No, because cancer always evolves. If you don't kill 100% of cancer cells, the remaining cells are likely resistant to whatever you used to kill the specific cell type, and they will grow up and continue causing problems.
What you would need is something that recognized cancer cells and only cancer cells with 100% specificity/100% selectivity, and killed all of them. This is near impossible for something that can evolve.
It's an arms race, but we can in principle stay mostly ahead in it. Evolution is not a perfect solution finder, and in the future, we may be able to constrain it by reducing the probability of a mutation we can't handle to arbitrarily low values. We know how the optimization process works here, what we lack today is enough control over the problem space to screw that process up.
Arms races are a great way to bankrupt yourself. as for "Evolution is not a perfect solution finder, and in the future, we may be able to constrain it by reducing the probability of a mutation we can't handle to arbitrarily low values", that's a nice speculation, however, it's just a speculation, and not one well-founded in fact.
I heartily encourage you to look at the current state of antibiotics and appreciate that while evolution may not be "smart", it has more advantages that just numbers.
As for cancer, it's not clear that more effective cancer will evolve in the near term if we don't apply selection pressure to it- cancer has existed in humans with little to no change for millions of years and in mammals for far longer than that.
It's great that we have all these research teams tackling the cancer problem from a machine-learning/genomics/personalised-medicine angle...
But do they have a common/open standard for technologies (ML, CV, etc), sharing datasets and findings, trials, etc? IMO something like HL7 [0] but for cancer research would be a game-changer.
I see a lot of promising approaches but not so good coordination, with everybody working in silos and developing their own framework... or maybe it's what I see on the surface.
There are. Data types are pretty much standardized on the basic level (eg. fastq, bam, bed, vcf etc.). Else, most is done by the big science consortia like TCGA, ICGC & PanCancer. And obviously the big databases from NCBI and EBI-EMBL. Trials are standardized by EMA and FDA.
FASTQ, BAM, BED, VCF, etc are all standardized in terms of file format, but the semantics of the content are not. Harmonizing data across multiple semantics is still a very challenging problem that few people take seriously.
Like the good old "chr1" == "1" problem? Well, yeah that’s true. But still I think we have those consortia in place, why do we need another authority. They just have to set (and somewhat enforce) the rules.
It goes far beyond "chr1" = "1". There are many other fields with next-to-impossible to interpret data. The consortia have, at best, advisory roles, they cannot enforce any rules. And given an extensible file format, third parties are going to extend it with poorly defined/standardized data. Then the consumers have to harmonize the data. Data harmonization (or cleaning) it typically one of the bottleneck steps in applying machine learning at scale.
I strongly suggest reading "The Biology of Cancer" by Weinberg to get a better understanding of why we will likely not understand cancer so completely we can render it a minor disease.
There is a very interesting dynamic at work in medical research strategy where the scientific urge to completely map our biochemistry meets the engineering urge to make tangible progress in repairing things when they go wrong.
The truth of the matter is that most of the mapping of molecular biochemistry being done in this age of enormous datasets and genetics is of very little immediate application. Meanwhile paths that can use what we already know to generate enormous impact on disease get little attention. This is an age of personalized medicine as the mainstream, but really the overwhelming majority of serious disease mechanisms, those that are age-related or infectious, are exactly the same for everyone. Personalized medicine is a wondrous machine for making money by the look of it, but won't deliver on its promises.
For example in cancer the best path forward is to turn off telomerase and ALT, preventing telomere lengthening. That will work for all cancers, one type of treatment for every cancer type, and will cost no more to bring to fruition than any one of the recent examples of cancer therapies brought to market. But that gets a fraction of the attention that goes towards mapping every last part of the genetics and cellular biochemistry of cancer. The efforts to create therapies that spin off from that type of mapping work are largely very limited, in that they involve targeting a mechanism or marker very specific to only one or a few of the hundreds of types of cancer, and even those types are capable of evolving away from that marker or mechanism when it is successfully targeted.
Cancer is hard because the research and development community largely follows a terrible high level strategy for implementation of therapies, even though in the long term the mapping strategy is exactly what scientists should be doing.
