#21 | A Remarkable Lung Cancer Drug

Quote of the Week 💬 

Lung Cancer Drug Demonstrates Remarkable Treatment Results

⌛ The Seven Second Summary: Lorlatinib is a relatively new and remarkably effective drug for treating patients with lung cancer caused by a specific type of error in the ALK gene. 

🔬 How It Was Done: A phase III clinical trial randomly assigned 296 patients with advanced ALK-positive lung cancer to receive either lorlatinib or its predecessor, crizotinib. The drugs work by preventing the ALK gene from overproducing enzymes involved in the abnormal growth and spread of tumor cells.

🧮 Key Results:

• After 5 years of treatment, 60% of patients treated with lorlatinib were alive and did not experience cancer progression, compared to just 8% of patients treated with crizotinib.

• Furthermore, the lorlatinib treatment reduced the risk of disease progression or death by 81% and reduced the risk of brain metastasis progression by 94% compared to crizotinib.

💡 Why This May Matter: As we have mentioned before, the latest generation of targeted therapy treatments are demonstrating signs of cautious optimism in the fight against advanced cancers and other virulent diseases. It will be interesting to see just how far AI and other emerging technological advancements can propel the next generation of targeted therapies forward over the next decade.

🔎 Elements To Consider: People who are ALK-positive with lung cancer tend to be younger than the average lung cancer patient, and they often do not have a history of smoking. Their genetic predisposition to developing lung cancer is one example to remember why so many distinctive medical interventions are required for cancer care. There are well over 100 types of cancers, and each disease has a unique biology, behavior, and treatment profile.

📚 Learn More: Peter Mac. ASCO Publications.

Stat of the Week 📊 

South Korean Launches Its Own Space Agency

2045

⌛ The Seven Second Summary: South Korea announced the creation of the nation’s new space agency, Korea AeroSpace Administration (KASA), to boost its space program and achieve goals like landing on the Moon and Mars

🔬 How It Was Done:

• The agency was created to fulfill a campaign promise by President Yoon Suk Yeol to spur economic growth.

• The country’s plan is to 2x spending on space-related programs over a 5 year time period until its cumulative space projects reach budget of $1.1 billion USD .

• This new agency unifies and builds upon the existing Korea Aerospace Research Institute, which was established in 1989, and was previously the primary space agency in South Korea.

🧮 Key Results: The government has already tasked the agency with landing a spacecraft on the Moon by 2032, and on Mars by 2045. There are also plans to develop satellites, telescopes and other commercial launch capabilities to help the country accomplish its goal and produce its own space research.

💡 Why This May Matter: South Korea successfully launched Nuri, a 200-ton rocket, into orbit last year. The country has been prioritizing and delivering on its space exploration goals in recent years, and it is worthwhile to track their incremental progress to their 2032 goal.

🔎 Elements To Consider: Last year, India became the 4^th country to land a spacecraft on the moon; in January, Japan became the 5^th country; in February, the United States successfully landed a spacecraft on the moon for the first time in 50 years. The space exploration industry is growing exponentially, as more countries invest in core capabilities to monitor and traverse the cosmos.

📚 Learn More: Science. Korea Times.

AI x Science 🤖

Will Multi-Modal Chatbots Actually Understand Us Better?

Since the release of Chat GPT-4o, we have been contemplating the future of user experiences (UX) with AI applications.

In order to foster trust in an AI solution, transparency, explainability, control and other features should be designed into the system. When done well, these attributes allow people using an AI tool to understand how the machine arrived at its decisions and provide feedback for the model to improve its performance and reliability.

In Large Language Models (LLMs), the written prompt is the input people control to guide a LLM to more helpful responses. There are numerous online resources and courses to teach people how to improve their LLM communication skills, including documentation from Anthropic and Google. What is interesting about the phenomenon of prompt engineering is it implicitly describes why the meaning and intent behind words and sentences are so important.

As we have mentioned before, the future development of LLM reasoning appears to be morphing into a type of tree search and evaluation problem. Meaning, when an LLM receives a specific question or task, it searches through a vast array of permutations of possible responses in a language, then evaluates and iteratively refines its answer until it provides a correct answer and is rewarded. Over time, the desired response from the LLM is reinforced, and the model becomes more efficient at arriving at a desirable answer with fewer searches. Importantly, when a LLM receives a prompt with more clarity, it becomes more effective at narrowing down its search and retrieving relevant information to inform its response.

However, since LLMs are predictive models, when they receive complicated, unclear or conflicting information, they tend to assume a person’s meaning and intent rather explicitly clarify with the person in a conversational manner. As a result, the model searches through a larger surface area to retrieve a relevant response, which makes it more susceptible to veering off course if it misinterprets what a person wants and why they shared certain information with it.

While there are ways to mitigate for this issue, it will be interesting to see if the engineers designing multi-modal models will have them behave fundamentally differently from their text-only predecessors. The best conversations are often dynamic exchanges, and we may have finally reached a point where conversational models are low latency, expressive, and inquisitive enough to provide a conversational experience that can flow with a natural, back-and-forth cadence. If this is true, multi-modal models may both sound more convincing and actually be more reliable. Not because research labs made foundational breakthroughs in model architecture or design, but because they implemented an elegant UX for machines to interact with and understand humans better. Do you think Apple will deliver this experience when they re-launch Siri?

