Cancer: A Computational Disease that AI Can Cure [video] | Hacker News Reader

Cancer: A Computational Disease that AI Can Cure [video]

videolectures.net

89 pointscaustic26 comments

I'm a bioinformatician at a translational cancer research institution & have some sense of the stumbling blocks and opportunities in this space.

First off, doing away with clinical trials as we perform them now would certainly speed up pairing drugs with patients but it's not going to happen unfortunately. Drug trials aren't designed to minimize search paths, they are designed to minimize risk.

With terminal diseases like many cancers the rules are bent a little but what's important to understand is that these are real people making the decisions ultimately. Sometimes the most information can be gained by withholding drug, and isolating another treatment's impact but if you are a patient with a terminal prognosis, or the doctor trying to treat that patient, are you going to choose not to take something that might help you?

This is a point of frustration for many in research fields because it means clinical data is hugely noisy. Patients are often cycled through different drugs quickly to find something that takes hold, while at the same time going in for as much chemo/radiation treatment as they can bare.

I'm not saying these approaches aren't relevant, they just won't happen as a grand reimagining of our drug approval system. What is starting to happen however, is reclassification of cancer 'type' based on genetic profiles meaning ovarian cancer may have certain genetic similarities to lung or pancreatic so instead of treating melanoma, you can treat a cancer with a disturbed MAP pathway.

All the same, I'm glad he's working on this problem because it's huge, and I look forward to seeing the point of view of the HN community.

edit: Though obvious to some, I should also mention this problem is complicated by the fact that genetic information is private and dissemination is highly restricted. Patients can release this info but it does often inhibit massive cross-patient research.

There are many new sorts of trials coming online that are closer to the model he's proposing here. (Which, BTW, is more like A/B testing than like a classical RCT.) See esp. adaptive trials, point-of-care trials, and global cumulative trails. Some of these new models are really being run by real clinical researchers.

This is a great point and I think many doctors are pushing for new ways to perform clinical trials but adaptive trials generally only change a single parameter during the course of a trial (dosage for example). And POC trials open up the number of patients that may be willing to participate by administering tests at local sites but the state of clinical trials still has a long way to go. Once it's firmly in drug companies' and insurance agencies' best interest to open things up then we may see more rapid change.

"opening things up" is definitely going to help, and is a huge part of the philosophy of my nonprofit science research org (our first project will be taking an anticancer through preclinical). There are, however, a ton ton ton of biological issues that still are, open or not, difficult to surmount, sometimes not even having to do with the cancer itself (bioavailability, side effects, etc).

Interesting discussion. At around 7 minutes in, he postulates that maybe we've already cured 20% of cancers. The problem is that we have cures that fail clinical trials because cancer is so varied and not generic enough to have one cure. One drug might only work in 10% of people while another cure might cure another 10%, and so on.

These drugs all fail clinical trails and are lost.

"One drug might only work in 10% of people while another cure might cure another 10%, and so on."... I believe this, to some degree. But I don't think it's that much. A lot of drugs get killed because of side effects, poor bioavailibility, unexpected responses, interference with other drugs that are common or necessary in the cohort of cancer patients, etc. Those things are much harder to predict. It is embarassingly easy to find compounds that kill cancer cells in vitro.

However, this:

"Cancer results from finite genomic mutations that biotechnology can easily list."

Well the first part is true. The second part, not so much.

Not even sure that the number of ways that a genome can be mutated is even finite. Genomes can be mutated not just by the substitution of single base pairs, but also by introducing copies, through deletions, translocations and insertions.

There is even the phenomenon known as Chromothripsis [1], which has the entire genome smashed into thousands or millions of pieces and then randomly put back together.

[1] http://en.wikipedia.org/wiki/Chromothripsis

Pedantically, it would have to be finite, because there is a size limit. Cancer couldn't transform your genome into a piece of DNA the size of the sun.

Just to out-pedant you, the size of the sun is still quite finite, most (all?) things are.

Perhaps colloquially, finite here is (not un-like literally) a metaphor, in this case for not-unfathomably-large, or more precisely, measurable.

It's not because it's finite that the number of combinations possible is limited. Just like for passwords, the more bits involved, the more possibilities you get, and a DNA or RNA is a molecular version of a very, very long password. So "finite" is a very poor way to describe the complexity.

"Time Travel: A temporal challenge faster than light travel can solve!"

Sorry to be snarky but my point is that it may be plausible that really strong AI could solve the cancer problem but that it doesn't matter.

The point that both AI and cancer are much harder and much different problem than even the optimistic researchers in the fields imagine (why they're still optimistic).

Most of the time cancer cure articles are met here by someone who posts a link describing how two side of even a single tumor will often involve substantially different cells with substantially dynamics. And each of these dynamics is a real challenge to the human (and the challenge varies because each person physiology varies more than is common recognized[1]) . A strong enough AI could be sending drugs separately to each side of the tumor, yes. But the AI would have to figuring out the dynamics of those particular cells. And I mean those particular cells, not the other cells kind of like them in some laboratory (yes, the presenter deals with this issue but I claim it's harder, much harder than he claims).

