It’s an all-too-familiar story. A laboratory somewhere announces that they have discovered a chemical formula that cures a previously intractable disease like cancer.
It works perfectly in Petri dishes, and in mice, and even in monkeys, and now the human trials begin.
Months later, we read that the formula not only didn’t work, but it made the patient sicker.
Why is pharmacology so difficult?
Biological evolution is nothing like human thinking. Biological evolution is much more like machine learning.
Consider the article titled, “Attacking Machine Learning with Adversarial Examples.”:
Adversarial examples are inputs to machine learning models that cause the model to make a mistake; they’re like optical illusions for machines.
To get an idea of what adversarial examples look like, consider this demonstration from Explaining and Harnessing Adversarial Examples: starting with an image of a panda, the attacker adds a small perturbation that has been calculated to make the image be recognized as a gibbon with high confidence.
To the human eye, both panda pictures seem identical, but the machine “sees” them differently.
Pharmacology is difficult and machine-learning can be inscrutable, for essentially the same reasons. They do not follow the straight-line logic of human thought.
Consider this A to B puzzle.
To the human brain, the algorithm for traveling from A to B might be:
- Begin at A
- Travel along the dotted line to the airplane.
- Pass through the airplane to B.
Now consider this maze. What is your algorithm for going from one opening to the other?
Your algorithm might be:
- Enter at either the top-left or bottom-right opening.
- Take any route.
- If that route gets you to the other opening, stop.
- If that route doesn’t get you to the other opening, go to #5.
- Enter at either the top-left or bottom-right opening, and take a different route from any route you ever have taken previously.
- Go to #3.
That’s the way you, as a typical human, might operate.
But this is the way a machine might begin:
- Erase half the lines.
- If you come to a dead end, open it and continue.
- Fold the paper in half so one opening touches the other.
- Draw a new maze.
- Write new rules.
- Jump up and down on the maze.
- Marry someone who owns a compass.
- Ask for the solution.
In short, both evolution and machine learning defy human rules and logic. but after many, many moves, might take you from one opening to the other.
For instance, if you marry someone who owns a compass (#7), that person also might happen to own a diagram that happens to show the solution to the maze. That’s how evolution operates.
The process may require millions of iterations, but computers work fast and evolution has plenty of time.
So those millions of iterations very well could provide a solution. It’s similar to the old story about the infinite number of monkeys pressing an infinite number of computer keys, and at some point (an infinite number of points, actually) they type all of Shakespeare’s plays.
Humans always are time-constrained. Biological evolution is not, which means that the paths to the survival of life can be mind-bendingly complex.
I thought about this when I came upon an article:
How tumors hide from chemotherapy
For decades, cancer researchers have wondered whether they could starve tumors into submission by choking off their blood supply and thus preventing their fast-growing cells from getting enough food and oxygen.
They developed a drug, Avastin, that blocks a molecular signal triggering blood vessel growth, or angiogenesis.
But, mysteriously, Avastin failed to improve survival unless patients received chemotherapy drugs at the same time — implying that Avastin was somehow helping the chemo to be more effective.
That piqued the interest of Rakesh Jain, a cancer researcher. “How can a drug that kills the blood supply help chemotherapy? You need the blood supply to get the drugs into the tumor.”
He started digging deeper, and what he found turned conventional wisdom on its head.
The blood vessels that deliver food and oxygen — and chemotherapy drugs — to a tumor tend to be highly abnormal.
Instead of the usual large, straight, simply branched vessels, the ones in and around a tumor are often unevenly distributed, misshapen, and tangled.
As a result, some parts of the tumor end up far from any blood vessels and thus have little exposure to chemo.
Those same regions become starved of oxygen, and this hypoxia suppresses the immune system and also acts as a signal for the tumor cells to metastasize, or disperse to new sites.
Moderate doses of Avastin don’t outright suppress formation of blood vessels around a tumor but can actually make them look more normal, so that they can deliver chemotherapy more efficiently and evenly.
Jain’s clinical collaborators found — surprisingly — that patients who responded with increased blood flow to the tumor lived longer than patients in whom the blood flow declined.
