AI Will Create More Jobs & Save Lives
By Dr. eMalick G. Njie, founder of Ecotone, using AI to cure genetic diseases.
Bio: Dr. Njie is the founder of Ecotone, a company with a vision of AI-designed medicines to cure rare genetic diseases.
Dr. Njie key background highlights:
Senior neuroscientist in genetics in Nobel Prize-winning laboratory at Columbia University
2x National Science Foundation SBIR winner
Founder of Genetic Intelligence, the first medical AI company from Columbia University
Edited by Lucy Danziger, formerly Editor-in-Chief of SELF magazine
Nala and Teddi
When Nala Shaw was born in 2020 in Northumberland, UK, her parents had no idea she was harboring a rare genetic condition. As a toddler, she appeared healthy, loved to laugh, spin around and play. Only when she reached the age of three did she begin to occasionally stumble or show signs of a tremor. The couple suspected something might not be right with their daughter. They took her to a neurologist, fearing a brain tumor. The doctor told them Nala didn’t have a brain tumor, but instead of this being good news, he delivered an even more devastating diagnosis.
Nala was born with Metachromatic Leukodystrophy, or MLD, essentially a faulty gene that meant her body was missing an important enzyme called arylsulfatase, which leads to the buildup of fatty or oily chemicals called sulfatides in the brain and nervous system. MLD causes the stripping cells of their outer layer, and destroying healthy neuron connections.
Eventually MLD would progress and end up robbing Nala of the ability to walk, talk, eat, see or hear. Most babies with MLD don’t reach the age of five.
When the Shaws found out they were both carriers of the MLD gene, they had their younger daughter Teddi tested. It was such crushing news that she, too, had the rare and debilitating condition.
But in Teddi’s case, then just 10 months of age, there was hope. She was young enough to qualify for an expensive new treatment that had been created over the course of 20 years of research by doctors in Milan, Italy. The gene for MLD had been identified, and a treatment called Libmeldy allowed doctors to replace the stem cells in Teddi’s still young system with new ones that had the genetic flaw fixed. It was experimental and costly.
The British health system footed most of the bill, (£)2.875 GBP (upwards of $3.5 million dollars) and with that single medical procedure, in which the faulty gene was swapped out with a working gene, Teddi was considered MLD free. In just one treatment, where the healthy gene was embedded into a harmless virus and delivered by a blood infusion, Teddi was cured.
The Shaw parents felt simultaneously grateful and heartbroken, since Nala’s diagnosis was made in time to save Teddi from the lifelong effects of MLD, but not in time for Nala. This is where the use of AI to create genetic medicines could eventually help: Doctors working in genetic treatments know the timeline for developing these long-sought cures has taken decades, and the treatments now underway are developed at an exorbitant cost to both patients and insurance companies, and in Teddi’s case, with government subsidies. It was the most expensive medical treatment in NHI’s history, a 2023 BBC News article reported. AI can change all this.
Developing genetic medical therapies such as Libmeldy requires first identifying the exact location of genes involved, and then developing a treatment for the faulty DNA sequence. It means finding a needle in the proverbial haystack or a few grains of sand on the ocean floor, since there are more than 3 billion possible combinations.
It is one area where AI can work faster than humans, to detect the exact combination of gene sequences responsible for any number of genetic ailments– from childhood blindness or any number of rare conditions present at birth, to diseases like Type 1 Diabetes, Alzheimer’s, Parkinsons, Crohn’s Disease, and other genetically-involved conditions that appear later in life.
This race against time can save lives, improve patient outcomes, and make treatments available at a fraction of the current costs, ultimately making it possible to treat and cure diseases that once had no such hope. Babies are not even genetically screened for MLD, as part of the Newborn Screening Panels, since until recently a treatment was not even an available option.
Given the costs, it is still out of reach for most people.
As sad as the outcome was for Nala, whose neurological condition will deteriorate every year for the rest of her life, the hope is that AI will help doctors and researchers complete the blunt-force data crunching in hours instead of years, allowing them to identify genetic flaws and then design cures such as Teddi’s in the future, and do it on an expeditious and more affordable time table.
