#32 | Major Lung Cancer Trials Begin

Quote of the Week 💬

Lung Cancer Vaccine Trials Launch Across Multiple Countries

⌛ The Seven Second Summary: The world’s first lung cancer vaccine launched its phase I clinical trial across 34 research sites in seven different countries: the United Kingdom, the United States, Germany, Hungary, Poland, Spain and Turkey.

🔬 How It Was Done: 

• This messenger RNA (mRNA) vaccine carries information about specific biomarkers found on a specific type of lung cancer cell.

• When the vaccine is administered, the mRNA enters the patient's immune cells and helps them recognize the biomarkers of the target lung cancer cells.

• Approximately 130 patients with stage 2 - stage 4 lung cancer will be enrolled across the 34 research sites.

🧮 Key Results: We have to wait several months for the initial results of this new trial to be reported, although there are two worthwhile data points to keep in mind:

• Early indicators of efficacy: BioNTech reported in their Q1 2024 earnings that when their mRNA lung cancer vaccine was administered in combination with another cancer treatment, 85% of patients had their cancer stabilized and controlled, and 30% of patients had their significantly cancer shrink.

• Other mRNA cancer vaccines are promising: 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 3 years compared to patients who did not receive a vaccine.

💡 Why This May Matter: Lung cancer remains the leading cause of cancer deaths worldwide, largely due to persistent air pollution and firsthand and secondhand smoke exposure. While systemic factors contributing to lung cancer still need to be resolved, a breakthrough solution to treat the illness can save countless lives in the years ahead.

🔎 Elements To Consider: The primary objective of the clinical trial is to assess the safety of the mRNA vaccine and understand its side effects at different doses and when taken alongside common cancer treatments.

📚 Learn More: UCLH. NIH Clinical Trials. BioNTech.

Stat of the Week 📊

AI System Predicts & Simulates An Interactive Game In Real Time

20

⌛ The Seven Second Summary: Google researchers developed an AI system to power a video game engine and generate an interactive first-person shooter game for their team to play.

🔬 How It Was Done:

• The team trained an AI agent to play DOOM, a popular video game that was first released in 1993. The team recorded the AI agent as it played DOOM for several hours. They did this so their system could catalog and familiarize itself with various gameplay strategies and scenarios.

• Next, the team used the agent’s recorded gameplay to train a model to predict the next DOOM frame to appear based on the previous frames and actions taken by the AI agent.

• Afterwards, the team fine-tuned the model to make its predictions fast enough and efficient enough to simulate DOOM in real-time for actual gameplay. This involved optimizing the model's architecture and training parameters to replicate the original game’s fidelity.

🧮 Key Results:

• Their model could generate the next frame of DOOM with an image quality similar to that of a compressed JPEG image. In fact, the model achieved a Peak Signal-to-Noise Ratio score of 29.4, which is remarkably close to the maximum possible score of 30, and is an indication of how well GameNGen recreated high-quality images of the original DOOM game.

• The team named their system GameNGen, and it can generate a 1993 video version of DOOM smoothly up to 20 frames per second on a single computer chip. For reference, the original game played at 35 frames per second, which is a large enough gap for some, but not all people to recognize.

• For example, the team asked 10 people to watch 130 short clips of the original game and the AI simulation playing side-by-side, and then guess which clip belonged to the original game. The raters correctly identified the actual game over the simulation only 58% - 60% of the time, which indicates the AI system's high video quality.

💡 Why This May Matter: The main aspects of this research involve image generation and video compression to create replicas of an original piece of art. We typically only mention this type of AI and machine learning development in cellular and other science-based contexts, but this study, as well as some photos we share in our Media section this week, are notable flashpoints of how close we are until certain AI generated images and videos become indistinguishable from human generated versions.

🔎 Elements To Consider: The clips the team showed people to rate were only 1-3 seconds long because the simulation's quality degrades over longer time periods. This means there is still a way to go for their model to generate an immersive gaming experience for the hours players dedicate to learning and perfecting an actual game in one sitting, let alone the lifetime of hours they spend playing a game.

