#20 | Gene Networks In The Brain

Quote of the Week 💬 

A Map Of Networks Regulating Gene Function In The Human Brain

⌛ The Seven Second Summary: A consortium of researchers from PsychENCODE published 15 papers mapping gene regulation networks in the brains of people with and without mental disorders.

🔬 How It Was Done:

• The researchers received brain tissue from over 2,500 donors in order to create intricate maps illustrating how genes interact and control different biological processes within the brain.

• The research’s scope mostly covered the outer layer and deeper regions of the brain essential for functions like decision-making, memory, learning and emotions.

• To validate the significance of different DNA segments in the brain, the researchers performed a variety of experiments to compare the DNA sequences, gene expression and regulatory patterns of people with and without specific traits or disorders, such as schizophrenia, PTSD, and depression.

🧮 Key Results:

• The United States’ National Institute of Health claims the results from this project are the “largest and most advanced multidimensional maps of gene regulation networks in the brains of people.”

• Since there is so much data to parse, analyze and explore, the research consortium developed an interactive web portal to help people visualize how various types of brain cells behave differently in people with and without mental disorders.

💡 Why This May Matter: As researchers gain a deeper understanding of how genetics influence the brain and contribute to mental illnesses, they become more adept at identifying molecular targets for new therapeutics to treat debilitating diseases.

🔎 Elements To Consider: There needs to be far more than 2,500 donors to develop a representative sample of how human brains are influenced by genes. The fact that such a small sample led to one of the largest and most comprehensive datasets in a research domain underscores just how much more work remains in the field.

📚 Learn More: National Institutes of Health. Science. Psych Screen.

Stat of the Week 📊 

The Success Rate Of AI-Discovered Drugs In Clinical Trials

≥ 80%

⌛ The Seven Second Summary: Analysts from Boston Consulting Group assessed if new drugs discovered by and designed through AI-based techniques are resulting in higher early-stage clinical success rates, and the preliminary findings are promising.

🔬 How It Was Done:

• The analysts identified 100+ pharmaceutical and biotechnology companies with specialties in AI and gathered data about their clinical trials.

• They categorized AI-based therapeutic solutions based on some criteria and recurring themes:

  • Drugs where AI helped find the specific biological targets in the body to interact with.

  • Small molecules, proteins, antibodies, vaccines and other substances identified, developed or optimized through the use of AI techniques.

  • Existing drugs where AI has found new uses for.

• Through this criteria, the analysts found 75 therapy treatments enter clinical trials between 2015 - 2023.

🧮 Key Results:

• As of 2023 December, the analysts found 24 AI-based therapeutic solutions that had completed Phase I trials.

  • 21 of those Phase I trials were successful, which represents an implied success rate of 87%. For reference, the industry average for Phase I trials is estimated from as low as ~40% to as high as 75%.

• Furthermore, the analysts found 10 AI-based therapeutic solutions that had completed Phase II trials.

  • 4 of those Phase II trials were successful, which represents an implied success rate of 40%. For reference, the industry average for Phase II trials is estimated from as low as ~30% to as high as 59%.

💡 Why This May Matter: More research labs around the world are incorporating AI-based solutions into their workflows, and it is worth understanding the results of this paradigm shift for patient outcomes in a clinical setting. It remains far too early to claim AI algorithms have solved molecular design problems for researchers. However, as we mentioned last week, solutions such as AlphaFold are reducing the time and costs involved in the design process, and they may decrease the severity of off-target effects from new drugs as technologies become more robust and precise.

🔎 Elements To Consider: Among the 3 AI-based therapeutic solutions that failed in Phase I trials, only 1 fell short of evaluation criteria; the other 2 were halted due to business decisions to re-prioritize their drug pipeline. Similarly, of the 6 candidates discontinued after Phase II trials, only 2 stopped because of negative outcome data; the remaining 4 stopped due to shifts in business priorities, clinical operations challenges, or other reasons. The success rate of AI-based therapeutic solutions may change dramatically as more treatments enter clinical trials over the years, and the explanations behind these initial failures is a reminder of the complex landscape organizations need to navigate in order to bring new treatments to market.

📚 Learn More: Science Direct.

