Transcription: Inflammation is all you (don’t) need
A trendy hypothesis is that chronic inflammation is responsible for many of the symptoms of growing older
I am trying something new here. To get myself in the habit of writing regularly, I’m launching a new series: Transcription, a weekly roundup of recent news and research that’s advancing our understanding of human biology. As with my usual long-form pieces, there will be an emphasis on the immune system and applications for artificial intelligence to understand the immune system. Please let me know what you think: This will be a huge work-in-progress for a while, and I need you to tell me what is most interesting for you to read. Welcome to Transcription!
Inflammation is all you (don’t) need
A trendy hypothesis is that chronic inflammation is responsible for many of the symptoms of growing older. And, in many ways, it makes sense. Inflammation, in the form of activated immune cells, causes a lot of deleterious side effects. Cells known as fibroblasts, which produce a lot of the connective tissue throughout our musculoskeletal system, go into overdrive and increase their production of fibrotic proteins during inflammation. This causes the buildup of scar tissue. Inflammation can cause heart attacks to be more frequent and more severe, and increases the risk of diabetes and Alzheimer’s, among other diseases. Inflammation also uses up a lot of energy, and that energy is taken from other bodily processes, resulting in fatigue.
Think of chronic inflammation like climate change: as the Earth’s oceans warm, the chance of more damaging weather events occurring goes up. Hundred-year-storms become ten-year-storms. As we age and chronic inflammation grows, the chance of dangerous diseases occurring goes up. We can handle the side-effects of inflammation when fighting an infection or disease. The immune cells are ultimately saving us! But chronic inflammation is devastating, especially when compounded over decades.
So some researchers, doctors, and civilians are interested in slowing aging. And one of their initial targets is to reduce chronic inflammation. Stop chronic inflammation from increasing people’s susceptibility to deadly diseases such as diabetes, cardiac hypertension, or Alzheimer’s. Get the immune cells to chill out. There is a careful balancing act, though. Too much immunosuppression can make patients susceptible to infectious diseases, like pneumonia, which also kill a lot of (particularly elderly) people worldwide. Their quest is to find the right balance: stop chronic inflammation that is ultimately bad, but maintain peak immune responses to defend from viruses and bacteria and cancers and other nasty illnesses.
The most tested, discussed, and well-known medication for chronic inflammation is a compound called rapamycin. Rapamycin was discovered in a species of bacteria found on Easter Island (and is named after the island, Rapa Nui). It has immunosuppressive properties that make it especially good at preventing kidney transplant rejections, which is what it was originally approved for by the FDA in 1999. But it turns out to be more than just a generic immunosuppressant. Rapamycin targets an important protein system in cells called the mammalian target of rapamycin (mTOR, named so aptly because it was discovered decades after rapamycin). Inhibiting mTOR makes immune cells less sensitive to inflammatory proteins, hence the immunosuppression, but also enhances a cellular process known as autophagy. Autophagy is basically the garbage disposal system of a cell: as a cell produces proteins over its lifetime, those proteins can build up and get in the way of normal cellular function. This buildup causes the cells to be stressed out because it cannot function properly, and this stress response can activate local immune cells, who think the cell may be infected. Autophagy cleans up the protein buildup, preventing the internal stress in the cells and stopping this immune response. The theory is rapamycin may be able to reduce chronic inflammation without impacting acute inflammation by 1) desensitizing immune cells to inflammatory proteins and 2) cleaning up protein buildup in cells throughout the body.
A piece in the New York Times this past week delves into the growing interest in taking rapamycin to slow or prevent aging:
On podcasts, social feeds and forums devoted to anti-aging, rapamycin is hailed as the “gold standard” for life extension. Longevity influencers Dr. Peter Attia and Bryan Johnson are believers, both saying they’ve taken rapamycin for years, and touting research to their millions of followers that shows the drug can extend the life spans of mice by over 20 percent.
Preliminary data presented earlier this year at the American Aging Association annual meeting suggested that rapamycin also works in closer cousins to humans: Marmosets given the drug showed roughly a 10 percent increase in life span.
