Inverse Vaccines, AI diagnosing Disease and Chatbots helping develop Autonomous Vehicles

September 21

This week we look at a new type of vaccine, an Inverse Vaccine. Standard vaccines train the body to recognize unwanted intruders, this does the opposite. It causes the body to forget certain intruders. It may become a viable treatment for many autoimmune diseases. We also investigate an AI that diagnoses some diseases from retinal images. We examine a self driving car software model that has added a ChatGPT type language model so that testers can talk to the autonomous vehicle and find out why it is doing what it does autonomously. Finally we discover a new type of Freeze Ray that NASA hopes to use in space.

Inverse Vaccines

A team at the University of Chicago have developed a new type of vaccine that may be able to completely reverse autoimmune diseases like multiple sclerosis and type 1 diabetes.

There are some viruses that cause our immune system to attack healthy tissue rather than the disease that has infected us. In recent years it was discovered that 99% of Multiple Sclerosis patents had Epstein Barr virus antibodies. The virus triggers the immune system to attack the body’s myelin, the protective coating around nerve cells. It does this by mimicking proteins in our body. A variety of other diseases are also suspected of having viral origins. The viruses are able to stop the body from defending itself.

A standard vaccine works by teaching the immune system to recognize a virus or bacteria as harmful as soon as it enters the body. T-cells can then recognize unwanted cells and molecules as foreign and get rid of them. The Inverse Vaccine does the opposite. It removes the immune systems’ memory of one molecule. This immune memory erasure can stop the autoimmune reactions like those seen in MS and rheumatoid arthritis where the immune system has been tricked to attack healthy tissues.

Our liver is responsible for labeling some molecules from broken down cells with a “do not attack” flag. These cells will die by natural processes and the marker prevents autoimmune reactions (we don’t want unnecessary autoimmune reactions). The researchers coupled a molecule being attacked by the immune system with a sugar molecule, N-acetylgalactosamine (pGal). This sugar resembles a fragment of an aged cell that the liver would recognize as a friend rather than foe. The team was able to show that the vaccine could stop the autoimmune reactions associated with diseases such as MS.

The researchers were able to attach any molecule that they wanted to pGal and it will teach the immune system to tolerate it. The team was able to attach pGal to an MS like disease in animals and stop the disease from attacking the myelin, allowing the nerves to function correctly again and to reverse the symptoms of the disease.

Currently autoimmune diseases are treated with drugs that shut down the immune system. Whilst effective this approach has side effects. It is expected that the inverse vaccine will have fewer side effects. Initial preclinical safety trials have begun in people with celiac disease. Phase 1 safety trials are underway in people with MS. It will take time before Inverse Vaccines are approved for general use however this is important progress toward dealing with these diseases.

AI identifying disease via Retinal Images

A new tool called RETFound has been developed to diagnose and predict the risk of developing multiple health conditions from people/s retinal images. A range of conditions from Parkinson’s Disease to heart failure to ocular diseases can be identified.

This is not the first AI tool that can detect disease from retinal images however this tool was built using self supervised learning. The 1.6 million images were not analyzed prior to being shown to the AI. This analysis is very time consuming and thus expensive. The AI was trained in a similar method to the way Large Language Models (LLMs) were trained. LLMs use a myriad of text examples scraped from the web to predict the next word in a sentence from the context of the preceding words. RETFound uses a multitude of retinal photos to learn how to predict what damaged portions of images should look like.

After scanning the 1.6 million images the model learnt the many features of a retina and what a healthy retina should look like. This forms the cornerstone of the model (knows as a Foundation Model). The model can then be adapted for many tasks.

The retina offers a window into our health. It is the only part of the body where the capillary network can be observed directly. This allows to see some systemic cardiovascular disease such as high blood pressure. The retina is also an extension of the central nervous system. This allows us to evaluate neural tissue.

