Generative AI in Pharmaceutical Drug Discovery and Design

Welcome to the 15th issue of the AstroFeather AI newsletter!

[ Daily AI News | Newsletter Signup | View Online ]

In the past week (or so), there have been some interesting stories about generative AI being used to speed up the pharmaceutical drug discovery and development process, as well as to improve military planning and warfighting capabilities.

On the other side of the industry, the Mayo Clinic (often ranked as the #1 hospital in the US across multiple specialties) is testing Google's AI medical chatbot in a clinical setting, there have been several headline-grabbing product announcements, including the release of Anthropic's Claude 2 (ChatGPT competitor), and the Associated Press and OpenAI have reached an agreement to share news content and technology.

You'll find these and many more trending discussions in this week's issue. I hope you enjoy reading this week’s updates and if you have any helpful feedback, feel free to respond to this email or contact me directly on LinkedIn [@adideswilliams]

Thanks - Adides Williams, Founder @ AstroFeather

In this week’s recap (10 - 15 min read):

• Generative AI in Pharmaceutical Drug Discovery and Design

• The US Military is Testing Generative AI for the Battlefield

• Product Previews and Launches

• Company Announcements and News Throughout the Industry

Must-Read News Articles and Updates

Update #1. Generative AI in Pharmaceutical Drug Discovery and Design.

In recent months, news headlines have been dominated by the rise of generative AI (genAI) services designed to perform a variety of tasks, including generating prose in different literary styles, hyper-realistic art, human-sounding music compositions, videos, and voice clones. Biotech companies, however, are using genAI to design drugs to treat debilitating diseases, adding to the growing list of use cases for the technology.

The latest: Insilico Medicine recently used an integrated genAI-based platform (Pharma.AI) to discover and design a drug candidate (INS018_055) for the treatment of a rare chronic lung disease (idiopathic pulmonary fibrosis, or IPF) that causes an irreversible decline in lung function, according to a report from Nvidia. INS018_055 has now completed Phase 0 and Phase I human safety studies and is believed to be the first drug discovered and designed using genAI to enter Phase II clinical trials.

How it works: Drug development is a difficult, time-consuming, and expensive process that (generally) consists of four steps:

1. Researchers must find a "target" (a malfunctioning biological mechanism) that causes disease.

2. They must develop a new drug to stop the disease-causing target without harming the patient.

3. The effectiveness and safety of the new drug must be determined through animal studies and clinical trials in healthy volunteers and patients.

4. Finally, research teams can apply for regulatory approval of the new drug if all tests show positive results in helping patients.

Given the complexity of all the steps involved in the drug discovery process, it can take an average of six years and at least half a billion dollars to develop a new drug. To address these challenges, Insilico Medicine developed Pharma.AI, an integrated platform consisting of three AI systems (PandaOmics, Chemistry42, and inClinico) that enabled its researchers to move from target discovery to human clinical trials (for INS018_055) in 30 months:

• PandaOmics: An AI target identification platform that analyzes scientific data from patents, research papers, and clinical trial databases to discover and evaluate new disease-causing targets.

• Chemistry42: Once a target has been discovered, researchers can input it into Chemistry42, a generative chemistry system that uses genAI to design a collection of drug candidates that have been optimized for specific properties.

• inClinico: Optionally, researchers can submit the newly designed drug to inClinico to help design clinical trials (including trial structure and patient eligibility criteria) and predict clinical trial success rates.

Clinical trial results for the AI-designed drug candidate:

• Phase 1 trials: The drug candidate (INS018_055) achieved positive topline results in Phase I trials and was tested in 78 healthy volunteers in New Zealand.

• Regulatory traction: Earlier this year, INS018_055 received the FDA's first Orphan Drug Designation for a drug developed using genAI. *Side note: Orphan Drug Designation refers to INS018_055's potential to treat a rare disease (in this case, idiopathic pulmonary fibrosis, or IPF).

• Phase II trial design: Phase II trials are randomized, double-blind, placebo-controlled studies designed to evaluate the safety and tolerability of the drug candidate.

Behind the news: Drug design with genAI appears to be gaining traction among research labs and biotech investors. Nvidia is partnering with both Insilico Medicine and Recursion Pharmaceuticals to enhance its genAI drug discovery and design platforms, Baidu Research is developing designer vaccines, and the University of Washington is developing designer proteins:

• Insilico Medicine’s genAI drug portfolio: In addition to tackling lung diseases, Insilico Medicine has also used genAI to design other drugs that have advanced to clinical trials, including one to treat COVID-19 (ISM3312) and another to treat breast cancer (USP1).

