> “I spent a long time on HIV vaccine development,” says Koff, adjunct professor of epidemiology at the Harvard T.H. Chan School of Public Health. “None of them worked
Really? I think I've heard about some successes in this area. Clearly there are contradicting statements, so how can we reconcile them? Remember taking statements like that with the grain of salt.
Sure you have. That's the way the corporate technology hype machine works. So many amazing things are perpetually just around the corner: cold fusion, cancer vaccines,cheap powerful batteries. Invest now!
Advances in energy density, reliability, and safety have revolutionized batteries over the last 30 years.
There are entire market segments where you are only as competitive as your battery and participants scramble to integrate the latest advancements but because you can't go to the store and buy a pack of AA batteries with the tech in them people think it doesn't exist.
The volumetric energy density of lithium-ion increased 8x between 2008-2020, doubling every four years.
Solid state batteries look to be taking up the improvement race. 450 watt-hours/liter was VERY good in 2020. A cheap YANGDA Thunder Stone solid-state drone battery purchased from a sketchy retailer in 2023 has an energy density of 537Wh/L, and you can buy a 18650 battery with 670Wh/L on amazon for $10. Specialty batteries 800-1000Wh/L are available to commercial clients, so the energy density has doubled once again in less than four years.
"To study this, Koff and his collaborators around the world are using existing samples, stored in biobanks, and sequencing vast numbers of crucial lymphocytes, known as B and T cells, each of which contains a unique receptor to recognise parts of viruses, bacteria, or tell-tale signs of cancer cells. Koff’s aim is to create an atlas of these receptors, and feed that information into an AI model to predict what the complete repertoire of B and T cells might look like in young, otherwise healthy individuals, what changes as we age, and what we might be able to do to modulate it."
This seems worthwhile to me but I'm skeptical of an AI's ability to accurately predict biology completely. This is (very loosely) sort of like an AlphaFold for antibodies however, we already know that AlphaFold does not predict unique protein domains well. Maybe this will improve with time - and more protein structures, but fundamentally I don't see how AI could predict the existence of something novel... Similarly, I don't see how this project could accurately map all of the possibilities for B/T cells.
I hope this leads to a renaissance in immune disease treatment.
Immune related diseases are absolutely hellish, if your immune system one day decides to make your life hard/painful through one of the million possible attack vectors, we basically have little treatment besides taking steroids to blunt your whole immune system.
Coming up with therapies that can shape your immune system would be an absolute game changer.
Renaissance? Immune therapy is the most active area for drug development in the last two decades, comparable to sequencing for diagnostics (except much more productive).
There are dozens of drugs suppressing and amplifying different aspects of the immune system. Multiple Sclerosis is one of the three most expensive disease in the world, and its drugs have multiple mechanisms of action. There are even dozens of cancer-specific immune therapies.
I have a degree in electrical engineering and a degree in medical microbiology and immunology.
It was one of my (many) dreams to model the immune system in software, using control systems to create a reference model that you could challenge and get legitimate responses (like climate models today). This was 30 years ago.
One of the challenges is that most immunologists have never even heard of the concept of "control systems" and likely have no concept of the math involved.
1) The immune system (and really the entire body) is a giant control system with many feedback loops that arent understood at all. I for example was studying something very simplistic which was the IL-2 vs Il-4 pathway in response to tuberculosis infection. I havent kept up on the research so I have no idea what state of the art is in immunology 30 years later.
2) The immune system is a giant random number generator. Your body randomly generates antibody patterns in the womb, so everyone is different. The system is analog instead of digital creating a tremendous amount of variation. It is unlikely you could collect data sets large enough to start training an AI to model the immune system except maybe at the grossest level.
It sounds to me that AI is being used as a buzzword to get more attention.
This is a summary article aggregating papers that are attempting to model biological systems as control systems
> The immune system is a giant random number generator. Your body randomly generates antibody patterns in the womb, so everyone is different.
I read about this in the book "Immune" [0], and it blew my mind. I just want to call attention to this part of your comment because people might breeze past it.
Your immune system is tasked with what might seem an impossible problem: be able to kill anything that doesn't match "you" / your genome. But it can't encode everything that isn't "you" in the DNA that your immune system came from, partly because DNA transcription is generative: any arbitrarily-long sequence of codons creates a protein.
