#19 | AlphaFold 3 Expands Its Model

Quote of the Week 💬 

AlphaFold 3 Can Now Predict DNA, RNA & Other Biomolecules

⌛ The Seven Second Summary: The AlphaFold team at Google built upon their protein structure foundation by predicting 3D structures of DNA, RNA, and other small molecules. Now that they have a representation of nearly every molecule known to science, they can predict how these different molecules interact with one another.

🔬 How It Was Done:

• The previous version of AlphaFold used existing protein structures and amino acid sequences to train a predictive model that can forecast the 3D structure of over 200 million proteins — nearly every protein known to science.

• The full details of the team’s latest methods was not released, but one notable change they did mention in their press release was having their model use diffusion networks to create its 3D structure. Diffusion networks are the architecture underlying most of the AI-generated art seen today, and its use allows an AI system to refine pixel coordinates through a series of iterative steps until it generates a realistic image.

• In the case of AlphaFold 3, their diffusion network receives 3D representations of amino acid sequences, chemical bonds, or other molecular details, and refines the spatial arrangements of these molecules until it produces a final structure of how they all might look like when in close proximity.

🧮 Key Results:

• The AlphaFold team claims that their system is at least 50% more accurate than existing methods at predicting the interactions of proteins with other types of molecule types.

• In some cases, the team says their system doubles the prediction accuracy.

• In fact, Dr. David Baker, a computational biophysicist at the University of Washington, leads a team developing RoseTTAFold All-Atom, one of the most prominent alternative systems to AlphaFold 3, and he told Nature that Deepmind’s new system is better than what his team has developed.

💡 Why This May Matter: AlphaFold 3’s expanded capabilities presents many new use cases for medicine, material science, and crops and agriculture. In the case of drug discovery, traditional development often fails due to unforeseen off-target effects that may be harmful. With this type of software advancement, researchers can now design, simulate and predict how a new drug might interact with various molecules beyond its designated target, and choose options with the least significant off-target effects.

🔎 Elements To Consider: Unlike with previous AlphaFold releases, Google will not open-source its latest work, and the web-based solution it is providing researchers cannot be used for potential drug interactions. This foreshadows Google’s intent to have its subsidiary company, Isomorphic Labs, commercialize the technology through licensing or partnerships with pharmaceutical companies.

📚 Learn More: Google. Nature. AlphaFold Server.

Stat of the Week 📊 

A Toddler Born Deaf Can Now Hear After Gene Therapy Trial

16 minutes

⌛ The Seven Second Summary: A baby girl born deaf in Oxfordshire, England now has “close to normal hearing levels for soft sounds“ after receiving a gene therapy treatment.

🔬 How It Was Done:

• The baby was born completely deaf because of a rare genetic condition known as auditory neuropathy. This condition affects auditory nerve cells and prevents sound information from transmitting to the brain effectively.

• Auditory neuropathy can be caused by a variation in a single gene, known as OTOF. This gene produces a protein called otoferlin, which helps hair cells in the inner ear communicate information with the brain’s auditory nerves.

• The 11 month old underwent surgery to receive a working copy of the OTOF gene in her right ear. Concurrently, a hearing device was fitted for her left ear to provide additional support for her newfound hearing development.

🧮 Key Results:

• The baby was able to respond to sound within 4 weeks of the gene therapy treatment. After 24 weeks, the research team confirmed the baby’s right ear had “close to normal” hearing levels for whispers and other soft sounds.

• It only took the surgeons 16 minutes to complete the operation.

• This was the first British patient in the world and the youngest child to receive this type of gene therapy treatment.

💡 Why This May Matter: Earlier this year, we highlighted a separate study utilizing a similar gene therapy treatment to treat a group of children diagnosed with a rare case of hereditary deafness. These experimental trials in hearing and vision are showcasing promising results thus far, which should bolster the funding and development of gene therapies for more common types of sensory impairments.

🔎 Elements To Consider: This child is participating in the Chord clinical trial. It consists of three parts, with three deaf children receiving a low dose of gene therapy in one ear only. A different set of three children will then receive a high dose of treatment on one ear. Then, if the high dose treatment is shown to be safe, 12 children will receive a dose in both ears at the same time.

📚 Learn More: University of Cambridge. The Guardian.

AI x Science 🤖

OpenAI Releases GPT-4o: Text, Vision, Audio In a Single Model

Last week, Open AI announced a significant upgrade to their AI model, GPT-4. Their new model, GPT-4o, synthesizes and integrates AI functionality across many modalities of data, including audio, video, image and text. This upgrade presents a model more capable of interacting with and understanding the world.

Soon after Llama 3’s release, we mentioned data quantity is one of the major attributes that yield logarithmic performance improvements in an AI system as it continues to scale. When a model increases the surface area of how it can learn, from text-only to a landscape of text, audio, video and image data, the result should eventually be a more competent system with an expanded toolkit to advance its skill progression.

We may already be observing this trend with GPT-4o. While Open AI only mentioned marginal improvements in various reasoning tasks, the aggregate performance jump has the model sitting comfortably atop the LMSYS Chatbot Arena Leaderboard. Furthermore, this model is 2x as fast and 50% cheaper than its GPT-4 Turbo predecessor, which allows a conversation with the system to have cadence similar to an interaction between two people.

Speaking of which, Open AI tuned the voice of their system to have the most anthropomorphic qualities we have yet to see from alternative AI systems. The end result are some impressive demos, including an example of GPT-4o as a math tutor.

