#26 | Ancient Chromosomes Discovered

Quote of the Week 💬 

Fossils Of Ancient Woolly Mammoth Chromosomes Discovered

⌛ The Seven Second Summary: A team of scientists from Baylor College of Medicine, the University of Copenhagen, and the Centre Nacional d’Anàlisi Genòmica discovered tiny chromosomal fossils in the remains of a 52,000-year-old woolly mammoth.

🔬 How It Was Done:

• The researchers obtained a skin sample from a 52,000-year-old woolly mammoth found in Siberian permafrost and used advanced microscope techniques to image the skin sample and identify fossilized chromosomes.

• Afterwards, the researchers used computational tools to analyze the microscope images and reconstruct the 3D structure of the fossilized chromosomes.

• Once their 3D chromosomal structure was built, the scientists were able to compare the ancient DNA molecules to modern elephant species to identify changes in their genetic code.

🧮 Key Results:

• Some of the differences the research team identified were the genetic variants associated with some of the skin and hair follicle changes between a woolly mammoth and modern elephant. This insight helps to explain why the woolly mammoth has its thick layers of fur.

• This research was possible because the ancient animals’ chromatin loops were preserved in a frosted state for thousands of years. Chromatin loops are typically 50 nanometers in size, and are twisted structures that help pack DNA into the nucleus of a cell.

💡 Why This May Matter: Fossilized chromosomes are extremely valuable discoveries because they contain orders of magnitude more data about an ancient artifact’s genetic code than any DNA fragments a fossil may leave behind. This research team believes they can assemble the entire DNA sequence of a woolly mammoth from their single chromosomal fossil.

📚 Learn More: Baylor College of Medicine. Cell Press.

Stat of the Week 📊 

AI Model Used To Diagnosis Different Forms Of Dementia

26%

⌛ The Seven Second Summary: Researchers at Boston University developed a machine learning-based tool to identify patterns in brain scans and medical history, and to help clinicians make more accurate diagnoses of dementia.

🔬 How It Was Done:

• The researchers gathered a vast amount of data from multiple sources, such as a patient’s personal details, medical history, test results, and brain scans. They combined all this multi-modal data together into one large dataset from 51,269 participants across 9 distinct geographical locations.

• Afterwards, the team trained a model with their dataset data by randomly hiding certain information for various tests. By doing this, they trained their model to piece together missing information and strengthen its ability to recognize relationships between data points.

• The team also used a 3D-based model architecture to analyze brain scans, identify complex patterns, and distinguish 10 different types of dementia from the images it received.

🧮 Key Results:

• One of the key metrics the team measured their model on was its receiver operating characteristic (ROC) curve. It is similar to an accuracy score insofar as both are point estimates, and a ROC score of 0.5 indicates random guessing while a ROC score of 1.0 indicates perfect performance.

• Neurologists improved their diagnostic performance by 26% across all 10 dementia types when they used the researcher's AI tool, compared to when they diagnosed patients without an augmented tool.

• The team's model also outperformed the assessments conducted by neurologists alone, achieving an average ROC score of 0.96 across all 10 dementia types. The model also showed promising performance in diagnosing mixed dementia cases, achieving a mean ROC score of 0.78 for patients with two types of co-occurring dementias.

💡 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, but thankfully, robust disease biomarkers and advanced machine learning techniques are two emergent techniques that can be sharpened to differentiate dementia diseases with high precision.

🔎 Elements To Consider: The datasets the model was trained on primarily consisted of Alzheimer’s dementia cases, which may skew the model’s performance towards introduce a performance imbalance with other types of dementia disease diagnostics.

📚 Learn More: Boston University School of Medicine. Nature.

AI x Science 🤖

AI Costs Continue To Decline

Over the past year, we have covered substantial progress the industry of Artificial Intelligence (AI) has made to improve hardware, software, data quality and algorithmic innovations. An advancement in any single component pushes the field forward in either accuracy, reliability, speed or efficiency. While AI systems have become more accurate and reliable over the past year, the most significant development of the year from our vantage point is how much cheaper AI systems are becoming.

For example, in a recent interview, Olivier Godement, OpenAI’s Head of API Product, mentioned GPT-4 costs have decreased by 85% - 90% since its launch last year. He anticipates these costs will continue to meaningfully decline for the foreseeable future, which aligns with what Microsoft’s Chief Technology Officer expects as well.

A separate example paints an even sharper picture about the industry’s pace of improvement over a slightly longer time horizon. Andrej Kaparthy, one of the founding engineers of OpenAI, released a tutorial about how to reproduce GPT-2. During his explanation, Kaparthy mentions an engineering team was required to develop GPT-2 when it was first released in 2019. But because the field is experiencing exponential improvements in hardware, software and data quality, an individual can now accomplish the same feat in less than 24 hours for less than $1000 USD.

Kaparthy did not mention algorithmic innovations in his tutorial discussion, but it is just as notable. Just a few weeks ago, a Google Deepmind research team shared a paper explaining a new technique they are experimenting with to drastically cut down the time and compute costs to train an AI model. Their methodology is called JEST (Joint Example Selection Technique), and the process involves two distinct models: one ‘learner’ model to identify and select the most valuable training data in a dataset, and a ‘reference’ model to serve as a proxy for how difficult a task might be to help the ‘learner’ model improve its data assessments.

Their ‘learner’ model is trained on a specific task, such as image classification or language translation, and calculates a learnability score for a batch of data by evaluating the difference between its performance and the ‘reference’ model's performance. A higher score indicates a batch of data is challenging for the ‘learner’ model but relatively easy for the reference model, which makes it a valuable learning opportunity for the ‘learner’ models to train on.