Not only can AI significantly shorten the amount of time it takes to develop new drugs and genetic therapies, but instead of developing one drug or treatment at a time, AI can create a platform of knowledge – a base from which to develop specific genetic therapies – that makes it possible to simultaneously create new drugs and treatments at scale.
This type of baseline, from which launching medicines can happen more rapidly and cost effectively, is the goal of scientists in the private and public sector. That’s the true promise of AI, which has been vilified of late. Instead of thinking of AI as a threat to jobs and livelihoods, we need to think of this brave new world as a game changer for scientific breakthroughs that will make genetic conditions as treatable as every other disease we currently have medical cures for.
AI needs to be seen not as a threat to humanity but as a tool to design genetic treatments and drugs that could allow genetic medical breakthroughs to happen quicker and cheaper giving doctors the tools they need to change patient’s lives.
AI: Awe and Distrust
The release of ChatGPT in 2022 demarcated the beginning of a new technological era for humanity, one not seen since the industrial revolution. Just as engineers replicated a horse's power when creating a motorcar and the shape of bird wings when designing an airplane, in Artificial Intelligence we have replicated the human brain’s ability to learn.
The reaction in the ensuing months as people began to play with what AI can do was rapid and widespread, with AI occupying the cultural zeitgeist. In just weeks, ChatGPT became the fastest adopted application in history, and since then AI has instilled an array of fear, ranging from job security concerns for writers and content creators, to the very question of human agency in the new AI centered world. When wide-spread fear emerges so rapidly, history has taught us, there are deeper truths behind the reaction.
First, Here’s How AI Works.
In May 2023, Sam Altman (the recently restored CEO of OpenAI) and other AI leaders were invited to appear before Congress to inform lawmakers on the potential of AI. Altman, instead of educating Congress on how AI works, further provoked fear by arguing that government regulators need to ‘save him and his AI compatriots from themselves.’ Such a fear-based narrative goes counter to human agency that democracy is supposed to represent.
When people feel as if a technology is too complicated to understand, they throw their hands up in frustration and become more willing to be guided or misguided by the powers that be. However, AI is not a complicated black box. Essentially it’s just an artificial brain.
As a scientist working on developing genetic treatments and medical uses of AI, I have been asked to explain in layman’s terms how to think about the future of AI— and I want people to understand this isn’t bad. It is not HAL from 2001 a Space Odyssey, nor the malevolent “Entity” from this summer’s Mission Impossible: Dead Reckoning, where the machine learns how to trick people.
AI will make available exciting new treatments for diseases that currently require us to work with and make sense of vast amounts of data. What used to take years to sort, understand and react to (such as the genome or millions of individual patient cases around the world) now takes AI a matter of minutes, since the artificial brain can process data much faster than humans can.
This is a quantum leap forward, made possible by machine learning. AI is fundamentally different from older generations of computation such as programming that runs the OS on your phone, computer, mainframes, and supercomputers. The difference: Computers need us to write the code, essentially a recipe for every action. AI learns from experience, teaching itself to cook.
The latest AI is closer to the human brain than to a computer. Here’s why.
An Engineer’s Perspective
An engineer will say ChatGPT-like AI is mechanistically based on next word predictions no different than the cloze tests we all took in the vocabulary section of our SAT tests in high school. In short, a cloze test asks one to fill in words that have been randomly removed from a paragraph to show that one understood the context of the paragraph. Do this several billion times, and what emerges is ChatGPT.
This description is technically correct as this is the objective that guides Large Language Models (LLMs) like ChatGPT to understand and become conversational in languages like English. The objective powers AI much like a V6 engine powers a race car.
The engineer’s description is accurate, but only explains how AI works, rather than why AI works. It lacks something deeper, that is borrowed from the animal kingdom, and is now in computers. A neuroscientist’s description is thus more apt.
Learning Through Positive Selection
In the growing brain of infants, biological neurons (BN) are born and try to form connections to each other based on factors such as outside stimuli (i.e., smell, taste, touch, sight, etc). This is why when a baby is first learning about the world around him or her, you’ll watch as he or she puts everything in their mouth: A rattle, a toy, then finally the bottle. The baby is learning what’s not food, and what is. This sorting by selection is the closest thing to watching neurons make successful connections, a largely random process.
