#22 | The Art & Science of Cancer Care

Quote of the Week 💬 

The World’s Largest Cancer Conference Shared New Findings

⌛ The Seven Second Summary: Last week, the American Society of Clinical Oncology hosted the largest cancer conference in the world. The theme of this year’s event was The Art and Science of Cancer Care: From Comfort to Cure, and a wide array of promising study results were shared.

🧮 Key Results: There were over 200 sessions at the conference. Here are a handful of highlights we found:

• Colorectal Cancer

  • A Phase II trial had 32 participants with high risk stage II or stage III colorectal cancer.

  • 59% of patients who received pembrolizumab treatment had no signs of cancer afterwards.

  • The remaining 41% of patients treated with pembrolizumab underwent surgery to remove their cancer, and 100% of these patients were cancer-free after surgery.

    • In contrast, when standard chemotherapy was given to patients with this genetic profile, fewer than 5% were cancer-free after surgery.

• Lung Cancer

  • A Phase 3 trial had 296 participants with stage III or stage IV lung cancer whose cancers did not improve from previous medical interventions.

  • After 5 years, 60% of patients who received a lorlatinib treatment have not seen their cancer worsen. In contrast, 8% of patients who received a different drug treatment have not seen their cancer worsen.

  • A separate Phase 3 trial reduced the risk of disease progression or death in patients with stage III lung cancer by 84%. These incredible results are why we mentioned the news when the FDA approved osimertinib earlier this year.

• Rectal Cancer

  • A Phase II trial had 41 patients with stage II or stage III rectal cancer.

  • 100% of patients who received the Jemperli treatment became cancer-free.

  • Furthermore, the first 24 patients who received this treatment remained cancer-free by a median of 2 years.

• Skin Cancer

  • Phase II trial had 157 participants with stage III or stage IV melanoma.

  • Patients who received an mRNA cancer vaccine after their melanoma was removed had a 49% lower risk of dying or having the disease reappear after three years compared to patients who did not receive a vaccine.

💡 Why This May Matter: We have mentioned before that medical professionals note it is rare for any medication to achieve a 30% reduction in patient mortality. Yet, many of the study results mentioned above decrease mortality risk or increase patient survival times by far more than 30%. The creativity and variety of approaches we’re seeing in cancer care today illustrates why the conference’s theme mentions the art and science of care. Imagine what we see in 10 years!

🔎 Elements To Consider: We are only scratching the surface of all the promising results mentioned at the conference. There were impressive early diagnostic tests, experimental AI research, and other patient-centric initiatives worth just as much praise as the treatment studies.

📚 Learn More: Penn Medicine. ASCO Meetings.

Stat of the Week 📊 

OpenAI Shares Its Methods To Interpret GPT-4’s Neural Network

16 million

⌛ The Seven Second Summary: OpenAI shared new research explaining methodologies they are exploring to scale up the resourcefulness of architectural systems to understand how Large Language Models (LLMs) conceptualize and interpret language.

🔬 How It Was Done:

• To inspect the activity inside neural network, OpenAI gives an LLM different inputs and searches inside the system’s inner circuitry to find specific neurons that activate when it processes the information it receives.

• Afterwards, they either turn off specific parts of the model’s inner circuitry or provide the model with approximations of previous inputs to see if the same neurons are activated.

  • If they are able to reproduce the same neural activation in different contexts, then they were able to isolate a feature or concept of language that the model processes. If different neurons activate during their verification steps, then they likely found correlating features, rather than causal features.

  • For example, the researchers may deliberately misspell the word “correct” to identify the model’s conceptual representation of misspelled words. To verify which activated neurons represent ‘misspelled words,’ they may need to deliberately misspell other words, misspell the word “correct” in different ways and in different languages, and check to see if the same neurons are activated when additional spelling or grammatical errors are present in the same sentences.

• Note: we explained how Anthropic performs a similar task when they shared similar interpretability research 2 weeks ago.