• Nvidia invests in AI drug discovery: Salt Lake City, Utah-based Recursion Pharmaceuticals announced that has raised $50 million from Nvidia to accelerate the development of its AI models for drug design and discovery. The biotech firm will reportedly use its biological and chemical data sets to train its AI models on Nvidia’s BioNeMo drug discovery cloud platform.

• Baidu Research’s designer vaccines: Baidu’s Silicon Valley-based research arm has created an AI tool called LinearDesign, which optimizes gene sequences in COVID-19 mRNA vaccines to create vaccines with greater stability (for storage until use) and efficacy (a measure of how well a vaccine works).

• University of Washington’s designer proteins: Researchers at the University of Washington have developed RFdiffusion, a neural network similar to image generators such as Midjourney, to design custom proteins for applications in therapeutics, vaccines, and biomaterials.

Why it matters:

• Perhaps one of the most exciting benefits of generative AI (genAI) is its potential to accelerate the drug discovery and development process. GenAI platforms can rapidly analyze vast databases (including chemical, biological, and clinical trial data) to identify novel targets of disease progression, design a collection of molecularly optimized drug candidates, and predict their clinical efficacy and safety profiles.

• It is well known that the process of developing new drugs is slow, complex, expensive, and laborious, requiring the evaluation of hundreds of thousands of drug candidates before a project reaches clinical trials. GenAI platforms have the potential to address these challenges and help research teams significantly shorten the drug development cycle, reduce costs, and safely advance potentially life-saving treatments into trials in less time than current methods.

Update #2. The US Military is Testing Generative AI for the Battlefield.

The latest: According to a recent Bloomberg report, the US Department of Defense (DoD) is exploring the use of generative AI (genAI) to support improved decision making and scenario planning on the battlefield.

How it works: While the exact details of the DoD's genAI tests have not been made public, the Bloomberg report offers some clues, suggesting that the DoD is testing five large language models (LLMs) trained on classified operational information as part of a series of experiments run by the DoD's Chief Digital and AI Office (CDAO). Of the five LLMs mentioned in the report, only Scale AI's Donovan is confirmed to be in use for the CDAO's testing:

• Scale AI's Donovan: In a recent interview with Bloomberg, Scale AI's CEO confirmed that the defense-focused Scale Donovan platform is the first LLM to be deployed on a classified network for the US government and is currently in use by the US Army XVIII Airborne Corps, DoD's CDAO, and the Marine Corps School of Advanced Warfighting (SAW).

• Four remaining LLM platforms: While these are unknown, potential vendors include Palantir Technologies and Microsoft/OpenAI. Palantir recently unveiled its artificial intelligence platform (AIP), which uses a variety of LLMs (Dolly-v2-12b, Flan-T5XL, and GPT-NeoX-20b) and is capable of coordinating military units in real time. In addition, the DoD is one of Microsoft's Azure Government users with access to OpenAI's AI models.

Experiment and test results:

• The military exercises are expected to last until July 26 (this year) and will focus on using the LLM platforms to help plan a military response to an escalating global crisis that will eventually shift to the Indo-Pacific region. Notably, both Scale Donovan and Palantir's AIP feature a conversational interface that allows operators to use text prompts to interact with the underlying LLMs and ask questions about enemy military activity, deploy surveillance drones to capture video of potential threats, and generate courses of action (COAs) that can be sent to commanders for approval.

• Early test results showed that one of the LLM platforms tested was able to complete a military plan in 10 minutes for something that would normally take hours or days. However, the DoD has expressed ongoing concerns about cybersecurity attacks on LLMs (specifically through "poisoned" training datasets), as well as hallucinations (a term commonly used to describe when LLMs confidently provide fabricated (and incorrect) answers to questions).