So rather than try to write down all the things that could kill you (even ones that have never been encountered in all of human history) and remember how to kill those things, it generates random proteins, checks whether they look like you, and if they don't, figures out how to bind to those and produces proteins (antibodies) that bind to those things.
(I'm heavily paraphrasing and I don't have the book handy to refer to, sorry)
I'm an electrical engineer, I don't know much about biology.
Electrical is well-suited to modeling with ordinary differential equations. By making design decisions such that linear time-invariance is a valid assumption, we unlock really simple, powerful, and broadly applicable techniques for modeling, measurement, and control design.
Most biology isn't designed, which is probably why those papers had to jump right to much more advanced and specific techniques for modeling and control. With no useful on-ramp for most biologists, modeling and control will have to be a deliberate collaboration between two specialized disciplines.
It does not merely include the legions of lymphocytes, neutrophils, and monocytes/macrophages, which biologists have pieced together and studied over many decades, but all of the interactions these cells have, over a person’s life, with pathogens, toxins, and the consequences of their diet and lifestyle.
I wish them well, but will note my pet peeve about the so-called immune system:
The antigens we produce are not the entirety of the immune system. It's like saying the alarm is all that's involved in securing a building and hand waving off the walls, doors, locks, etc.
There are myriad ways to trick the white blood cells and the body also has ways of killing stuff other than the white blood cells.
Perhaps it will pan out, but it seems more like a private intellectual-property play than a real effort. It will likely slow progress by wrapping up information and access.
Methodologically, no one solves a hugely complex system by looking at the whole thing.
We've made tremendous progress when investigators focus on one interaction and do the bench work to isolate and prove features of it.
It's true that researchers need interaction data banks, e.g., to identify all the components in a metabolic cycle, to see what might be affected when part of that is amplified or suppressed.
Epidemiology is notoriously unhelpful in immune research because there are not enough numbers/powers to overcome individual variability.
Researchers largely work on models that operate immediately. There are no in-vitro models for immune learning, and few even for the relevant gene regulation.
People have tried and failed to build a control-system model for even the simplest interactions of a single cell.
Synthetic biology companies have burned billions with few results. I wouldn't expect someone with 40 years in HIV research to have the technical or leadership skills to pivot to the immune system.
The gain-of-function virology researchers who very likely played a leading role in the creation of Sars-CoV2 and its escape from the Wuhan Institute of Virology will see this as a map of vulnerabilities that can be exploited, i.e. a complete map of the attack surface for the human operating system.
Sure, there will also be many medical benefits from gaining this knowledge, but at the same time, there needs to be a global treaty-enforced ban on the deliberate biotechnological production of new infectious diseases by either private or public research labs.
17 comments
[ 2.9 ms ] story [ 51.9 ms ] threadReally? I think I've heard about some successes in this area. Clearly there are contradicting statements, so how can we reconcile them? Remember taking statements like that with the grain of salt.
Advances in energy density, reliability, and safety have revolutionized batteries over the last 30 years.
There are entire market segments where you are only as competitive as your battery and participants scramble to integrate the latest advancements but because you can't go to the store and buy a pack of AA batteries with the tech in them people think it doesn't exist.
The volumetric energy density of lithium-ion increased 8x between 2008-2020, doubling every four years.
Solid state batteries look to be taking up the improvement race. 450 watt-hours/liter was VERY good in 2020. A cheap YANGDA Thunder Stone solid-state drone battery purchased from a sketchy retailer in 2023 has an energy density of 537Wh/L, and you can buy a 18650 battery with 670Wh/L on amazon for $10. Specialty batteries 800-1000Wh/L are available to commercial clients, so the energy density has doubled once again in less than four years.
But nobody notices.
"To study this, Koff and his collaborators around the world are using existing samples, stored in biobanks, and sequencing vast numbers of crucial lymphocytes, known as B and T cells, each of which contains a unique receptor to recognise parts of viruses, bacteria, or tell-tale signs of cancer cells. Koff’s aim is to create an atlas of these receptors, and feed that information into an AI model to predict what the complete repertoire of B and T cells might look like in young, otherwise healthy individuals, what changes as we age, and what we might be able to do to modulate it."
This seems worthwhile to me but I'm skeptical of an AI's ability to accurately predict biology completely. This is (very loosely) sort of like an AlphaFold for antibodies however, we already know that AlphaFold does not predict unique protein domains well. Maybe this will improve with time - and more protein structures, but fundamentally I don't see how AI could predict the existence of something novel... Similarly, I don't see how this project could accurately map all of the possibilities for B/T cells.