When the team scaled up their JEST system, they were able to improve state-of-the-art training processes for a large-scale multimodal model by an order of magnitude: 13x fewer training data examples were needed, which resulted in a 10x more compute efficient training process. These sorts of breakthrough efficiency gains are meaningful, because lower compute costs can unlock important safety features for AI research labs to holistically interpret the behaviors of their largest models, as just one example.

As far as the industry has come, it is still at the surface of its potential in many ways.

Our Full AI Index

• Models For Robot Reasoning: Researchers from UC Berkeley and Stanford University developed a new way for robots to improve how they learn and reason through a new methodology. Their approach is called Embodied Chain-of-Thought Reasoning (ECoT), and helps robots break down complex tasks into smaller steps. We have mentioned similar methodologies to build reasoning models in software, and it is a welcome sign to see counterparts to address physical tasks. This sort of reasoning architecture appears to be promising for robotics tasks as well, because the machines that used the ECoT system were able to outperform robots using a different reasoning system by 28% across a variety of tasks. arXiv. 

• Self-Driving Cars: The Wall Street Journal reports Shanghai, China is opening a 205 km (127 mile) portion of its city for residents to book free rides in robotaxis from four different self-driving car companies. This news follows Waymo’s recent announcement that San Francisco, California will now allow the company to hail rides for anyone in the city.

• AI Mathematical Olympiad Competition: The Artificial Intelligence Mathematical Olympiad (AIMO) competition concluded its recent challenge for AI systems to solve intermediate-level high school math challenges. The winning entry, NuminaMath, scored an impressive 29 out of 50 on both public and private test sets, which far surpasses the baseline 3 out of 50 score from small-sized Language Models (LLMs) like Gemma 7B. Kaggle. Hugging Face.

Other Observations 📰

A Rare Total Larynx Transplant Helps Patient Speak Again

A team of doctors at Mayo Clinic in Arizona performed a rare total larynx transplant on a patient with active cancer, marking a medical milestone as the first known such procedure in the United States. The surgery was part of a clinical trial, and the patient's recovery and progress has exceeded the research team’s expectations thus far. The transplant involved replacing the patient’s larynx, pharynx, upper trachea, upper esophagus, thyroid and parathyroid glands, blood vessels, and nerves. With so many components for surgeons to replace, it makes sense why the procedure is so uncommon.

The patient lost the ability to speak, swallow, and breathe on their own due to a rare form of laryngeal cancer. After the transplant, the patient can now speak again with a voice quality estimated to be around 60% of normal. They can also swallow small amounts of liquids and solids with minimal difficulty or discomfort. Once they can fully breathe on their own, the doctors plan to remove their breathing tube, which is expected to happen in the coming months based on the patient’s current recovery progress.

These sorts of stories are inspiring for us to read. It is remarkable to learn about somebody who lost something so essential to human life most people cannot imagine what it is like to live without it. Through the wonders of modern medicine, that same individual has mostly regained their sought-after independence. We know the goals of this clinical trial is to research the safety and efficacy of laryngeal transplants and the field is likely many years away before it becomes a standard option for patients, but it is fantastic to witness important advances in this domain, because there are thousands of people worldwide who are in need of similar support. Mayo Clinic. Mayo Clinic Proceedings Journal.

Our Full Science Index

• Renewable Energy Transition: Bloomberg reports battery prices in China are down 51% from last year, thanks in large part to a drop in raw material costs and an overcapacity of batteries for their region after such aggressive clean energy measures over the past 2 years. The lower prices are expected to fuel more clean energy demand, and allow a virtuous cycle to cascade. In other news, Europe reported more than 50% of its energy in the first 6 months of 2024 came from renewable sources. PV Tech.

• Modifiable Cancer Risks: Researchers at the American Cancer Society conducted a study to estimate the proportion of cancer cases and deaths attributable to modifiable risk factors in adults 30 years old and older in the United States. The study found 40% of cancer cases and 50% of cancer deaths could be attributed to modifiable risk factors. Cigarette smoking was the leading risk factor, contributing to nearly 20% of all cancer cases and 30% of all cancer deaths. Excess body weight, alcohol consumption, UV radiation exposure, and physical inactivity were some of the other major modifiable risk factors noted by the researchers. American Cancer Society. ACS Journal.

• Dark Matter: A team of astronomers used NASA's Hubble Space Telescope to study the movements of stars within the Draco dwarf galaxy, a system located roughly 250,000 light-years from Earth. The team analyzed 18 years of observational data to create a 3D representation of the galaxy's cusp-like structure. They hope their research can provide new insights into the distribution of dark matter within galaxies and how it influences a galaxy’s structure. NASA. The Astrophysical Journal.

Media of the Week 📸 

A Robot Designed With A Vision & Speech Integration

The team of Deep Robotics shared their latest demo of their robot with a multimodal AI model integration. Looks like Chat GPT-4o is coming to a machine near you.

JWST Captures A Penguin & Egg Galaxies

On the 2^nd anniversary of NASA's James Webb Space Telescope, a team of astronomers observed the interacting galaxies which are collectively known as Arp 142. The silhouette of Arp 142 resemble a penguin and an egg, and is located approximately 326 million light-years from Earth. NASA.

Hubble Telescope Finds Evidence Of Black Hole In Omega Centauri

A team of astronomers used over 500 images from NASA's Hubble Space Telescope to has found evidence for the presence of a black hole 17,700 light years away from Earth, in the globular cluster Omega Centauri. The team measured the velocities of 1.4 million stars in the cluster and found 7 stars are moving at escape velocities, which indicates a massive object is pulling on them through its gravity. This discovery provides the most direct evidence for a black hole in Omega Centauri, and it suggests a black hole is closer to Earth than the supermassive black hole at the center of the Milky Way. NASA. Nature.

This Week In The Cosmos 🪐

July 21: a full moon.