Here’s how positive selection learns. Four BNs are born that each have the potential of detecting the color green. One of these neurons happens to be right and chooses the correct color, which is rewarded when the caregiver exclaims, “Good! Yes, the trees are green!” This positive feedback reinforces the connectivity and firing pattern of that neuron to remember the correct answer.
Our genetics cause this BN to grow since hearing “yes, green!” affirms correctness, while the other three incorrect ones fade and die. But other neurons are rewarded for recognizing blue, or red, or yellow, or purple, until the brain learns which neurons are positively selecting the correct answers, and future learning gets built on these pathways.
From Random Guesses to Machine Learning
Positive selection is initially based on random guesses, but when it gets it right, and then repeats that selection multiple times, the results are no longer random and now those neurons specialize in sensing green, while other neurons can identify red, or blue or yellow, and so forth.
The fundamental biological phenomenon of positive selection is the reason for the most fabled term in neuroscience: “neurons that fire together, wire together.” These pathways grow stronger and create more connections as we learn, age, and take in more information. The more information, the more we learn. But for humans, reading, watching and learning takes time.
Artificial intelligence is inspired by these same principles of positive selection, but the difference is that it can be achieved much faster. Artificial neurons (AN) undergo positive selection based on random draws of numbers to match training data. Let’s take four ANs that have the potential for detecting the color green from pictures of trees that were given to the AI as training data.
Since green is represented by a numerical value in a computer, such as the arbitrary number 5. Our four AN’s job is to come up with the number 5, thus making them a green detector. How AI works is by giving these AN’s numbers randomly, say 7, 3, 5 and 9. The AN’s with 7, 3 and 9 have the incorrect value, and are thus discarded, while the AN with 5 is kept and built upon.
Any time a new picture of a tree is given to the AI, this AN, which is sensitive to 5 (i.e., the color green), will turn on. Do this many times over and you will have ANs that detect green, red, blue and so forth. Thus, scientists can create very straightforward parallels between biological neurons and artificial neurons. It’s only a matter of time before we input enough data that the AI learns and teaches itself more complex levels of “thinking” based on earlier correct answers.
AI Is Not Just a Better Version of Your Mainframe Computer
Positive selection based on random draws of numbers is a paradigm change from programming such as in Windows, MacOS, websites, etc. These older generations of computers required numbers to be input in manually by humans.
To make a computer program that allowed a webcam to detect the color green, a human being would input into the webcam’s software the number 5. The webcam’s software would not itself find the number 5. The older programs were hardcoded to directly detect green. While AI is softcoded to eventually “learn” to detect green.
The game-changing application of AI is its ability to sensibly integrate millions and billions of data points. This feature is the missing link that ties the Big Data era of the last decade to the exponentially rising use-cases of AI today. It is the electricity that we cannot see, but we can understand is underlying the era of AI we are entering.
Deploying AI In Service of Humanity
The ability of AI to process massive amounts of data and yield an output that directly helps humans is demonstrated by the work being done by the team at Flatiron Health in New York. Flatiron, which focuses on cancer therapeutics, has been scanning vast amounts of health record data so that their AI can provide bespoke insight and allow doctors to understand how to better treat individual patients.
Learning From Millions of Medical Records
Some of these medical records are electronic, however many of them, such as prescriptions, are handwritten by doctors (who notoriously have terrible handwriting). To give an idea of the scale of what it takes to digitize these records, Flatiron has hired approximately one million nurses around the globe to transcribe these records into a database.
The records are then used as training data for AI to look for patterns. For instance, doctors in Iowa may prescribe a particular medication at a specific dose, while doctors in Austria or Asia prescribe this same medication at a different dose. Which is working better, and which patients are responding?
The Correct Dosing of Medications
Dosing accurately is a matter of life and death for some patients and oftentimes is set arbitrarily based on calculations made by pharmaceutical companies that conducted animal tests –which may or may not necessarily directly translate perfectly to humans.
The fact that doses prescribed by doctors in various parts of the world are often different presents a natural experiment, which one could use to determine the most effective dose for future patients. However, individual doctors are often unaware of the activities of their counterparts in other parts of the world.