AI x Science 🤖

Anthropic Researchers Map The Mind Of A Large Language Model

When we discussed Open AI’s Chat-GPT4o demos last week, we mentioned the impressive anthropomorphic qualities of their new system. As AI models become more charismatic, persuasive, and personable in their text and speech patterns, they will appear more convincing and trustworthy, even if their responses are inaccurate or misleading. Thus, the ability to understand and interpret the reasoning behind AI systems becomes more important as the models become more advanced and sophisticated. This is why our newsletter focuses so much attention on research about model reasoning.

Today’s newsletter issue is no different. Anthropic recently released a paper expanding upon their previous efforts to identify patterns about how Large Language Models (LLMs) process language.

When an LLM receives language inputs, different strings of numbers are activated in response to specific words, phrases, or concepts. To understand how an LLM internally processes and represents language, algorithms can be developed to identify and analyze the underlying patterns that cause certain strings of numbers to activate while others remain inactive.

Over time, a trend emerges, and researchers can translate the model’s patterns of activated numerical strings into a map to reveal how the LLM represents language concepts such as objects, emotions, and relationships, as well as more abstract ideas like justice, freedom, and love. The model’s conceptual representation of language is referred to as features and appears to share some similarities with how humans group and consider concepts. For example, the model’s feature about “friendship" might be closely connected to other features about "support", "trust", and "loyalty".

By analyzing millions of features, Anthropic can identify which ones might be related to harmful or unsafe behavior. For instance, they might find features connected to discrimination, prejudice, and bigotry, or features the model activates in association with deception, manipulation, and misinformation. Understanding which features are involved in harmful behavior enables Anthropic to develop strategies to detect and mitigate for dangers, thereby making their LLMs safer and more reliable to interact with.

In their paper, Anthropic also mentioned the ability to fine-tune features and connections between features to become more or less pronounced in their largest models. This is just one of many avenues they can now utilize to remove, reduce or counterbalance undesirable features and associations the model may otherwise learn, such as intentionally lying or harboring biases towards a group of people.

Anthropic notes that applying this interpretability technique to their largest models across an entire language corpus costs far more than the costs to train the same models. This means there is still a lot of important work remaining to make compute cheaper so people can holistically understand how AI models generalize information and behave in different contexts. Nonetheless, this is a great step forward for AI research. Anthropic. Transformer Circuits Publication.

Our Full AI Index

• Gemini’s Technical Report: While we know few details about Open AI's Chat GPT-4o's model architecture, the recent technical report of Google's Gemini 1.5 model discloses a mixture-of-experts design, where multiple specialized models are combined to create a sophisticated AI system. Their report also mentions a special component called Gemini 1.5 Flash to help their system process and understand natural language at faster speeds. Between the news about Gemini 1.5 Flash and Google’s focus on improving long context-lengths in their AI systems, it appears the company believes memory and speed / low-latency are two major keys to make the conversational assistant experience they demoed in Project Astra a reality. Google Deepmind.

• Machine Learning & 3D Printing: Researchers at Boston University developed a robot named MAMA BEAR that uses machine learning and 3D printing to create and test structures for energy absorption efficiency. After conducting over 25,000 experiments, MAMA BEAR was able to design a shape with 75% energy absorption efficiency, which is reportedly an optimization record. Boston University. Nature.

• Digital Twins & Heart Surgery: The Wall Street Journal shared the story of researchers from John Hopkins University who are developing digital twins of patients' bodies and organs to predict how they will respond to treatments and procedures. While the existing technology cannot capture all the minute details and intricate complexity in the human body today, its approximations of actual patient scenarios will become increasingly useful for testing drugs, predicting disease progression, and designing prosthetics, so that people can benefit from highly personalized and optimized care. Wall Street Journal Archive.

• The European Union’s AI Act: The European Council has officially approved its groundbreaking Artificial Intelligence Act into law. The law follows a risk-based approach to categorize and regulate AI systems differently based on the potential harm they may cause to society. The law also generally prohibits certain technology practices, such as behavioral manipulation tactics and predictive policing to profile people by using their biometric data. European Council.

• Microsoft’s Flurry Of Developments: Last, but not least, Microsoft shared a ton of news and product updates at their Build Conference. Some of our favorites were the following:

  • Atmospheric Forecasting: A new AI model called Aurora that uses vast amounts of weather and climate data to make 5-day global air pollution predictions and 10-day weather forecasts. arXiv.