Behind the news: The US military's interest in leveraging AI on and off the battlefield has been on the rise since earlier this year, including multi-billion dollar budget allocations for AI-related initiatives, renewed efforts to revive its JADC2 cloud-like network, and the advancement of genAI to the "plan" phase of the DoD's technology watch list:

• Historic research and development (R&D) budget requests: The DoD recently proposed its largest-ever budget for innovation, asking Congress for $145 billion in fiscal year 2024. Part of this proposal included investments in science and technology ($17.8 billion), AI (including machine learning and genAI) ($1.8 billion), and Joint All Domain Command and Control (JADC2) implementation ($1.4 billion). Taken together, these programs (along with the DoD's recently established Chief Digital and AI Office (CDAO)) signal the US military's increased focus on improving its warfighting (and technological) capabilities by introducing AI-enabled systems on its secure platforms.

• US DoD revives JADC2: The DoD has revived efforts to implement the JADC2, which is described as a system that will connect all the services, including the Navy, Air Force, Marines, and Space Force, "into a single cloud-like network," allowing teams to seamlessly share data and "deploy the full force of military capabilities during current and future conflict." The JADC2 will essentially help the DoD quickly act on information across the battlespace using genAI, machine learning, and predictive analytics.

• DoD Technology Watchlist Update: The DoD's Defense Information Systems Agency (DISA) recently updated its technology watchlist to include generative AI (genAI). DISA's technology watch list is considered a window into the DoD's vision for a future digital battlefield. The list highlights emerging technologies that can enhance military capabilities and now includes genAI in its "PLAN" phase to determine technical implications for DISA and DoD missions.

Why it matters:

• Advances in lethal autonomous weapons used in the Ukraine-Russia war, ongoing tensions with China and Russia, and the recent Cambrian-like explosion of generative AI use cases are believed to be some of the driving forces behind the US military's renewed interest in enhancing its warfighting capabilities with AI and other advanced technologies.

• The DoD's integration of Scale AI's Donovan (a defense-focused LLM platform) into its secure networks, inclusion of genAI on its technology watch list, and multi-billion dollar budget requests specifically for AI-related initiatives suggest that the US military is focused on leveraging AI to advance its military planning and warfighting capabilities.

Update #3. Product Previews and Launches.

Beehiiv AI writing assistant: Beehiiv has launched an AI writing assistant to improve the newsletter writing process. Features include text generation with tone of voice options, language translation, "1-click" spell check, and image generation.

Shopify’s new AI assistant: Shopify has introduced its own AI assistant called Sidekick, which will be built directly into the Shopify platform to help business owners. Sidekick will have several features, including the ability to manage tasks, design storefronts, provide details on sales trends, and apply discounts to all items on the site. Shopify's CEO also provided additional details on his Twitter page.

Anthropic’s Claude 2: The second generation of Anthropic's text-generating AI model is now available in beta in the US and UK via a paid API and a web interface. Claude 2 is better than v1.3 at giving harmless responses, explaining its thinking, and performing coding, math, and reasoning tasks.

Google NotebookLM: Google's AI-powered note-taking tool, NotebookLM, recently launched for a “small group of users in the US”. The app allows users to select and ask questions about their documents, create new content, and is designed to provide personalized AI assistance based on the user's own data and notes.

ChatGPT Code Interpreter: OpenAI has made its proprietary plug-in called Code Interpreter available to all ChatGPT Plus subscribers. Code Interpreter allows users to run code and access uploaded files to perform tasks such as data analysis, charting, file editing, and mathematical operations.

Update #4. Company Announcements and News Throughout the Industry.

Employment and Jobs Updates: While there have been several reports on the impact of AI on jobs, some notable examples from the past week include Hollywood actors joining screenwriters in a strike over the use of AI, NYC enforcing its AEDT law to combat bias in AI-driven hiring, and the OECD reporting that over 25% of jobs in its member bloc are at risk of automation:

• Hollywood actors go on strike: Hollywood actors recently joined screenwriters in a strike over wages and the use of AI. During a press conference where members of SAG-AFTRA (Screen Actors Guild - American Federation of Television and Radio Artists) confirmed their strike, it was revealed that Hollywood studios have proposed a controversial AI agreement. The proposal suggests that background actors could be scanned, paid one day's wages, and their digital likeness would be owned by the companies indefinitely for use without further consent or compensation.

• NYC AI hiring law: New York City (NYC) has begun enforcing the Automated Employment Decision Tool (AEDT) law, which aims to reduce bias in AI-driven hiring and employment decisions. Under the law, employers and employment agencies in NYC are prohibited from using AI and algorithm-based technologies to evaluate job candidates and employees without conducting an independent bias audit.

