#29 | A Gene Therapy To Treat Cancer

Quote of the Week 💬

FDA Approves First Gene Therapy to Treat A Rare Form Of Cancer

⌛ The Seven Second Summary: The U.S. Food and Drug Administration (FDA) approved TECELRA (afamitresgene autoleucel) to treat a rare and aggressive form of cancer called synovial sarcoma once it has metastasized or cannot be removed with surgery. 

🔬 How It Was Done: 

• The FDA’s approval was based on the results of the SPEARHEAD-1 trial that included 44 patients with an inoperable and metastatic form of synovial sarcoma.

• The enrolled patients were also people who received prior therapy treatments with little to no success, and whose tumors contained a MAGE-A4 antigen.

• The patients received a single dose of TECELRA, which is an immunotherapy treatment that modified their body's own immune cells to recognize and target the MAGE-A4 protein marker found on the surface of their cancer cells.

🧮 Key Results:

• Out of the 44 patients who received TECELRA, 19 (43.2%) saw their cancer meaningfully shrink.

• Of those who responded to TECELRA, 2 (4.5%) had their cancer completely disappear.

• The median effects of this treatment lasted for at least 6 months, and 39% of responders were still seeing benefits after 12+ months.

💡Why This May Matter: This was the first immunotherapy approved to treat this specific type of cancer, which is a notable milestone. Also, recall the many jaw-dropping clinical trial results from meetings at the largest cancer conference in the world back in June. A number of these treatments are immunotherapies, and the FDA’s TECELRA approval suggests several more promising cancer treatments may receive approvals in the months ahead.

📚 Learn More: FDA. The Lancet.

Stat of the Week 📊

Brain Waves & AI Systems Diagnose Neurodegenerative Diseases

6

⌛ The Seven Second Summary: Researchers at the Mayo Clinic’s Neurology Program developed an AI-powered tool to analyze and differentiate the brain waves of cognitively normal individuals from those with Alzheimer's disease, Lewy-body dementia and other forms of neurodegeneration.

🔬 How It Was Done:

• The scientists gathered a dataset of 12,176 brain wave recordings (EEG) from over 11,000 patients for their model to analyze and identify patterns in brain activity.

• Their tool was able to identify 6 distinct patterns derived from the patient dataset of brain wave activity and determine if abnormal cognitive patterns were indicative of cognitive issues.

• The results of the tool were also compared to traditional measures of dementia, such as cognitive exams, biomarker tests, and brain imaging scans.

🧮 Key Results:

• The team’s tool was highly accurate (83%) in distinguishing between healthy individuals, those with mild cognitive impairment, and those with a type of dementia disease.

• The tool's results were closely associated with established methods to diagnose dementia, although it did not outperform these other methods in most cases.

• The study suggests Alzheimer's disease, Lewy-body dementia and different types of cognitive decline have unique brain wave patterns that are difficult for humans to notice but possible for AI systems to detect. For example, the pace, variability and rhythm of a person’s brain waves are all indicative of their brain’s cognitive health, and can reveal how severe their neurodegenerative disease is.

💡 Why This May Matter: As we mentioned in our Alzheimer’s deep dive article, one of the challenges with dementia diagnostics is people often have mixed dementia, which is when a brain has changes associated with more than one type of dementia. It is exceedingly difficult for clinicians to determine which symptoms are due to which dementia, and advanced machine learning techniques are a fantastic use case to improve the precision in which clinicians diagnose and differentiate patients with one or more dementia diseases.

🔎 Elements To Consider: These AI-based tools still need a lot more work to become as reliable as other types of dementia exams, but their resourcefulness is a reminder of how useful tools + expert decision making = better outcomes than expert opinion alone.

📚 Learn More: Mayo Clinic. Brain Communications.

AI x Science 🤖

Distilling The Knowledge Of Large Models Into Small Models

Back in June, Google Deepmind released Gemma 2, which is the company’s next generation lightweight and open source language models.

This past week, the company shared the technical report to accompany their new family of open-source models, and a major takeaway we wanted to highlight from the team’s research was the use of knowledge distillation to train their small models to reach performance levels close to their largest, state-of-the-art models.

Deepmind achieved its feats in their small models through a student-teacher setup, where the small Gemma 2 model attempt to emulate much larger models. This is possible by having the large teacher model and the small student model train through different methodologies.

Next token prediction is the predominant way large models understand language and predict the most likely character, word, or phrase in a language sequence. Since the methodology is so granular, the models are able to identify countless patterns and relationships about a language in its training data, and can produce highly coherent and contextually relevant text when prompted. In contrast, Deepmind’s small student model is not trained on token prediction. Instead it is trained to predict the probability distribution of its teacher model so it can predict the most likely answer without searching through its more limited corpus of knowledge.

For example, if a teacher model is prompted to complete the phrase “green eggs and” its probability distribution of the next token may be something like {ham: 0.8, spinach: 0.1, grits: 0.05, …}. If a student model were to complete this same phrase through next token prediction, its probability distribution may be weaker, like {ham: 0.5, avacado: 0.25, magic: 0.13, …}, because it has fewer parameters to find relevant patterns from its training data. To overcome the small model’s limitations with token predictions, the research team has it predict the probability distribution its teacher will produce, rather than produce its own probability distribution from scratch.

The benefit of this technique is it allows a student model to learn important language patterns and associations by proxy of the teacher’s behavior. These details include the most likely next token in a language sequence {ham}, as well as possible token alternatives {spinach, grits} and the relative likelihood of each possible token {0.8 vs 0.1 vs 0.05}. This process is referred to as knowledge distillation because the student model is learning to distill the knowledge of the teacher model into a smaller, more efficient form.

