Why computational predictive toxicology is hard | Hacker News Reader

Why computational predictive toxicology is hard

owlposting.com

24 pointsabhishaike13 comments

> Some drug side effects, like mild nausea or headache, might be considered acceptable trade-offs for therapeutic benefit. But others, like liver failure or birth defects, would be considered unacceptable at any dose. This is particularly true when it comes to environmental chemicals, where the effects may be subtler and the exposure levels more variable. Is a chemical that causes a small decrease in IQ scores toxic? What about one that slightly increases the risk of cancer over a lifetime (20+ years)?

This is an interesting question. For example we know that exposure to traffic pollution reduces fertility and life expectancy, and that seems to be acceptable. For example city centers are usually the most polluted, but can have some of the most expensive square meters to live in. Although it is true that maybe most people are not aware of that.

>For example city centers are usually the most polluted, but can have some of the most expensive square meters to live in. Although it is true that maybe most people are not aware of that.

Or it's a known tradeoff. If you live in the city center you save a lot of time on commuting if your office is there, too. You save a lot of time going out because you can walk over to the next restaurant in 5-10 minutes.

That's measurable time savings in the 1-2h per day range depending on what you compare it to.

If it doesn't decrease your lifetime by a similar amount it may be a worthy tradeoff.

Decreased mortality in order to walk to a food establishment quickly is a worthy tradeoff?

That depends. Do you value sitting in a car or doing literally anything else more? Restaurant probably isn't the best example because it's a luxury. What about the grocery store? Work?

Walking to and from the grocery store carrying a family’s worth of groceries, inhaling exhaust fumes and trying not to get hit by a car, sounds fucking miserable.

Sitting in a car in traffic might not even be that healthy either. And it definitely costs time.

But I don't want to tell you how to live.

Sitting at a restaurant however, no biggie.

Biology is messy, quite hard to fit into neat lock-and-key paradigms.

Furthermore, the data we do collect on relationships between drugs and X (whether fine/coarse grained toxicity, activity, preference for a target, etc.) is disorganized and biased, both intentionally and unintentionally.

I was working on affinity models using ML for a while (whether drug sticks to X). I spent quite some time imagining cool architectures to handle the task, and at one point even thought I might have SOTA on a common benchmark.

It took me a bit to realize that not only was I not SOTA with a proper "hard" split, but that this whole zoo of models coming out - claiming to have an inch over the previous best model - all perform more-or-less the same. The "better-performing" ones often added a lot of bells & whistles and smart theory only to result in no pragmatic edge.

The cherry on top was a study which found that a common benchmark was so biased that models perform eerily well even when you remove the drug or the target from the datapoint. Yup, your model could predict relatively well whether Mike and Alice like each other if you only show it Mike (or Alice) at training and inference. The exact reference evades me, sadly.

This is all to say: The space of interactions between drugs and targets/cells/tissues/organisms is so sparsely explored, that "foundation models" of this sort still seem to me a thing of science fiction. They're that far out.

If we're to make groundbreaking discoveries, my best guess is it would be in very restricted problems, as opposed to applying a general solution to a particular case.

As Paracelsus already wrote, "All things are poison, and nothing is without poison; the dosage alone makes it so a thing is not a poison." (https://en.wikipedia.org/wiki/The_dose_makes_the_poison) That might be one of the reasons why predictive toxicology is hard?

All chemicals, not all things. Things that physically disrupt life systems are not poisonous despite being deadly. Concrete can kill but isn't really a poison. Drink enough of it and it will physically disrupt your digestive tract. Inject it into the space between your brain and skull and it will kill. But so too will cheese. Whether something is a poison also depends on circumstance. If one eats dry ice (a thing that does happen in the real world) then one is not being "positioned" by CO2. Your ruptured stomach is a physical injury. But a buildup of CO2 in the blood, that is CO2 as a poison.

Water is an interesting example. Strictly speaking, it does not act chemically. It disrupts chemistry. It physically prevents other chemicals from acting. But water always ends up being a special case.

Well, pure water causes physical damage to the cells by changing osmotic pressure (now, is that distinct from toxicity in your view?).

Or asbestos, which causes physical damage to cells which in turn may case cancer that leads to death. There are no bright rules. As a practical matter, I personally would draw the line between physical actions that are perceptible on a human scale (concrete blocking up the digestive tract, water blocking the lungs of drowning people) and "poisons" that operate on a smaller scale that we cannot see. Such an approach would include asbestos as a poison despite it doing only physical damage.

  All chemicals, not all things.

While you have a point on physical injury, all things are chemicals and your water example is really bad. Dissolution and ions transport are chemicals process, so are dilution and concentration.

Reply on news.ycombinator.com