After the 1.6 million images had been viewed by the AI, a small number of labelled images were introduced. For example 100 images from people who had developed Parkinson’s Disease and 100 from people who had not. Having learnt what a retina should look like, the AI can quickly learn the retinal features associated with a disease.

The system performed well at detecting ocular diseases (a range of 0.822 to 0.943 correlation depending upon the data set used). It was not as good at predicting the risk of systemic disease such as heart attacks, heart failure and Parkinson’s however it was superior to other AI models trained to identify these conditions.

The authors of the model have made it publicly available for others to use and adapt. The hope being that training on further patient populations will improve the predictive outcomes.

Using Chatbots to train Autonomous Vehicles

Many of you will have tried chatGPT or one of the many other large language models that have recently come onto the market. These models are considered to be learning models. As we use the models they learn from the reactions to our output and improve. This recent example shows how a user asked chatGPT a simple math question.

chatGPT initially gave a wrong answer however it showed its’ working and convinced itself that the original answer given by Morgan was correct. Morgan asked the same question again.

chatGPT has learnt from its’ mistake and can now provide the correct answer. A similar approach is being trialed by UK Autonomous Vehicle startup Wayve. They can interrogate their vehicles and get answers back about the vehicles driving decisions.

The company added a large language model to their existing self driving software. LINGO-1 synchs video data and driving data with natural language descriptions that capture what the car sees and what it does.

In a demo the CEO asked the car “what’s the weather like”, the car responded “the weather is cloudy”. “What hazards do you see?”, “there is a school on the left”. “Why did you stop?”, “Because the traffic light is red”. By quizzing the software every step of the way the company hopes to understand why the car is making various decisions. Most of the time autonomous vehicles drive without incident, it is the edge cases that cause problems. This system may help Wayve discover the underlying cause of issues more quickly.

Additionally Wayve is having their trained instructors drive cars and talk out loud as to why they took every action. This input to training models may help train cars more quickly.

Wayve have made a few technological breakthroughs in recent years. In 2021 an AI trained on the streets of London was used to drive cars in four other cities in the UK. That would typically take a lot of reengineering and new training. In 2022 they were able to use an AI trained on one car type to drive a different car. Another industry first and one that will be vital if self driving cars are to become widely adopted.

Google is using natural language to train robots in domestic and industrial settings. In the future the way we interact with intelligent machines will be through language. This is likely an important first step in that direction.

Freeze Ray

A team at the University of Virginia are developing a plasma freeze ray that will allow engineers to cool electronics in the vacuum of space. I know this will disappoint the aspiring super villains amongst you, but it is not the Freeze Ray from comic books.

Plasmas are ionized gasses that can reach temperatures many times that of the sun. Using a plasma to freeze something doesn't make intuitive sense however plasmas have some surprising properties. Despite the high temperatures, when a plasma is first generated it can interact with other materials to produce a cooling effect.

The energy flux of a pulsed plasma interacts with the surface of a target physically, chemically and electromagnetically to generate an effect that evaporates water and carbon dioxide molecules that have been absorbed by the target’s surface. This rapidly cools the surface by tens of degrees. Pulsing the plasma keeps it from counteracting the cooling effect.

The US Air Force and Space Force are interested in the technology due to the difficulties with cooling electronics in space and at very high altitudes. There is no air in space to circulate around electronics to carry away any heat. Currently electronics are set on metal cooling plates which then conduct the heat to radiators.

The goal is a robot arm with sensors that can zap electronics any time they need cooling. The current prototype uses helium. The next step is to make the equipment lighter and more compact whilst investigating any other gases that might be even more effective.

Paying it Forward

If you have a start-up or know of a start-up that has a product ready for market please let me know. I would be happy to have a look and feature the startup in this newsletter. Also if any startups need introductions please get in touch and I will help where I can.

If you have any questions or comments please comment below.

I would also appreciate it if you could forward this newsletter to anyone that you think might be interested.

Till next week.