The promising data in many animal experiments have led to a growing population of those looking to slow down their own aging process by enrolling in clinical trials or by taking rapamycin off-label.
After taking a small weekly dose for approximately 15 months, Mr. [Anthony] Holman, who lives outside Raleigh, N.C., said he hasn’t experienced many changes, positive or negative, though he has noticed a subtle decrease in daily aches and pains. “It’s almost like taking vitamins,” he said. “You don’t take vitamins because you’re expecting some immediate benefit. You take it to hopefully see a benefit over time.”
But while users are optimistic and the evidence that rapamycin can increase longevity in animals is promising, the research in humans is thin and long-term side effects are uncertain. In the few studies in which rapamycin has been compared to a placebo, tangible benefits are hard to come by.
The challenge with finding a drug to treat aging is the clinical data takes forever to come in. Without fancy anti-aging drugs, we still expect to live more than 70 years. Figuring out if a drug enhances that number means the clinical trials take decades to obtain their data. And that has to happen before companies can sell and doctors can prescribe the anti-aging drug. So, many companies and researchers use some proxy data as their endpoint, like the incidence of heart attacks or diabetes or other diseases at least partially attributed to inflammation. And that data has been less compelling. But aging impacts everyone, so there are bound to be some people who see promising data elongating animals’ life spans and want to take matters into their own hands, because their life can’t wait.
Automating drug discovery with artificial intelligence
Many people, including myself, believe that biology and biotechnology is one of the most promising applications for artificial intelligence. There is a tremendous amount of biological data out there, collected by automated experiments that can generate massive amounts of training feed to plow into cutting-edge artificial neural networks. And there is a huge need: Biology is hard. The complex nature of biology, particularly human biology, is difficult to fathom.
And lots of researchers and companies and investors are betting that artificial intelligence can, among other things, understand the complex human biological system using laboratory or clinical data. Many applications of biological artificial intelligence are in drug discovery, where researchers like myself are trying to design new medications. In this field there is a hierarchy of problems that these models can attempt to solve. In order of least to most valuable:
Screen a pre-designed library of target compounds before laboratory experiments
Design novel compounds to target particular proteins chosen by the researchers
Propose which proteins to target in the first place to treat a particular disease
The better the model is at narrowing the solution space of possible therapeutic medicines, the more valuable that model is to drug discovery.1
One strategy to build models that can understand complex systems is what’s called a “foundation model”. These models are given unlabelled sequential data—massive amounts of it—and apply a tremendous amount of electricity and computational power to find relationships between different parts of the sequence. ChatGPT and the other GPT models from OpenAI are an example of a foundation model; in this case, the sequence is a sentence or paragraph or book of human writing.
And biology has a lot of sequential data! The language of biology is in DNA, a sequence of nucleotides A, C, G, and T that encode our proteins, how much of them to produce, and in what conditions to produce them in. Proteins are just sequences of amino acids encoded by that DNA, and their sequence determines how the protein folds, which ultimately governs how that protein functions.
There are numerous organizations out there working to apply this sequential data to a foundation model as a way to uncover human biology’s complexity. One of the most advanced is EvolutionaryScale, who recently unveiled their latest-generation biological foundation model: ESM3. The ESM3 model was trained on billions of sequences of biological data with many powerful computers analyzing those sequences. And they showcase ESM3’s power by asking it to generate a new sequence for a protein that fluorescenes green, from scratch, and compared it to a naturally occuring protein called green fluorescent protein (GFP) that is very frequently used in cellular biology experiments. The sequence it generated was not only a protein that fluoresces, it was also so different from natural fluorescent proteins that it would’ve taken biological evolution 500 million years to mutate an equivalently obscure sequence.
We could use a model like this to design better antibodies, for example. A paper from earlier this year used the previous generation of ESM models to identify mutations in therapeutic antibodies to make them better at binding to their target. The model was able to pick mutations in the antibody sequence that improved their binding-affinity to their target protein by seven-fold. This model, in other words, could 1) screen sequences given a starter sequence (the existing antibody) to propose which variants are most likely to succeed in downstream lab experiments.