🧮 Key Results: By improving their mechanism to find activated neurons in an LLM, OpenAI was able to identify and visualize over 16 million features in GPT-4. They also built an interactive webpage to visualize some of the features they found from their work.

💡 Why This May Matter: The ability to holistically interpret the behaviors of AI models will be essential to improving the accuracy and reliability of more advanced systems. Without this insight, models may fester, associate and reinforce information that can lead to unhelpful, deceptive or discriminatory outcomes when applied in the real-world.

🔎 Elements To Consider: Despite OpenAI’s scaling and efficiency improvements to interpret larger portions of an LLM, they are still only able to apply the methodology to a small fraction of their model. The cost of compute is still too high for even the most well-funded research labs to interpret their largest models.

📚 Learn More: OpenAI. arXiv. Github. Web Visual.

AI x Science 🤖

Mamba-2: Next Gen LLM Model Architecture Trains 50% Faster

Last month, Microsoft held its annual Build Conference. At the event, one of the points of emphasis from their CTO, Kevin Scott, was the rate of improvement they are seeing in model costs and efficiencies. “It’s 12x cheaper and 6x faster“ to make calls to GPT-4 since its initial launch ~18 months ago.

Moreover, Scott mentioned AI models are “nowhere near the the point of diminishing marginal returns on how powerful we can make AI models as we increase the scale of compute.” This is noteworthy for a couple reasons. First, not only does more compute lead to more capable models, more compute can also lead to a better understanding of a model’s innate capabilities, as the interpretability research we just highlighted from OpenAI implies. Second, it suggests the field’s longstanding logarithmic scaling laws will still be capable of predicting future model improvement for quite some time, which provides assurances for research labs around the world to invest more in research.

The latest example to compound AI model efficiency belongs to an engineering duo from Princeton and Carnegie Mellon, where they recently released the highly anticipated Mamba-2 architecture.

The researchers improved upon their previous Mamba architecture in two main ways. First, they simplified the system’s core processing to make the main part of the model more straightforward and efficient. Second, they also adopted parallel projections to allow their model architecture to process information and run multiple tasks simultaneously, rather than sequentially. These improvements make Mamba-2 2-8x faster than Mamba-1 at processing various tasks, and their new system is able to train models 50% faster while handling far larger quantities and more complex data. Goomba Lab. arXiv. Github.

Our Full AI Index

• Personalizing Parkinson’s Treatment: Researchers at Mass General Brigham developed an algorithm called Cleartune that uses deep brain stimulation to treat Parkinson's disease symptoms. From an analysis of 237 patients, the team was able to identify specific regions in the brain associated with improving four major Parkinson’s symptoms, such as tremors and muscle rigidity. In a preliminary test of the algorithm, 4 out of 5 patients who received treatments from Cleartune settings showed improved symptoms compared to standard treatments. Mass General Brigham. Nature.

• Atmospheric Forecasting: Researchers at Microsoft developed Aurora, a new AI foundation model to predict weather patterns and atmospheric forecasting. The model was trained on over 1 million hours of weather and climate data, and outperformed other state-of-the-art atmospheric chemistry models. It’s also projected to be faster and more precise than traditional forecasting systems, which should lead to earlier and more accurate extreme weather predictions for the future. Microsoft. arXiv.

• Predicting New Antibiotics: Researchers at the University of Pennsylvania used an AI system to identify and test 63 new drug candidates with promising capabilities to combat disease-causing bacteria and other antibiotic-resistant bacteria strains. Penn Medicine. Cell.

• Enhancing AI Reasoning: A research team from Peking University, UC Berkeley and Stanford developed a new machine reasoning technique called Buffer of Thoughts (BoT). This approach enhances LLM’s accuracy, efficiency, and robustness by storing and adapting high-level thoughts for problem-solving, which allows AI systems to achieve significant performance improvements (11-51%) on 10 challenging reasoning-intensive tasks. Notably, BoT requires only 12% of the cost of multi-query prompting methods and has the potential to enable smaller models to surpass the performance of much larger models. arXiv. Github.

