [S31542] Imran Haque, VP, Data Science at Recursion Pharmaceuticals  
• CUina: An Efficient Implementation of Autodock Vina Specially Crafted for CUDA GPUs [S32046] Adrian Morrison, Senior Software Engineer at Atomwise 
• Accelerating Drug Discovery with Advanced Computational Modeling [S33182] Robert Abel, Chief Computational Scientist at Schr?dinger 
• OrbNet: Quantum-Enabled Deep Learning to Accelerate Molecular Simulation and Discovery [S32585] Thomas Miller, CEO at Entos 
• Augmented Curation of Clinical Notes from a Massive EHR System Reveals Symptoms of Impending COVID-19 Diagnosis [S32580] AJ Venkatakrishnan, Senior Director, Scientific Research and Partnerships at Nference 
• Utilizing Data Fabrics to Enable GPU-Accelerated Unstructured Text Analytics [S32123] Chris Bouton, CEO at Vyasa Analytics 

Register for GTC today to get access to all of these sessions and workshops, absolutely free.