The new ESM3 model may be able to 2) design new therapeutic sequences to fit a particular function, like bind to a certain protein, or inhibit a target receptor, or… glow green? Such a tool would be a huge step forward in foundation models for drug development. Imagine if a researcher can ask the model to generate an antibody that targets a particular protein, for example, and the model can create a completely unique protein to do that. From scratch, without the researcher giving it a starting position. “What an amazing time to be a researcher working in biotechnology!”, I think to myself, a biotech researcher, as artificial intelligence threatens to take my job.
No one wants to take the groundbreaking CRISPR therapy
We recently dived in to gene editing, using the novel CRISPR therapy for sickle cell anemia as an example for the technological marvel and clinical challenge it embodies. Detailing the clinical application of CRISPR therapy, I wrote:
The patient has a bone marrow sample taken to collect as many blood stem cells (known as HSCs, or hematopoietic stem cells) as possible[, and then undergoes] intensive chemotherapy to kill off their residual blood cells and HSCs. The [CRISPR] edited HSCs will then be transplanted back into the patient’s bone marrow. The patient will be monitored for any signs of immune rejection to ensure the cells are engrafting and beginning to produce healthy blood cells. This whole process from start to finish can take 6 months or more.
Not only is the process time-consuming, labor-intensive, and enourmously expensive, it is incredibly taxing on the patient. Too taxing, it turns out, for many would-be patients. Reuter’s reports:
The new one-time treatments, approved in the U.S. last December, have so far been used on around 100 people globally, including in clinical trials. They require chemotherapy, which raises the risk of cancer and can cause infertility.
Some patients say the time involved – up to a year – is a daunting prospect for anyone whose condition is not critical.
[…]Dr. Michael DeBaun, director at Nashville's Vanderbilt-Meharry Center of Excellence in Sickle Cell Disease, questioned the logic of recommending a new gene therapy used so far on so few people.
"You wouldn't do that for cancer," he said. "You would only offer that to people who had the most severe disease who were likely going to die."
Dr. Mark Walters at UCSF Benioff Children's Hospital in Oakland, California expects the therapies to initially be used for about 10% of sickle cell patients, noting the field is moving quickly as researchers aim for therapies that may not require chemo.
"The chemotherapy drug we use causes infertility in most," Walters said.
The technology behind the CRISPR therapy is awesome. In patients treated during clinical trials, 93.5% of them are free of the painful symptoms of sickle-cell anemia. It really is a cure for the disease. But the clinical application of CRISPR is far from polished. Trading sickle-cell for infertility is not a fair trade for most patients, especially when there are other treatment options available. It does not help that the CRISPR therapy costs $2.2 million, or that there is an ongoing lawsuit over whether the drugmaker (Vertex Pharmaceuticals) or the government is on the hook to pay for fertility treatment in patients under Medicaid.
CRISPR as a gene editing tool is astounding and has changed laboratory research forever. But its first clinical product is just not an enticing solution for many patients living with sickle-cell anemia.
What does that mean? Well, consider there are 20,000 protein-coding genes in the human genome. That means there are (roughly) 20,000 protein targets to inhibit, or activate, or replace, to treat diseases. If you are targeting just one protein. If you instead target two proteins, the number of targets (20,000 choose 2) is 199 million. If you target three, that number jumps to 1.33 trillion. And that is just for the target protein. You then have to design and test your therapeutic(s) to make sure they actually target your particular protein. Companies often screen hundreds or thousands of compounds to test if any of them target their protein of interest before finding a hit. So they have developed optimized high-throughput experiments to try and test as many compounds as quickly as possible. But all this takes time and money. If a computer can perform these screening experiments beforehand, thats a big save on would-be wasted materials and time. If it can help design the compounds to screen, that’s even better (it saves a lot of chemists and molecular biologists a lot of time). But if it can propose which protein to target in the first place to treat a particular disease? That model will change the world.