You turn off telomerase, your HSCs stop proliferating, you have no blood cells and game over (a bit exaggerated, but not too much). Cancer is not hard because scientists are stupid, it is hard because it is highly heterogenous and very difficult to even determine common ground, expensive so that study cohorts are allows on the lower end of what is necessary and can be hard to interpret as you have only a single attempt with every patient, mouse model results can be nontransferable to humans etc. We will eventually get it but we will consider it solved after it has been solved for 5 years, not before we even reached a mere 90% cure rate.
Hats off to the new Microsoft for a worthy endeavour.
I just hope that the eventual solution is cheap for those that need it and that the tools and techniques are readily available to the hacker community (i.e. open sourced).. otherwise I'll put my hat back on.
For those not that deeply involved into this: There is the whole field of bioinformatics that is involved in this kind of research, creating data in the range of multiple peta-bytes, using some of the biggest computing clusters around. That a big enterprise like Microsoft wants their part of the cake is obvious, especially as this is going to be a billion dollar market. But there is no reason to especially belief in Microsoft, they are just doing what everyone else is doing (and only with more money and/or expertise if they spent more than a few billion dollar on it, else this is mere marketing, nothing really note worthy). Maybe its a nice touch for the field that now even Microsoft has noticed it.
Indeed there's a whole field, and I think Microsoft has a strong potential to radically improve it.
Most bioinformaticians I know are poor developers, have limited knowledge of existing algorithms, and/or have critically limited bio knowledge. As a result you get buggy, slow and unmaintainable software (see: all the MRI analysis software posts recently), reinvention of classical algorithms (anything from basic statistical tests to full on clustering methods) or tools lacking the full scope required to actually be useful.
I think MS could change this by having strong CS/SWEs with longer tenures than you see in academia, backed by more money and essentially unlimited computing resources, with a high profile name to win them collaborators who can share good datasets. (Compared with Google, they're not trying to generate data in-house as far as I've heard.) Some of the most notable advances in bioinformatics have come from applying decades-old algorithms that never made it to bio before then. Even if MS just accelerates this by applying people who know about these algorithms, that would be very helpful.
Cancer cannot only be solved with algorithms and machine learning. Remember that it's biology, not informatics. A lot of smart people have been working on cancer for decades. It's unlikely MS will 'solve' cancer.
With all due respects, the claim of the article is overblown. Those are the kind of articles a very good department would publish, nothing special or game changing.
Derek Lowe's take, as usual, is a good counterpoint and a giant "Take With Lump Of Salt" NB.
>I have beaten on this theme many times on the blog, so for those who haven’t heard me rant on the subject, let me refer you to this post and the links in it. Put shortly – and these sorts of stories tend to put actual oncology researchers in a pretty short mood – the cell/computer analogy is too facile to be useful. And that goes, with chocolate sprinkles on it, for all the subsidiary analogies, such as DNA/source code, disease/bug, etc. One one level, these things do sort of fit, but it’s not a level that you can get much use out of. DNA is much, much messier than any usable code ever written, and it’s messier on several different levels and in a lot of different ways. These (which include the complications of transcriptional regulation, post-transcriptional modification, epigenetic factors, repair mechanisms and mutation rates, and much, much, more), have no good analogies (especially when taken together) in coding. And these DNA-level concerns are only the beginning! That’s where you start working on an actual therapy; that’s what we call “Target ID”, and it’s way, way back in the process of finding a drug. So many complications await you after that – you can easily spend your entire working life on them, and many of us have.
And that’s why many of us who have actually been working on diseases like cancer get a little testy when we see folks from computer science coming in with this “Gosh darn it fellows, do I have to do everything myself?” attitude. Years of working with (human-designed) hardware running (human-written) code have given many people in that field what I think is an exaggerated idea of human capabilities (at least as they are right now). When you can write code that gets used by hundreds of millions of people in their daily lives, it’s understandable to think that you’re able to just reach in and change reality by sheer braininess and force of will. Unfortunately the world of code and computational hardware, as important, useful, and lucrative as it is, is just a sandbox compared to the real physical universe, of which living creatures are just a tiny little part. But biology has no debugging programs, no annotations, no manuals. It wasn’t written by humans – in fact, as far as we know, it wasn’t written by anyone at all, it “just grew” in a process that has no good counterpart to the ways that humans generally get things done.