There an exciting number of potential use cases for multi-modal AIs, and these demos are just the beginning. While GPT-4o’s reasoning abilities is comparable to other AI systems, this may change as people and businesses use the product in interesting ways for it to learn from additional types of data. Open AI.

Our Full AI Index

• Robotics Simulation: We often share news about advancements in reinforcement learning since it is a powerful technique to compound AI capabilities. Recently, researchers from the University of Pennsylvania shared an algorithm to help robots learn tasks more efficiently by automating the process to define what constitutes success and adjust its training environment. This reduces the need for human involvement and makes it more viable for robots to reliably simulate tasks, learn from its virtual settings, and apply its acquired knowledge and skills in the real world. Github.

• Hospital Simulation: To help Large Language Models (LLMs) learn about best practices to follow in a hospital setting, researchers from Tsinghua University developed a virtual hospital environment called Agent Hospital. Here, autonomous LLM agents played the roles of patients, nurses, and doctors, and learned how to effectively treat illnesses and care for patients. arXiv.

• Spatial Reasoning: Microsoft researchers shared a new technique to improve how LLMs reason through tasks involving spatial awareness. The process is similar to chain-of-thought reasoning, where the LLM is prompted to break down a reasoning problem into steps. arXiv.

Other Observations 📰

Crops Engineered To Inherit 100% Of Genetic Traits From Parents

Plant breeding follows a scientific method to methodically and precisely change the traits of specific crops in order to produce desirable characteristics in future generations of offspring. Higher yield, improved disease resistance, increased drought tolerance, and better taste, are some of the attributes these agricultural scientists attempt to optimize.

One of the challenges with crop breeding is ensuring desirable genes are inherited in a parent’s offspring, since typical plant reproduction results in a random selection of 50% of genes from both parents. To solve this problem, a research team from Max Planck Institute designed a system to change how crops reproduce so an offspring inherits 100% of its genes from both parents.

They accomplished this by genetically modifying some reproductive genes in parent tomatoes. When these crops reproduced, their cell division followed mitosis instead of meiosis to avoid genetic recombination and segregation, resulting in offspring with sex cells that are exact clones of the parent plant.

Interestingly, Ohalo, a private company, had its patents published recently, and a presentation by their CEO mentions using specific proteins to induce a similar reproductive outcome for plant breeding. When Ohalo conducted experiments to optimize for higher crop yield, they reportedly produced offspring with 50%+ more yield than their control groups. If comparative gains can be made in other crops and for different attributes without sacrificing the nutritional content of the crops, then agriculture may be able to achieve unprecedented levels of crop resiliency and yield in the future. 

It’s easy to imagine this use case for other plants as well, such as scientists who want to optimize trees for carbon sequestration, wood properties and other reforestation purposes for extreme weather. What a time to be alive. Max Planck Society. Nature.

Our Full Science Index

• Falling Emissions: Coal and gas provided less than a quarter of the European Union’s energy in April 2024, down to a record low of 23%. This marks a 24% year-over-year improvement for the region, which keeps it on its path to fully eliminate fossil fuels by 2035. Ember.

• Public Health: The World Health Organization has certified Belize, Jamaica, and Saint Vincent and the Grenadines for eliminating the transmission of HIV and syphilis from mothers to their children. A huge win for child mortality. UNICEF.

• Spinal Cord Implant: Researchers from the University of Cambridge are developing an implantable device to treat spinal cord injuries. Their approach seems complementary with a different research group using spinal cord implants to help Parkinson’s patients walk. University of Cambridge. Science.

• Beethoven’s Health: A consortium of researchers shared research results of an analysis Beethoven’s hair. They discovered his blood lead levels were several times higher than an adult in modern times. Clinical Chemistry. NYT.

Media of the Week 📸 

Project Astra Has An AI System Tour The Office

Google held its major developer conference last week, and the Big Tech company announced ~100 new product and app updates. One of the announcements that caught our attention was their Project Astra demo, a multi-modal model similar to Chat GPT-4o. Project Astra did not release any products alongside its announcement, so it’s difficult to know how far Google actually is from making this vision a reality. Nonetheless, it’s impressive to imagine. Google Deepmind.

A Surgical Operation On A Kettle Of Corn

Microsurgery is one of the many captivating ways to showcase technological advancements and surgical precision. Watch this expert operate on a kettle of corn! Sony.

A Cubic Millimeter Of The Human Brain Mapped In Detail

Researchers from Harvard, Google Research and elsewhere recently completed a 10 year project to construct a 3D map of 1 cubic millimeter of the human brain — roughly one-millionth of a whole human brain. Their map contains roughly 57,000 neurons and 150 million synapses, with each neuron colored according to its size. Science. YouTube.

The Dawn Of A Star In The Cosmos

NASA’s Hubble telescope delivers again with its latest photos of a young star before it blossoms into a mature form, like our galaxy’s Sun. Does this photo remind anyone else of the sky in Van Gogh’s Starry Night? NASA.

The Strongest Solar Storm In 20 Years Swept The Planet

Last but not least, the world experienced the strongest solar storms in over 20 years, which created beautiful auroras at much lower latitudes than normal. This was our favorite photo from the spectacle. We hope you were able to enjoy the glistening sky alongside us! ESA.

This Week In The Cosmos 🪐

May 23: A full moon.