Our Full AI Index

• AI Reads Intensive Care Charts: Researchers at Duke University developed an assistive machine learning model to help medical professionals read electroencephalography charts of intensive care patients. The model helps medical professionals visualize and identify patterns in brainwaves that may lead to seizures and brain damage. In a test, the model improved the accuracy of medical professionals' diagnoses from 47% to 71%. Duke University. NEJM AI.

• Synthetic Data: Google researchers shared a paper detailing some of the work they are pursuing to generate synthetic, high-quality data that is tailored for specific tasks without relying on human feedback. The framework relies on a model to generate the synthetic data as well as other services to annotate, filter and tailor the generated data for specific purposes. arXiv.

• Smart Bandages: Researchers from the University of Arizona are developing smart bandages to remotely monitor wounds, detect infection, and deliver precise treatments to the body when necessary, such as doses of electrotherapy or antibiotics. The bandages use advanced technologies like flexible electronics, biosensors, and wireless transmission systems to provide real-time information to doctors who are monitoring the recovery process of a patient’s wound. The team’s ultimate goal is to create bandages that can detect and treat wounds autonomously, reducing the need for medical intervention. Wall Street Journal Archive. Nature.

Other Observations 📰

Positive Updates On The World’s Renewable Energy Transition

In January, we shared a report from the International Energy Agency projecting “global energy-related CO₂ emissions will peak by 2025.” In the months since, a number of encouraging data points are suggesting society may be ahead of schedule.

For example, two separate analyses last week suggest China’s CO₂ emissions peaked in 2023 and will start declining from 2024 onwards. First, the Center for Research on Energy and Clean Air cited China’s CO₂ emissions fell by 3% in March, and Bloomberg has reached similar conclusions. China’s unprecedented renewable energy transformation has been one of the most significant developments in the world’s clean energy transition — they contributed 63% and 65% of new wind and solar energy in 2023, respectively, and their growing adoption has dropped the costs of these technologies to the point where they are now the cheapest source of electricity by a large margin.

Furthermore, after an extended period of project delays, India installed over 10 GW of solar in Q1 2024, which was a 414% year-over-year growth rate. There are plenty of other projects in their pipeline, which is a promising sign that their renewable energy transformation may accelerate from here.

Also, in Q1 2024, the United States’ Energy Information Administration recently announced the nation’s renewable energy generated 40.5% of the country's electricity. For the first time in generations, the country is getting close to the point where, fossil fuels will provide less than half its electricity.

All this is to reinforce the Carbon Brief analysis predicting global emissions may fall as much as 2.5% in 2024. There is still a lot more work ahead to reach the 2030 COP 28 goals, but these are all encouraging signs the world’s ambitious climate aspirations are still well within reach with more investment and a greater sense of urgency.

Our Full Science Index

• The Thinnest Lens On Earth: Physicists from the University of Amsterdam created the world's thinnest lens — just 3 atoms thick — by using a unique material called tungsten disulphide. University of Amsterdam. ACS Publications. 

• Complete Chromosomal Genome Of Primates: Researchers from Penn State, the National Human Genome Research Institute and the University of Washington generated complete reference genomes for the sex chromosomes of 5 great ape species and 1 lesser ape species. This achievement helps researchers study and understand the evolution of sex chromosomes and related diseases in some of our closest relatives in the animal kingdom. For example, the study discovered the Y chromosome is highly variable across ape species. Penn State. Nature.

• Breast Cancer Blood Test: A team from The Institute of Cancer Research in London developed a blood test to detect tiny amounts of cancer DNA in the blood of breast cancer patients. The test utilizes whole genome sequencing and was able to predict the recurrence of the disease up 41 months before symptoms appeared. The Institute of Cancer Research. JAMA Network.

• Synthesizing Rare Earth Elements: Researchers at Oak Ridge National Laboratory determined the properties of promethium, a rare earth element, for the first time. To accomplish this feat, the scientists bound promethium with special organic molecules and used X-ray spectroscopy. This allowed the team to determine properties such as the length of the material's chemical bond and its proximity to neighboring atoms. Oak Ridge National Laboratory. Nature.

Media of the Week 📸 

Teaching A Robot To Keep Its Balance In Difficult Situations

Researchers from ETH Zurich developed a way to test how well legged robots can move around in unexpected situations. They used a machine learning technique known as an adversarial network to identify weaknesses in the robots' decision-making systems. Afterwards, they refine the weaknesses in the system, leading to a robot that can navigate its environment more safely and reliably. Github. arXiv.

The First Manhattanhenge Of The Year

Researchers from the University of Plymouth developed an AI model to identify and analyze key developmental events in embryos from videos. The model can detect features such as heart function, crawling, and hatching, and may be used to study how climate change and other external factors affect embryo development in humans and animals. University of Plymouth. Journal of Experimental Biology.

NASA’s JWST Finds Most Distant Known Galaxy

NASA's James Webb Space Telescope recently discovered the most distant galaxy ever observed. JADES-GS-z14-0 is a galaxy that existed just 300 million years after the Big Bang, and is surprisingly bright and large for how young it is. The galaxy has a mass of hundreds of millions of times that of our Sun, and its discovery challenges current theoretical models of galaxy formation and evolution in the early universe. Webb Space Telescope. arXiv.

This Week In The Cosmos 🪐

June 6: A new moon. The best time to stargaze!