....
>If you remove the hubris from the Microsoft announcement, though, which takes sandblasters and water cannons, you get to something that could be interesting. It’s another machine learning approach to biology, from what I can make out, and I’m not opposed in principle to that sort of thing at all. It has to be approached with caution, though, because any application of machine learning to the biology literature has to take into account that a good percentage of that literature is crap, and that negative results (which have great value for these systems) are grievously underrepresented in it as well. I think that machine approaches to understanding biological pathways will, in the end, probably be the way to go, because it’s too complex for us to keep it all straight in our minds (not human!) But we’re not there. There are many, many important things that we simply don’t understand very well, and many others, I’m sure, that we just flat-out don’t even know exist yet. Debug that.
So if Microsoft wants to apply machine learning to cancer biology, I’m all for it. But they should just go and try it and report back when something interesting comes out of it, rather than beginning by making a big noise in t...
“We’re in a revolution with respect to cancer treatment”
Reading this quote from the original article make me nauseous. We're in a revolution with respect to cancer RESEARCH. PEOPLE are the ones being treated. You can't only focus on a math equation. You also need to focus both on palliative care and explainable research. The article seemed to save that acknowledgement for the last two sentences.
My father has been an oncology physician/researcher for 35 years. He regularly has to tell to some of his non-physician researchers that their work needs to be explainable to both his patients and those who will continue funding the research, yet they consistently show up with black box neural networks saying "we found it!"... found what exactly? He has to be able to convince his master-of-the-universe patients that he knows more than WebMD or their favorite holistic blog, see: Steve Jobs. It's incredibly more common than one would hope, see: anti-vax.
Just to clarify, there is nothing wrong with ML research into cancer. However, research isn't treatment. Oncology is depressing. Fewer and fewer doctors are interested in specializing in it. The overall suicide rate of physicians is double the general population, and that's not even oncology-specific. There's a revolution with respect to cancer treatment, and it's a pretty scary downward spiral.
Announcements such as these can end up taking money AWAY from cancer treatment. The government sees this stuff and jumps on the bandwagon. "Why fund humans when computers can do it?" Government research grants pay my father's salary along with all of his physician-researchers and students. Without that funding, people just get to read about the upcoming research while they wait to die.
The last two projects mentioned in the article (Literome and better interpretation of medical images) sound like they may be useful. The projects that are described as cell-debugging and cell-reprogramming sound like the sort of hubristic projects that someone with a good computing background but less than an undergraduate knowledge of cell biology might propose (at least as the projects are described in the article).
About a decade ago I was involved with a US Department of Energy effort to model radiation damage to DNA in silico, and more peripherally involved with a proteomics effort working out of the same institution. They didn't work very well. The number one problem was that we didn't have good baseline data from experiments to work with. IMO, you'd get better mileage out of automating in vitro experiments and ensuring that the data is reproducible before you start trying to leverage the latest in computer science. How much can you machine-learn from garbage inputs? Biology experiments even at the isolated cell level are notoriously fiddly. Let the biologists (or at least people who have worked with them) explain the problems that need solving in biology, and then bring in CS people as required. Starting with a computing-oriented perspective is just begging to misunderstand what biological research needs.
40 comments
[ 4.4 ms ] story [ 90.5 ms ] threadThis has been picked up by the media (how I heard about it) - but with the usual hyperbole, e.g.,
- Microsoft will 'solve' cancer within 10 years by 'reprogramming' diseased cells: http://www.telegraph.co.uk/science/2016/09/20/microsoft-will...
- Microsoft is reprogramming cancer: http://www.pcauthority.com.au/News/437929,microsoft-is-repro...