Even if they could see each other’s data, one would need to be able to find outcome data from thousands of doctors in one part of the world compared to other parts of the world, where populations, lifestyles, treatments and protocols may vary.
Billions of Data Points, Saving Lives
These data sets are composed of billions of data points, and until AI became widely available, it was impossible to tie them together or interpret the data in a practical way. However, with modern AI tools, scientists make life-saving discoveries regarding dosages on a daily basis.
Beyond dosing, it is also critical that doctors have the best information at their fingertips as they pace busy hospital floors making decisions such as diagnoses that can impact life and death.
When we consumers are searching for a primary care physician or a specialist, the most important factor one relies on is experience. For instance, when I see my physician, I want her to have seen a thousand patients that had the same condition as I do, so she is informed to make the best decision regarding my health.
Helping Doctors Make Clinical Decisions
Doctors rely on a heavy backend of human systems architecture that the patient never sees in order to do their jobs well. For instance, in the US, most hospitals use Clinical Decision Support software from companies like Epic Systems to help doctors who are writing prescriptions make sure the drugs they are giving their patients will not have negative interactions with other drugs the patient may be on. This is called medical “reconciliation.”
Medical reconciliation has only recently become automated. As recently as 2008 the average hospitalized patient was subject to at least one medication error per day, according to research published by the NIH. Medication errors once were the most common cause of patient safety errors. By 2018 these reconciliation processes had improved, but were still subject to human error. Currently the system in place is designed to automate that, but it still takes longer than it needs to.
Typically a doctor will have a one-on-one with a patient while continuously inputting their medical information into a tablet and uploading it to Epic’s software. The doctor must list all the medications the patient is on, add medications to be prescribed and check against possible adverse interactions. Epic checks against thousands of data points input by other specialists who are doing research on drug interactions and this information allows the doctor to make informed decisions about which drugs to prescribe.
The average time it takes to reconcile the medicines and check for adverse effects as of writing can be many minutes. AI could help reduce this time while providing better results. This is critical since for some patients, the speed and accuracy of treatment is a matter of life and death.
There are now AI approaches currently being developed to shorten the medication reconciliation process. OpenEvidence, based in Nevada, is doing work that hints at what this future looks like. Every day, millions of medical articles and clinical trial results are published, some of which contain key information needed by doctors for evidence-based decision making.
Genetic Treatments & Gene-Based Medicines
Researchers publish new and promising findings in about three thousand articles every day The average time it takes a medical professional to thoroughly read each of these articles is approximately one hour.
For one specialist in the backend of Epic to read all the articles coming out and relay key information to a doctor can take more than 50 days. This obviously is not good enough when doctors are pressed to make life and death decisions and treatment recommendations in seconds.
OpenEvidence is tackling this problem by using AI to read newly published medical information. This data, which pushes beyond the billions of data points, is at a scale that can only be processed by AI, and is critical for putting new health discoveries at doctors’ fingertips.
Next Generation of Treatments for Parkinson’s, Alzheimer’s & More
An aspect of medicine that AI will be essential for is the next generation of medicines. These are colloquially called ‘Precision Medicines,’ as they are based on personal characteristics of patients, such as their genetics.
The current generation of medications we are accustomed to were made using a generalist approach. For instance, when a pharmaceutical company releases a new drug, they typically say it is for everyone, rather than for a select few. (This is market driven since drugs are expensive to produce so finding a drug that large swaths of the population needs is a gold mine.)
Oftentimes these drugs work best on some patients while causing major side effects on other patients. Indeed, we live in a world where the “side-effects” are considered normal and accepted by doctors and patients alike.
And it is not unusual to have a doctor prescribe another medication to treat the side-effect. And this medication itself has a side effect that yet another medication has to be prescribed for, and so on and so on. Moreover, we are accustomed to our medications treating rather than curing.
Precision Medications hold the promise of changing these expectations. They are designed to be customized to small groups of patients than can number in the hundreds, and offer a lifetime cure without causing side-effects. There have been several examples of successful Precision Medicines. These include the new treatment that was used to cure Teddi Shaw’s MLD.