📚 Learn More: GameNGen. arXiv.

AI x Science 🤖

The Energy Required To Scale AI

Last week we mentioned the detailed report by Epoch AI researchers attempting to forecast the scale of AI models at the end of the decade. Today we want to focus a bit more on this paper as a way to explain energy consumption — the most significant limiting factor to AI progress today.

As a reminder, the Epoch AI teach estimates the AI industry will be able to scale their models by 10,000x from where we are today by 2030. They reached this conclusion because Meta already detailed in their Llama 3.1 paper that they needed 27 megawatts (MW) of power in order to train their largest 405 billion parameter AI model. The engineers at Meta also explained how temperature changes fluctuated the power consumption of their data center while they were training their model, and these fluctuations were “on the order of tens of MW, stretching the limits of the power grid.“ This means there is not enough regional power for AI research labs to centralize all of their model training in the data centers of today without causing power outages and other unintended consequences.

To resolve this issue, research teams are experimenting with ways to train their AI models in a geographically distributed manner. If successful, they can split power consumption across multiple locations instead of relying on one grid. Google first mentioned progress with this challenge in its initial Gemini paper, however the problem remains unsolved. Since there are at least half a dozen mega research labs in the United States (US) alone with insatiable power demands, the companies have decided to address their energy constraints through as many means as possible instead of wait for one solution.

This is why AI research labs are working with government entities to build massive data centers with substantial energy provisions. While similar news has trickled in throughout the year, its culmination is a report a few days ago by The Information. The article claims Josh Teigen, US Commissioner of Commerce, was approached by multiple AI companies with separate $125 billion plans to build data centers initially consuming 500-1,000 MW and growing to 5,000-10,000 MW.

Once online, these data centers will provide enough energy capacity for AI research labs to train models equivalent to GPT-6. For reference, Epoch AI estimates 6,000 MW are needed to train a GPT-6 level model, equivalent to ~30% of global data center power not used for AI training today. If logarithmic performance improvements continue to scale with historical trends, a GPT-6 system will be astounding. However, if improvements diminish as models grow more complex, new breakthrough research will be needed to advance AI progress. We have previously mentioned the industry’s exponential efficiency curve that Microsoft’s CTO Kevin Scott explained. While Scott’s assessment remains true, Mark Zuckerberg is also correct when he says the industry does not know when its exponential curve will flatten out. The only way to find out is to sign a blank check.

Our Full AI Index

• AI System Differentiates Various Types Of Cells: A cross-functional team of researchers from the Center for Genomic Regulation and other organizations developed an AI-based tool to differentiate cancer cells from normal cells. They also trained their system to detect early stages of viral infections by enhancing medical images to have higher resolution photos than their original copies. This imaging technique allowed their AI system to develop remarkable displays of precision at a nanoscale. In fact, after training, their tool was able to detect and classify cellular changes as small as 20 nanometers (nm) in size within an hour of exposure. For reference, 20nm is much smaller than many infamous viruses, such as HIV, influenza and coronavirus, although it may be slightly larger than some of the proteins, molecules and receptors that cancer cells and viruses interact with to damage healthy cells. Nonetheless, it is a great achievement for researchers and clinicians around the world to be able utilize these sorts of tools to understand how to diagnose and treat ailments. Center For Genomic Regulation. Nature.

• Machine Learning System Predicts Major Earthquakes: Researchers at the University of Alaska Fairbanks used a machine learning algorithm to analyze seismic data from two major earthquakes (6.4-7.1 magnitude) in Alaska and California during 2018 and 2019. They found abnormal levels of low-magnitude seismic activity across 15-25% of the impacted regions ~3 months prior to each major quake. Their algorithm also learned to predict whether a major earthquake would occur within 30 days or less with up to 85% probability scores. Hopefully, this research and insight can help seismologists and emergency responders identify precursors to destructive quakes, and provide communities at risk with days, if not weeks, of advance notice to minimize damages and casualties. University of Alaska Fairbanks. Nature.