- Microsoft Shortly to Invent Electric Car, and Probably Outsell Tesla by a Factor of 3
- Microsoft Only Two Years Away from Composing a Song So Beautiful it Revolutionizes Music Industry
- Oracle launches cloud service, predicts Amazon's demise
(joke)
What you would need is something that recognized cancer cells and only cancer cells with 100% specificity/100% selectivity, and killed all of them. This is near impossible for something that can evolve.
As for the speculation part, it kind of follows from the math of things. Evolution is very dumb compared to us, it just has the numbers advantage.
As for cancer, it's not clear that more effective cancer will evolve in the near term if we don't apply selection pressure to it- cancer has existed in humans with little to no change for millions of years and in mammals for far longer than that.
But do they have a common/open standard for technologies (ML, CV, etc), sharing datasets and findings, trials, etc? IMO something like HL7 [0] but for cancer research would be a game-changer.
I see a lot of promising approaches but not so good coordination, with everybody working in silos and developing their own framework... or maybe it's what I see on the surface.
[0] http://www.hl7.org/
I take it to mean that we will understand cancer so completely that we can render it a minor disease.
Oh man, I had to call in sick last week 'cause I had a nasty cancer, but I'm feeling fine now.
The truth of the matter is that most of the mapping of molecular biochemistry being done in this age of enormous datasets and genetics is of very little immediate application. Meanwhile paths that can use what we already know to generate enormous impact on disease get little attention. This is an age of personalized medicine as the mainstream, but really the overwhelming majority of serious disease mechanisms, those that are age-related or infectious, are exactly the same for everyone. Personalized medicine is a wondrous machine for making money by the look of it, but won't deliver on its promises.
For example in cancer the best path forward is to turn off telomerase and ALT, preventing telomere lengthening. That will work for all cancers, one type of treatment for every cancer type, and will cost no more to bring to fruition than any one of the recent examples of cancer therapies brought to market. But that gets a fraction of the attention that goes towards mapping every last part of the genetics and cellular biochemistry of cancer. The efforts to create therapies that spin off from that type of mapping work are largely very limited, in that they involve targeting a mechanism or marker very specific to only one or a few of the hundreds of types of cancer, and even those types are capable of evolving away from that marker or mechanism when it is successfully targeted.
Cancer is hard because the research and development community largely follows a terrible high level strategy for implementation of therapies, even though in the long term the mapping strategy is exactly what scientists should be doing.
It's actually pretty interesting.
I just hope that the eventual solution is cheap for those that need it and that the tools and techniques are readily available to the hacker community (i.e. open sourced).. otherwise I'll put my hat back on.
Most bioinformaticians I know are poor developers, have limited knowledge of existing algorithms, and/or have critically limited bio knowledge. As a result you get buggy, slow and unmaintainable software (see: all the MRI analysis software posts recently), reinvention of classical algorithms (anything from basic statistical tests to full on clustering methods) or tools lacking the full scope required to actually be useful.
I think MS could change this by having strong CS/SWEs with longer tenures than you see in academia, backed by more money and essentially unlimited computing resources, with a high profile name to win them collaborators who can share good datasets. (Compared with Google, they're not trying to generate data in-house as far as I've heard.) Some of the most notable advances in bioinformatics have come from applying decades-old algorithms that never made it to bio before then. Even if MS just accelerates this by applying people who know about these algorithms, that would be very helpful.
Google X is this kind of thing 1000x.
With all due respects, the claim of the article is overblown. Those are the kind of articles a very good department would publish, nothing special or game changing.
Derek Lowe's take, as usual, is a good counterpoint and a giant "Take With Lump Of Salt" NB.
>I have beaten on this theme many times on the blog, so for those who haven’t heard me rant on the subject, let me refer you to this post and the links in it. Put shortly – and these sorts of stories tend to put actual oncology researchers in a pretty short mood – the cell/computer analogy is too facile to be useful. And that goes, with chocolate sprinkles on it, for all the subsidiary analogies, such as DNA/source code, disease/bug, etc. One one level, these things do sort of fit, but it’s not a level that you can get much use out of. DNA is much, much messier than any usable code ever written, and it’s messier on several different levels and in a lot of different ways. These (which include the complications of transcriptional regulation, post-transcriptional modification, epigenetic factors, repair mechanisms and mutation rates, and much, much, more), have no good analogies (especially when taken together) in coding. And these DNA-level concerns are only the beginning! That’s where you start working on an actual therapy; that’s what we call “Target ID”, and it’s way, way back in the process of finding a drug. So many complications await you after that – you can easily spend your entire working life on them, and many of us have.