Immune related diseases are absolutely hellish, if your immune system one day decides to make your life hard/painful through one of the million possible attack vectors, we basically have little treatment besides taking steroids to blunt your whole immune system.
Coming up with therapies that can shape your immune system would be an absolute game changer.
There are dozens of drugs suppressing and amplifying different aspects of the immune system. Multiple Sclerosis is one of the three most expensive disease in the world, and its drugs have multiple mechanisms of action. There are even dozens of cancer-specific immune therapies.
It was one of my (many) dreams to model the immune system in software, using control systems to create a reference model that you could challenge and get legitimate responses (like climate models today). This was 30 years ago.
One of the challenges is that most immunologists have never even heard of the concept of "control systems" and likely have no concept of the math involved.
1) The immune system (and really the entire body) is a giant control system with many feedback loops that arent understood at all. I for example was studying something very simplistic which was the IL-2 vs Il-4 pathway in response to tuberculosis infection. I havent kept up on the research so I have no idea what state of the art is in immunology 30 years later.
2) The immune system is a giant random number generator. Your body randomly generates antibody patterns in the womb, so everyone is different. The system is analog instead of digital creating a tremendous amount of variation. It is unlikely you could collect data sets large enough to start training an AI to model the immune system except maybe at the grossest level.
It sounds to me that AI is being used as a buzzword to get more attention.
This is a summary article aggregating papers that are attempting to model biological systems as control systems
https://www.frontiersin.org/articles/10.3389/fbioe.2021.6779...
I read about this in the book "Immune" [0], and it blew my mind. I just want to call attention to this part of your comment because people might breeze past it.
Your immune system is tasked with what might seem an impossible problem: be able to kill anything that doesn't match "you" / your genome. But it can't encode everything that isn't "you" in the DNA that your immune system came from, partly because DNA transcription is generative: any arbitrarily-long sequence of codons creates a protein.
So rather than try to write down all the things that could kill you (even ones that have never been encountered in all of human history) and remember how to kill those things, it generates random proteins, checks whether they look like you, and if they don't, figures out how to bind to those and produces proteins (antibodies) that bind to those things.
(I'm heavily paraphrasing and I don't have the book handy to refer to, sorry)
[0] https://www.amazon.com/Immune-Journey-Mysterious-System-Keep...
What is the antonym of "machine learning"? because this looks like machine learning to me.
Electrical is well-suited to modeling with ordinary differential equations. By making design decisions such that linear time-invariance is a valid assumption, we unlock really simple, powerful, and broadly applicable techniques for modeling, measurement, and control design.
Most biology isn't designed, which is probably why those papers had to jump right to much more advanced and specific techniques for modeling and control. With no useful on-ramp for most biologists, modeling and control will have to be a deliberate collaboration between two specialized disciplines.
I wish them well, but will note my pet peeve about the so-called immune system:
The antigens we produce are not the entirety of the immune system. It's like saying the alarm is all that's involved in securing a building and hand waving off the walls, doors, locks, etc.
There are myriad ways to trick the white blood cells and the body also has ways of killing stuff other than the white blood cells.
https://news.ycombinator.com/item?id=34639722
Which is to suggest that their focus on white blood cells seems overly narrow to me for a "groundbreaking plan to map the entire human immune system."
Methodologically, no one solves a hugely complex system by looking at the whole thing.
We've made tremendous progress when investigators focus on one interaction and do the bench work to isolate and prove features of it.
It's true that researchers need interaction data banks, e.g., to identify all the components in a metabolic cycle, to see what might be affected when part of that is amplified or suppressed.
Epidemiology is notoriously unhelpful in immune research because there are not enough numbers/powers to overcome individual variability.
Researchers largely work on models that operate immediately. There are no in-vitro models for immune learning, and few even for the relevant gene regulation.
People have tried and failed to build a control-system model for even the simplest interactions of a single cell.
Synthetic biology companies have burned billions with few results. I wouldn't expect someone with 40 years in HIV research to have the technical or leadership skills to pivot to the immune system.
Sure, there will also be many medical benefits from gaining this knowledge, but at the same time, there needs to be a global treaty-enforced ban on the deliberate biotechnological production of new infectious diseases by either private or public research labs.