Ultimately, Deepmind’s insights are informative because they partially explain how research labs like Apple, Microsoft and OpenAI are innovating their ways to smaller, cheaper and more efficient AI models. Thos ongoing trend is a welcome sign to see because it means more capable systems are approaching the point where they can run locally on devices, rather than on remote servers or cloud computing platforms, which pose greater risks to privacy, security, and latency.

Our Full AI Index

• A Visual Model To Identify Objects: Shortly after the release of Meta’s Llama 3.1 models, Meta open-sourced SAM 2, which is a computer model to identify and track objects in videos. They built a special memory system to help the model remember what it has seen before, so it can accurately identify objects even when they move or change. The model was trained using a large dataset of 51,000 annotated videos, and it learned to improve its predictions through repeated cycles of feedback and updating. Meta Paper. SAM 2. Github.

• Drones To Look Inside Hurricanes: The BBC shared a cool story about how scientists from the United States’ National Oceanic and Atmospheric Administration are partnering with the private company Saildrone to use drones to collect data on oceanic and atmospheric conditions of developing hurricanes. The department’s hope is to understand how they intensify to build a more complete picture of storms and improve their computer models. BBC.

• White House Supports Open-Source: After reviewing the benefits and risks of open-source AI models with publicly available weights, The National Telecommunications and Information Administration (NTIA) issued its policy recommendations for more open AI development. This is an encouraging sign to see because Big Tech companies are largely driving the market forward with their immense data, hardware and resource advantages that small research departments struggle to compete with. NTIA. Full Report.

Other Observations 📰

Using Insects To Turn Dangerous Waste Into Valuable Products

A new blood test called APS2 was developed to diagnose Alzheimer's disease by measuring levels of two key proteins — beta-amyloid and tau — in the blood. The test was evaluated in a study of 1,213 patients with cognitive symptoms who were also assessed by primary care physicians and dementia specialists, and the results were comparable to current gold standard tests for Alzheimer's, such as spinal fluid tests and PET scans.

The APS2 test was developed by researchers at Skane University Hospital and Lund University, and was found to be 91% accurate at diagnosing Alzheimer's disease. There are now over a dozen different Alzheimer’s blood tests at various stages of development, and it is becoming more commonplace to see new blood tests that can correctly classify 90% - 95% of patients as either positive, negative or borderline cases.

The field is fast approaching the day where the first Alzheimer's blood test is approved by a major regulatory authority. When the day arrives, Alzheimer's dementia can evolve into manageable conditions like diabetes and HIV/AIDS, because clinicians will be able to detect the dormant disease years before serious problems arise and track how people respond to specific treatments in near-real-time with cheap, convenient, and reliable tests. Lund University. JAMA.

Our Full Science Index

• Half Of Cars Sold China Are Electric Vehicles: China's plugin vehicles accounted for a 49.9% market share of new car sales in June. This is the first time China’s market reached this milestone, and the country’s pace of renewable transformation suggests the largest auto market in the world may be fully electric by 2030. As a reminder, the International Energy Agency estimates 60% of global new car sales need to be electric by 2030 in order to keep pace with its 2050 net zero emissions target. So far so good. Clean Technica.

• A Digital Twin For Human Blood Circulation: Quanta Magazine featured a fascinating interview of a computer scientist who uses 3D images of a patient's blood vessels to simulate the fluid dynamics in their body and forecast their expected blood flow. Their system can accurately simulate and forecast a person’s blood flow about a week in advance, and may be an important piece in a scientist’s future toolkit to personalize medical treatments at an individual level. Quanta.

• How Flies Can Turn Waste Into Useful Products: Researchers at Macquarie University shared a paper documenting how to genetically modify the diet of black soldier flies to include contaminants like PFAS and other waste as an effort to produce industrial enzymes, specialized lipids, animal feed, and crop fertilizers. There are specific genes in the insect’s digestive system that allow the species to eat many types of organic waste already, and with some fine-tuning, the scientists demonstrate how the insects are able break down more complex organic materials and contaminants for biomanufacturing purposes. Nature.

Media of the Week 📸

Playing 20 Questions With OpenAI’s Advanced Voice Mode

Last week, OpenAI allowed a select cohort of paying customers to trial their voice feature for the first time. There are already countless online videos of people sharing their experience with the company’s forthcoming product, but this example stood out to us. In order for a model to reliably solve 20 questions it must have sufficient deductive reasoning skills, a long enough context window to remember all relevant information from an ongoing conversation, and ask efficient questions to reach the answer within its 20 question limit.

Some Large Language Models (LLMs) are already be capable of this feat over text conversations, but the system’s anthropomorphic qualities continue to amaze.

A Map Of How The Brain’s Blood Vessels Age Over Time

Researchers from Penn State created detailed maps of a mouse brain's vascular network, revealing age-related changes in blood vessels to identify vulnerable areas that may contribute to neurodegenerative diseases. Their results showed a 10% decrease in vascular length and branching density in aged brains, as well as "leaky" blood vessels. Penn State. Nature.

New Details Of The ‘Screaming Woman’ Mummy Discovered

Researchers from Cairo University "virtually dissected" a 3,500-year-old "Screaming Woman" mummy by using CT scans, electron and infrared microscopes, and other advanced techniques. Their analysis found the woman was likely 48 years old, had a case of mild arthritis, and a few other notable characteristics. Scimex. Frontiers.

This Week In The Cosmos 🪐

August 12/13: The Perseids Meteor Shower will be at its peak. It is one of the brightest and most active meteor showers of the year, and the visibility will be decent with an accompanying half moon in the sky.