Basically, there is no AI of sufficient quality to have the flexibility to figure out each person's cancer. There's nothing flexible enough in theory so there's nothing to apply. Modern AI can recognize a lot of patterns and can use fixed rules to solve problems but it's far from the human ability to put all these things together. And our human ability here is, itself, not up to the challenge. You can have a fifty experts treating a single person's cancer and the chance of success won't go that much, etc.

[1] See: http://www.anapsid.org/aboutmk/biochem.html etc

Imagine what NSA's mathematicians (and budget) could do, if they were focused on these tasks, instead of spying!

One could argue that all the brain resources currently used on Wall Street are also wasted, when compared to the magnitude of the problems we're facing in other places.

Hopelessly optimistic. Spoken like someone who has never taken an experimental design course imho.

1) The search space is practically infinite dimensional. All methods suck when trying to extract a causative model from an infinite search space. There's a reason why in experimental design we change as little as possible.

2) SNP's are not the entire story. If it were this simple we would have progressed much further already. See dismal failure of all other high throughput sequencing and microarray technology. We still don't know how to analyse this stuff properly, if it will ever be possible.

3) The metabolome is adaptive! While we each have different enzyme kinetics due to slight differences in protein makeup, overall metabolic flux rates are amazingly consistent. See Oliver Feihn et al for more details.

There's an unsaid implication that it is possible to sub-type cancers based on a patients genetic type. E.g. what makes a particular drug work on a particular cancer form, is determined by that patients genetics. Is this a generally accepted assumption?

This technique is coming into prominence but genomic is still a very young science with the cost of sequencing only recently coming down enough to have it become regular practice. What we have now are essentially a shortlist of known mutations that are seen across many cancers, a huge list of closely associated mutations in the same pathway or impinging pathways that are sometimes seen in conjunction with other mutations, sometimes not and then a ton of noise from which researchers are trying to pick out possible correlation. The vast majority of mutations are not cancer causing, so picking out the ones that are is difficult. That being said, frequently broad characterization can be done. This person has an effected PI3Kinase, or MAP pathway. We only have drugs for a few of the major mutations so once the cancer mutates to avoid drugging, patients are often out of luck.

To answer your second question, cancer is generally accepted as a genetic disease ie mutations, copy number alterations, insertion and deletion of genetic information, however, recent research has shown that epigenetic processes (that is what your cells are doing with your genes such as alternative splicing or DNA methylation) are also an important factor and these can be effected by environmental factors such as diet, exercise, even mood.

it's worth mentioning that just because a mutation is seen across cancers, it may not necessarily help us cure it. Like, for example, if the downstream effect is "increases mutation rate". That's kind of "well I can see how this causes cancer", but targetting it would be too late.

> This person has an effected PI3Kinase, or MAP pathway. We only have drugs for a few of the major mutations so once the cancer mutates to avoid drugging, patients are often out of luck.

Yeah, patients relapsing into cancer usually do not have much alternatives. Combination therapies often help since they can attach several pathways at once, but combo therapies are usually limited to 2 compounds at the same time.

Another huge problem is linked to medical practice. So many doctors treating cancer patients are just NOT aware of the latest studies and developments in terms of Standards of Care. If you have a cancer, the most relevant factor is the necessarily the drug you take but who's treating you and how much he knows about your condition.

Helping doctors make decisions may be another area for disruption, where Software may help.

Very, very impressive. I'm excited about the possibilities! FWIW using machine learning approach to try and individualize treatments and treat people one at a time (linking 'what works' for the individual rather than trying to find the single cure for all) seems very similar to what we tried to do for autism spectrum treatments with http://autism360.org (thought that some on here may find this interesting).

Link to the company he founded: http://www.cancercommons.org

thanks for this. I didn't know about this... I'm looking for an "outside scientific advisor" to audit my anticancer compound research project. I think I'll be contacting these guys.

I recall seeing a fascinating paper which did address some medical malady purely as a programming problem, and offered a viable solution. Made a lot of sense, seemed to work, am mentioning it in hopes somebody might remember it too and find a link (I don't remember enough keywords).

you may be thinking of emerald therapeutics and the work by Brian Frezza.

The future of Cancer therapy could exist today! But why are do we still rely on decades old medical technologies? I cannot say.

Check out this gene therapy called Gendicine approved in China to reverse cancer.

http://scholar.google.com/scholar?q=Gendicine+p53

Apparently, the treatment has been available for years, and yet the product has failed to penetrate to the U.S. It would be funny if it were not true.

Perhaps because there haven't been any wide scale clinical trials that prove it's effectiveness? Of the papers I read through at the given link, most seem to come to conclusion that Gendicide p53 is safe to give to humans and don't comment on whether or not it's an effective therapy, or if it's more effective than existing treatments.

This was invented by onyx pharma in the US years ago. The clinical trials didn't work. Onyx moved onto the greener pasture of small molecule drugs. It seems the viral p53 delivery may work out better with liver cancers, which is a bigger problem in China.

Reply on news.ycombinator.com