Angiogenesis inhibitors work, in other words, for exactly the opposite reason than scientists initially thought.
Because we are the result of evolution, and evolution uses counter-intuitive “thinking,” solutions to our physical problems can come from that same, counter-intuitive thinking.
Consider, for example, the “illogical” the use of stimulants to combat ADHD:
Most parents wouldn’t give a child with attention deficit hyperactivity disorder (ADHD) a caffeinated drink, for fear that their hyperactivity would only worsen.
So why do doctors give stimulants to kids with ADHD? It seems so counter-intuitive.
Evolutionary “thinking” and machine learning are not straight-line. They both are the product of “random walk,” processes, in which each step is non-directed, but evaluated for success or failure.
If a step has some immediate value, or at least is not fatally harmful, it is retained as the basis for a next step, even though that earlier step may not have seemed to bring one closer to an ultimate solution.
Human thinking seeks solutions. We don’t have the time or capacity to retain and use all the various steps that might have had immediate value or not been fatallly harmful.
There are many routes to “survival of the fittest” (or more accurately, “survival of the fit enough“). Within our lifetimes, we can’t try them all.
Machine-learning works fast, but it doesn’t display its individual steps. It can tell you where it is, but not how it got there.
In the “panda/gibbon” example, above, we do not see the steps that led to the misidentification. We do not see the “logic.”
Because our brains are made by natural selection, while logic is an artificial construct, we do not know the details of why the following illusion works:
Biological evolution and machine learning are similar in that they operate via the random walk, a seemingly illogical, unpredictable, and indirect process, alien to human thinking.
Unless, sometime in the distant future, we know, then program a computer with, the cause/effect of every atom and every linkage in the human body, while overcoming the limits of the Uncertainty Principle, we repeatedly will be amazed that when starting from A, B, C and reaching X, Y, Z, we did not pass through L, M, N.
And that is why pharmacology is so difficult, really almost impossible.
Rodger Malcolm Mitchell
Search #monetarysovereignty Facebook: Rodger Malcolm Mitchell
The most important problems in economics involve the excessive income/wealth/power Gaps between the richer and the poorer.
Wide Gaps negatively affect poverty, health and longevity, education, housing, law and crime, war, leadership, ownership, bigotry, supply and demand, taxation, GDP, international relations, scientific advancement, the environment, human motivation and well-being, and virtually every other issue in economics.
Implementation of The Ten Steps To Prosperity can narrow the Gaps:
Ten Steps To Prosperity:
2. Federally funded Medicare — parts a, b & d, plus long-term care — for everyone
3. Provide a monthly economic bonus to every man, woman and child in America (similar to social security for all)
4. Free education (including post-grad) for everyone
5. Salary for attending school
6. Eliminate federal taxes on business
7. Increase the standard income tax deduction, annually.
8. Tax the very rich (the “.1%”) more, with higher progressive tax rates on all forms of income.
9. Federal ownership of all banks
10. Increase federal spending on the myriad initiatives that benefit America’s 99.9%
The Ten Steps will grow the economy, and narrow the income/wealth/power Gap between the rich and you.
2 thoughts on “The surprising reason why pharmacology is so difficult, really almost impossible.”
Great post, Rodger. Not your usual fare; it’s a very nice change of pace, not to mention an important topic.
BTW, on my browser (Opera w/Windows 7 Pro) the random walk illustration is repeating the some pattern rather than showing all five.
I do have a theory about why the optical illusion works. It seems to be related to residual images in the brain as the picture moves. In a sense, your brain is not processing the image as fast as it’s changing. The previous position of the lines is sorta combining with the new position to make it appear as though there is movement of the lines. Maybe.
Understanding the way that optical illusions work has been an interest of mine for at least 50 years. It’s fascinating how the brain can be fooled, as can I. I still haven’t figured out what the evolutionary advantage is.
John, I suspect the evolutionary advantage relates to reaction time. See –> React is how we survive. Generally better than See –> Think about it –> React.