  • Biomedical Image Parsing: A new AI system called BiomedParse that analyzes biomedical images. In testing, BiomedParse outperformed other systems on a massive dataset of 102,855 different images, at a number of tasks, including detecting and recognizing objects from text prompts. arXiv. Github.

  • New Small Language Models: Microsoft expanded upon their Phi-3 family of small language models with the addition of Phi-3-vision, a multimodal model that combines language and vision capabilities similar to Chat GPT-4o. Azure. arXiv.

Other Observations 📰

Majority Of Patients With Crohn's Disease Achieve Remission

Eli Lilly recently shared positive results from their Phase III VIVID-1 study, which evaluated the efficacy and safety of mirikizumab in patients with moderate to severe Crohn's disease. Mirikizumab works by binding to a specific protein in the digestive system, which prevents the body from releasing certain chemicals that can cause inflammation.

The randomized, double-blind study found patients treated with mirikizumab achieved statistically significant and clinically meaningful improvements after one year compared to their placebo group. Specifically, 54% of patients on mirikizumab achieved clinical remission, and 48% experienced meaningful healing in their digestive tract, compared to just 27% and 28% in the placebo group, respectively.

Furthermore, after one year of treatment, 39% of patients new to biologic drugs for Chron’s disease and 37% of patients who had not responded to previous biologic treatments achieved meaningful clinical responses with mirikizumab. This compares to just 11.8% and 6.2% of the placebo group. It is uncommon to see Phase III trials outperform placebo groups by such a significant margin, which may serve as yet another indicator that targeted therapy treatments are becoming more effective and precise than in the past. Lilly. Clinical Trials.

Our Full Science Index

• Kidney Cancer Survival: The immunotherapy drug pembrolizumab has been shown to improve overall survival in patients with early-stage kidney cancer, reducing the risk of death by nearly 40% compared to its placebo group. This milestone marks the first time a post-surgical treatment has been proven to extend lives in kidney cancer patients. Cancer. NEJM.

• Replication Tracker: The Framework for Open and Reproducible Research Training (FORTT) has launched the Replication Database to track medical and behavioral studies that have undergone replication attempts. To date, their database includes 2,173 replication attempts spanning 656 original studies, though only 54% of the replication attempts thus far have been able to reproduce its original findings. Replication Database.

• Material Design & Computer Simulations: Researchers from the University of Liverpool and the University of Southampton are using computational design methods to construct a new type of material called non-metal organic porous. The material is made from salts, and has demonstrated the ability to capture pollutants, such as iodine, as well as its metal-organic alternatives. Nature.

• The Features Of Traditional Music: A study involving 75 researchers from many different cultural backgrounds analyzed 300+ audio recordings from traditional songs around the world and discovered they share similar features. Some of the shared characteristics they identified were slower music tempos, and higher pitch vocals. The team is now wondering if certain evolutionary benefits are responsible for the universal musical attributes found across distinctive cultures. Science.

Media of the Week 📸 

World Record For Fastest Robot To Solve A Puzzle Cube

Mitsubishi Electric’s machine only took 0.305 seconds to complete the 3×3×3 Rubik’s Cube — quite literally as fast as a blink of an eye. For reference, the human record for solving a Rubik’s Cube is 3.13 seconds, set by United States champion Max Park in 2023.

European Space Agency’s Euclid Mission Captures Its First Data

The European Space Agency’s Euclid space mission has released 5 new images of the universe in stunning detail, including a galaxy cluster, star-forming regions, spiral galaxies, and galaxy groups. The mission is less than 1 year old, and their research findings have already led to 10 scientific papers, in addition to these beautiful photos. ESA. Cosmos.

Comparing Visual Similarities At Macro & Micro Scales

Kim Arcand, a visualization scientist for NASA’s Chandra X-ray Observatory, recently took on a project to illustrate some of the common patterns often seen in the universe. The gallery she curated covers 9 different image pairings, and it’s fascinating to see the similarities of design in its largest and smallest dimensions. Chandra.

This Week In The Cosmos 🪐

No major events this upcoming week. How disappointing…