Other Observations 📰

The Largest Catalog Of Bacteria, Fungi & Microbes Found In Food

Researchers at the University of Trento created the most comprehensive catalogue of microorganisms found in food to date. They examined yogurts, cheeses, meats, fermented foods and more to build a dataset of 2,533 foods from around the world. They compared the microorganisms in these foods to those found in human gut samples to see what sorts of relationships they could find in their data.

The team discovered some overlap between microorganisms in food and human gut samples, although the data is too noisy to determine if food microbes enter the human gut through direct consumption or other environmental factors. Interestingly, the study found a greater overlap between gut microbes and certain foods in younger individuals. For example, adult gut microbes showed a 3% overlap with food microbes, compared to 8% in children and over 50% in newborns. The reason for this is unclear, but it may be partly due to the fact that the human gut has fewer microbes at younger ages, making any overlap more pronounced. Nonetheless, the dataset released by this study will be a valuable resource for other researchers to explore correlations between food microbes and human health. Nature. Cell.

Our Full Science Index

• New Region Discovered In The Earth’s Core: Researchers from The Australian National University discovered a previously unknown doughnut-shaped region within Earth's liquid core. They found it by using special tools and techniques to monitor and detect an area with slower seismic wave speeds than the surrounding mass. After an investigation, the team realized this portion of the Earth’s core has a distinct composition and magnetic field, although it is far too early for the full implications of their discovery to be fully understood. Australian National University. Science Advances.

• More Promising Renewable Energy Trends: In December 2020, China set a goal to increase its clean energy by ~2.5x within 10 years. The country has now officially reached its target, 6 years ahead of schedule. China’s recent energy transformation is the main reason why global emissions will likely peak this year and continually fall from here on out. It is also the main reason why the world still has a solid chance to achieve the International Energy Agency’s 2030 goals to reach net zero emissions by 2050. In other news, India installed 15 GW of solar power in 1H 2024, which is a 282% increase compared to India’s solar installations in 1H 2023. Battery costs have also declined by over 90% in the past decade, and reportedly over 50% in the last 18 months. These factors should continue to compound and grow the world’s solar power production at a rapid pace for years to come.

Media of the Week 📸

Watch An Insect Complete The World’s Fastest Backflip

Researchers from North Carolina State University used high speed cameras to study the jumping behavior of springtails in slow motion. The team’s camera can capture images at 40,000 frames per second, and this fidelity allowed the scientists to calculate the insect accelerates into a jump in just 0.001 seconds and reaches a peak torque rate of 368 rotations per second. The springtails are also able to launch themselves over 60 millimeters into the air, which is more than 60x their own height.

Simone Byles is the GOAT human gymnast, but these insects are on another level. If only they could clean up their sloppy landings. North Carolina State University. Integrative Organismal Biology.

JWST Captures Another Beautiful Photo Of A Star Cluster

The James Webb Space Telescope captured another photo of a star-forming cluster with its superb sensitivity and infrared technology. The gas and dust clouds are hallmarks of a new star taking shape, and astronomers estimate the region is only 1 - 3 million years old. This star cluster is approximately 960 light-years away from Earth and is the type of space beauty we can admire all night. ESA.

Image Generation AI Models Are Getting Better & Better & Better

As we alluded to in our Stat of the Week section, AI image generation has already reached exceptional levels in 2024. There is still a specific sheen or texture apparent in many AI generated images of people and animals, but the evidence of computer generation is much harder to notice in certain objects and stylized designs.

If you’re wondering what the giveaways are in these two examples: the letters on the car tires are gibberish; the embers of the fire have awkward shapes and drawn in unnatural places.

In spite of these minor imperfections, they are easy to overlook at first glance, and even easier for a trained professional to edit these first drafts into something they may actually use for their work. With this type of image fidelity, it’s no surprise AI systems are able to identify cancer cells and virus interactions at nanometer scales. Ideogram.