And that’s why many of us who have actually been working on diseases like cancer get a little testy when we see folks from computer science coming in with this “Gosh darn it fellows, do I have to do everything myself?” attitude. Years of working with (human-designed) hardware running (human-written) code have given many people in that field what I think is an exaggerated idea of human capabilities (at least as they are right now). When you can write code that gets used by hundreds of millions of people in their daily lives, it’s understandable to think that you’re able to just reach in and change reality by sheer braininess and force of will. Unfortunately the world of code and computational hardware, as important, useful, and lucrative as it is, is just a sandbox compared to the real physical universe, of which living creatures are just a tiny little part. But biology has no debugging programs, no annotations, no manuals. It wasn’t written by humans – in fact, as far as we know, it wasn’t written by anyone at all, it “just grew” in a process that has no good counterpart to the ways that humans generally get things done. ....
>If you remove the hubris from the Microsoft announcement, though, which takes sandblasters and water cannons, you get to something that could be interesting. It’s another machine learning approach to biology, from what I can make out, and I’m not opposed in principle to that sort of thing at all. It has to be approached with caution, though, because any application of machine learning to the biology literature has to take into account that a good percentage of that literature is crap, and that negative results (which have great value for these systems) are grievously underrepresented in it as well. I think that machine approaches to understanding biological pathways will, in the end, probably be the way to go, because it’s too complex for us to keep it all straight in our minds (not human!) But we’re not there. There are many, many important things that we simply don’t understand very well, and many others, I’m sure, that we just flat-out don’t even know exist yet. Debug that.
So if Microsoft wants to apply machine learning to cancer biology, I’m all for it. But they should just go and try it and report back when something interesting comes out of it, rather than beginning by making a big noise in t...
Reading this quote from the original article make me nauseous. We're in a revolution with respect to cancer RESEARCH. PEOPLE are the ones being treated. You can't only focus on a math equation. You also need to focus both on palliative care and explainable research. The article seemed to save that acknowledgement for the last two sentences.
My father has been an oncology physician/researcher for 35 years. He regularly has to tell to some of his non-physician researchers that their work needs to be explainable to both his patients and those who will continue funding the research, yet they consistently show up with black box neural networks saying "we found it!"... found what exactly? He has to be able to convince his master-of-the-universe patients that he knows more than WebMD or their favorite holistic blog, see: Steve Jobs. It's incredibly more common than one would hope, see: anti-vax.
Just to clarify, there is nothing wrong with ML research into cancer. However, research isn't treatment. Oncology is depressing. Fewer and fewer doctors are interested in specializing in it. The overall suicide rate of physicians is double the general population, and that's not even oncology-specific. There's a revolution with respect to cancer treatment, and it's a pretty scary downward spiral.
Announcements such as these can end up taking money AWAY from cancer treatment. The government sees this stuff and jumps on the bandwagon. "Why fund humans when computers can do it?" Government research grants pay my father's salary along with all of his physician-researchers and students. Without that funding, people just get to read about the upcoming research while they wait to die.
About a decade ago I was involved with a US Department of Energy effort to model radiation damage to DNA in silico, and more peripherally involved with a proteomics effort working out of the same institution. They didn't work very well. The number one problem was that we didn't have good baseline data from experiments to work with. IMO, you'd get better mileage out of automating in vitro experiments and ensuring that the data is reproducible before you start trying to leverage the latest in computer science. How much can you machine-learn from garbage inputs? Biology experiments even at the isolated cell level are notoriously fiddly. Let the biologists (or at least people who have worked with them) explain the problems that need solving in biology, and then bring in CS people as required. Starting with a computing-oriented perspective is just begging to misunderstand what biological research needs.