Precision Medicines Are Changing Lives
Genetic treatments such as this take decades to produce, require years of research, and once they are available, the timetable to use them can be short, as witnessed by Teddi and Nala’s parents, who expressed feeling both blessed and heartbroken.
Teddi’s life is forever changed by the gene treatment, while Nala’s condition deteriorates by the day, since she was not treated in time. For genetic conditions, the timetable, costs and feasibility of such life-changing medicines will be revolutionized with the use of AI.
However, these medications have yet to become widespread because there is a limiter that has proved insurmountable for the massive deployment of Precision Medicines. This limiter is our own knowledge of the genetics that underlie many diseases, particularly inherited diseases.
The number of genetic diseases is estimated to be at around 10,000, ranging from well-known diseases such as Parkinson’s to many that are so rare most people have never heard of them.
Searching the genetic code is like trying to find a spec of sand at the bottom of the ocean.
For context on the size of this challenge, the English language has approximately one hundred and seventy thousand words. The genome of a single human has approximately three billion parts. Researchers have made important inroads such as the Human Genome Project, but much of the genetic causes of disease is yet to be discovered.
The Human Genome Project, from the Clinton era in the early 2000’s mapped all of the parts in the human genome (known technically as bases). However, the meaning of any one of these parts in relation to diseases remains an unmet challenge. The private sector is stepping up. Just as Flatiron and OpenEvidence have leveraged AI to solve Big Data challenges, several companies now see AI as critical to finding out which parts of our genetics are causing inherited diseases.
For babies born with genetic conditions, including certain types of blindness, and other patients with genetic diseases—simply by patching or replacing a specific gene—it is possible to offer a lifetime cure. This type of breakthrough is not just a treatment, since if we can identify and replace the faulty gene, it is a cure.
Ecotone is Using a New AI to Cure Rare Genetic Diseases
My work at Ecotone involves building foundational AI, called a Large Genome Model (LGM). This generative model is similar to Large Language Models such as ChatGPT, but instead of predicting the next word in sentence, it predicts the next genetic sequence.
The core problem that Ecotone is solving is discovering which of the parts in the human genome are causing genetic diseases. To do this, its Large Genome Model uses training data composed of human genomes from various human heritages. There are approximately 760 human heritages globally, each with slight differences from one another.
Multiply the number of heritages by the three billion parts in the human genome and you get 2.3 trillion possible parts, any one of which can potentially be causing genetic diseases. What is remarkable is that a Large Genome Model processes the mystery DNA letter code of life intuitively as if it’s a first language, and thus can handle this type of Big Data quickly.
Inside of the model, we can begin to draw insights regarding the movement of genes between heritages. If a person of a blue-eyed heritage moves to a part of the world where there are only green-eyed people, and the blue-eyed person has children there, the LGM traces the flow of genes that cause blue eye pigment from the blue-eyed heritage flowing into the green-eyed heritage.
This tracing of gene flow that is modeled inside the LGM can also be done for genes causing diseases as these also move into new heritages with the natural movement of people over millennia. We of course do not have the genome of that first blued eyed person or the first person manifesting a rare genetic disease as they have been deceased for millennia. However the history of these individuals is carried in scatterings within the genomes of thousands of descendants present and alive today. This is Big Data that our LGM is precisely built to model.
Ecotone’s mission is to make AI-designed Precision Medicines for rare genetic diseases that potentially save babies’ lives and improve outcomes for patients of all ages. This is made possible by using the power of AI to track vast data of gene flow.
Challenges Remain, Such as “Hallucinations”
Even the best AI systems, like ChatGPT, will produce falsehoods. These so-called ‘hallucinations’ are able to confidently capture sentence structure and nuances and misstate facts, especially in specialized areas such as legal precedent. One such incident led a lawyer in South America to submit an entire legal brief to court based on fictitious cases made up by ChatGPT.
In my experience, asking ChatGPT to describe a gene that I know does not exist has resulted in ChatGPT outputting a scientific abstract and describing the gene with supporting information in the form of published medical manuscripts. It even included author lists that are names of reputable scientists I personally knew.