Other Observations 📰

A Major Cause Of Inflammatory Bowel Disease Is Discovered

Researchers at the Francis Crick Institute and other universities identified a new pathway driving inflammatory bowel disease (IBD) and other conditions of chronic inflammation in the digestive tract.

They accomplished this by exploring an area of DNA that does not code for any proteins but has been previously linked to IBD. From their search, they discovered a sequence of DNA that boosts the ETS2 gene's activity in a type of immune cell known as macrophages. After a series of gene edits and drug tests on gut samples they received from IBD patients, the team discovered when they increased the activity of ETS2, macrophages became inflamed and resembled the condition of IBD patients. Conversely, when they inhibited the activity of the ETS2 gene, the same immune cells became less inflamed.

These findings offer a new potential therapeutic target for IBD treatments, and the researchers are already working on ways to deliver ETS2 inhibitors directly to macrophages in patients. The Francis Crick Institute. Nature. YouTube.

Our Full Science Index

• China’s Lunar Landing: China's Chang'e 6 mission successfully landed on the far side of the moon. The goal of the mission is to collect up to 2 kilograms (~4.4 pounds) of soil and rock samples to learn more about the moon's geological history and diversity. BBC.

• Gene Therapy For Deafness: A gene therapy treatment restored the hearing ability in 5 children born with inherited deafness. Mass Eye and Ear’s trial differs from previous studies we have covered insofar as it treated children in both ears as opposed to one. This dramatically improved the children’s hearing ability and did not result in any serious adverse events, which provides promising evidence for the researchers to expand their trial to treat more patients. Mass Eye and Ear. Nature.

• Public Health: A joint report from the WHO and UNICEF shared progress on global access to clean drinking water, sanitation, and hygiene in schools. Between 2015 - 2023, these metrics increased from 66% to 77%, from 68% to 78%, and from 58% to 67%, respectively. This translates to 200+ million schoolchildren worldwide who are now living in better conditions to learn and grow. UNICEF.

• Reducing Air Pollution: We noticed a series of stories last week documenting progress to reduce air pollution in many parts of the world. First, Europe recorded the largest annual decline in PM2.5 — the air pollution most closely linked to harmful health effects — of any region of the world between 2010 - 2019. As a result, deaths in the region from heart disease attributed to pollution fell by 19.2% and from strokes by 25.3%. We also saw a separate report from Paris officials detailing how the city reduced their air pollution by 40% after a series of actions to discourage motor vehicle transportation. Finally, a third analysis from Cell Reports Sustainability estimates the rapid increase of wind and solar generation in the United States has improved the country’s air quality and reduced people’s risk for asthma. These two metrics alone are projected to be worth ~$249 billion in climate and healthcare costs for the country.

Media of the Week 📸 

Robots Completing Multi-Step Tasks From Voice Commands

We’ve shared some of the captivating robotics videos from 1X before, and their latest demo did not disappoint. You can tell there is still a way to go in the robotics voice commands, but the field has come a long way since the Chat GPT boon.

Research Team Discovers Frost Forming Atop Mars’ Tallest Volcano

Scientists from the European Space Agency discovered frost forming on top of the tallest volcanic mountains on Mars, despite the thin air and sunlight that was thought to keep temperatures too high for frost to form. The frost is estimated to weigh at least 150,000 tons and is only about 1/100th of a millimeter thick. Its existence suggests a remnant of an ancient climate cycle on Mars and could help identify any remaining sources of water on the planet.

SpaceX’s Starship Completes Test Flight For The First Time

SpaceX's Starship rocket and Super Heavy booster successfully launched and returned to Earth for the first time, marking a significant milestone in the development of a reusable rocket system. The rocket reached a peak altitude of 132 miles (213 km) and withstood temperatures of up to 2,600° Fahrenheit (1,430° Celsius) during reentry. Time will tell how many more times the system can handle those conditions!