• Jobs at risk of automation: According to the Organization for Economic Cooperation and Development (OECD), more than 25% of jobs in its 38-member bloc (which includes the US) could easily be automated by advances in AI. An OECD survey also found that 60% of workers fear losing their jobs to AI in the next decade.

OpenAI Updates: Since the release of ChatGPT, OpenAI has been in the news constantly, and this week is no different. In the past week, OpenAI has found itself in the middle of an FTC investigation, recently expanded its partnerships with Shutterstock and the Associated Press and is also forming a team to deal with superintelligent AI systems:

• FTC OpenAI investigation: The Federal Trade Commission (FTC) is investigating OpenAI for possible violations of consumer protection laws. The investigation focuses on OpenAI's handling of personal information, potential inaccuracies in the information provided by ChatGPT, and the risk of consumer harm, including reputational harm.

• Shutterstock partnership: Shutterstock has announced a six-year extension of its partnership with OpenAI, allowing OpenAI to continue training its models using Shutterstock's extensive library of images, videos, music, and metadata. The partnership initially began in 2021, allowing OpenAI to use Shutterstock's images to train its AI image generator, DALL-E. Shutterstock also integrated OpenAI's image generator into its website.

• Associated Press (AP) partnership: OpenAI and the AP have announced a groundbreaking partnership. OpenAI will license text content from the AP archives to train its large language models (LLMs). In return, the AP will benefit from OpenAI's expertise and technology, although it has clarified that it will not use generative AI (genAI) to write news stories.

• Creating a Superintelligence team: OpenAI is forming a new team to develop methods for guiding and controlling “superintelligent” AI systems. OpenAI predicts that AI that surpasses human intelligence could emerge within the next decade, raising concerns about the need to control and constrain its behavior. Current techniques for moderating AI rely on human oversight but may not be sufficient for systems that are much smarter than humans.

Language model updates: In the past week, some interesting stories have been shared about LLMs, including the poor working conditions some contractors endure while training LLMs, AI detectors often misclassify non-native English writing as AI-generated text, and LLM performance is highest at the beginning or end of the context window:

• Overworked contractors are training AI models: Contract workers play an important role in improving the responses of AI chatbots like Google's Bard. However, a recent Bloomberg report highlights how these contractors face challenges such as low pay, minimal training, tight deadlines, and are even responsible for reviewing and refining AI-generated responses in areas outside their expertise. Contractors have raised concerns about their working conditions and the impact this may have on the quality of AI products.

• AI text detectors show bias: A recent study found that AI detectors often misclassified the writing of non-native English speakers as AI-generated. More than 50% of the time, the AI detectors incorrectly assumed that the writing of non-native English speakers was AI-generated.

• Limitations of LLM architecture: An interesting paper on LLMs and context input suggests that LLM performance is highest when relevant information is at the beginning or end of the input context, and that performance degrades when accessing information in the middle of long contexts.

HealthCare Updates: Neko Health has raised millions for its AI-powered full-body scans that help detect health conditions. Meanwhile, Google's Med-PaLM 2 AI chatbot is being tested at the Mayo Clinic, showing promise for improving healthcare conversations:

• AI full-body scans: Neko Health recently raised $65 million to build out its AI-driven full-body scans for preventative healthcare. The scans cost roughly $200, take about 10 minutes, and help doctors detect various health conditions, including cancer, cardiovascular disease, and diabetes. Headquartered in Stockholm, the company has a team of 35 doctors, researchers, and technicians.

• Mayo Clinic trials Google’s medical AI chatbot: Google's Med-PaLM 2, an AI tool designed to provide medical information, has been undergoing testing at the Mayo Clinic research hospital since April. Med-PaLM 2 has been trained on a curated set of medical expert demonstrations, making it more suitable for healthcare conversations than other chatbots such as Bard, Bing, and ChatGPT. While still in the early stages, Google sees the potential for Med-PaLM 2 to significantly benefit healthcare with its AI capabilities.

Thanks for reading this issue of the AstroFeather newsletter!

I’m always looking for ways to improve and would love to hear your constructive feedback about the format and content of the newsletter. You can reply to this email, and I’ll be sure to respond.

See you in the next issue!

If you enjoy AstroFeather weekly content, be sure to share this newsletter!

Adides Williams, Founder @ AstroFeather (astrofeather.com)