Yet it was wrong, or worse, it was trying to pretend it knew information that it didn’t know because the gene doesn’t exist. ChatGPT is childlike in this way: It would rather make up an answer than say: I don’t know. Or I can’t find anything on that.
All the information ChatGPT returned to me was false – the authors did not write the manuscripts, and the manuscripts did not exist. This is alarming, to say the least, and when the stakes are human lives (not just legal briefs) AI becomes as much of a liability as an asset.
Why AI hallucinates is a topic of ongoing research. What is clear is that in the realm of web tools like ChatGPT, hallucinations can be amusing or a nuisance; however, there is no room for them in medicine as patient health is at stake.
Human intervention is needed
Another challenge of AI is how it will impact jobs. At FlatIron Health, AI actually helped create more work, as one million nurses were required to help input the doctors’ data into the system. These humans were needed to help AI do its job.
So of course AI will take up some tasks, but as far as input, and fact checking output, human beings are still needed to be on both ends of the AI functionality. There will be similar examples in the future, but unfortunately there will be many jobs permanently lost, or transformed, by the introduction of AI into industries.
Change is coming. Just as the dressmaker during the Industrial Revolution leveled up from making one garment per month to hundreds of garments a day, employees will see dramatic improvements in their productivity due to the assistance of AI.
This increase in productivity will in turn lead to more economic activity and the making of new industries that will produce new types of jobs. Unfortunately many people will be caught in the weaves of this change before they can react to the fact that their industry is permanently changed. We fortunately have foresight thanks to the Industrial Revolution. We can act to protect the livelihoods of those who will inevitably lose their jobs by creating new opportunities.
Conclusion: AI Is a Gamechanger In Medicine and beyond
Whether it’s medicine or eminent existential threats such as climate change, AI is helping us solve problems that would have taken years to wrap our minds around. With climate change, we have little time to waste.
In climate applications, AI is helping to calculate the carbon sequestration of oceans, which naturally absorb more than half of the earth’s carbon. Sometimes called Blue Climate or Blue Carbon, companies such as TrueOcean are using hydroacoustic technology together with AI to map vast areas of underwater seagrass meadows that are absorbing carbon at a rapid rate.
Blue Carbon is thought to be 50 to 60 times more efficient than Green Carbon (i.e. trees on land) at sequestering carbon, thus determining it’s value concretely with the help of AI is essential. This includes Blue Carbon-rich island nations, who are often most vulnerable to rising ocean waters. Using AI they can better participate in the lucrative and growing global carbon credit marketplace, where companies purchase carbon offsets, meant to stymie climate change.
The field of medicine is one example of how AI can improve lives and power humanitarian science, but it is only the beginning. The concerns about AI replacing jobs in creative fields are valid, but there will always be the need for humans to work with AI. It is up to us to see the possibilities in time to change and save lives.
Sidebar:
We Asked ChatGPT for Suggestions for Medical Uses of AI. ChatGPT’s Answer:
Drug Discovery: AI algorithms can rapidly analyze massive datasets to identify potential drug candidates for various This accelerates the drug discovery process and holds promise for finding treatments for rare diseases.
Personalized Treatment: AI-driven models can analyze patient data to tailor treatment plans based on individual genetics, medical history, and other This leads to more effective and personalized medical interventions.
Remote Patient Monitoring: AI-powered wearables and sensors can continuously monitor patients' health parameters and provide real-time alerts to healthcare providers, enabling early intervention for critical
Predictive Analytics: AI algorithms can predict disease outbreaks and identify at-risk populations by analyzing patterns in data such as symptoms, geographic locations, and environmental
Telemedicine: AI-enhanced telemedicine platforms can provide virtual consultations, diagnosis, and treatment recommendations, especially beneficial for underserved and remote
Sources:
UK’s most expensive drug Libmeldy saved Teddi Shaw, but is too late for her sister BBC 2023
Senator Fine expands newborn screening test to include MLD 2023
OpenAI’s Sam Altman To Congress: Regulate Us, Please! Forbes 2023
At 24, Two Entrepreneurs Took On At 32, They’re Worth Hundreds Of Millions Forbes 2018
Assessing the Value of ChatGPT for Clinical Decision Support Optimization medRxiv 2023