Why, yes, pattern recognition is what ML really is and should be used for.
Aside from pattern recognition everything else is a new astrology, with estimated (straight from one's ass) probabilities instead of planets and constellations.
Scientists use ML to find an antibiotic able to kill superbugs in mice
41–50 of 93 posts
Re: Scientists use ML to find an antibiotic able to kill superbugs in mice
#42A lot of the more cynical commenters here are misunderstanding the breakthroughs that have lead to the novel discovery, and underpin the emerging ML revolution that we are just beginning to witness. Yes, partly the results of this study are due to increased compute, but I'd say that's only about 50% of the secret sauce. The other 50% is attributable to many very recent developments in the field ML which are gradually coming together in solutions to a range of problems - deep learning, convolutional networks, new architectures (GAN, transformer, RNN, LSTM, autoencoders, etc), regularization techniques, gradient control, hyperparameter optimization...the list is quite long, and every SOTA neural network inevitably incorporates a large proportion of these small steps towards the massive leaps that are being taken in the applied ML space. The concepts of perceptrons and gradient descent may be 50+ years old, but an entire body of theory has been explosively developed in the last decade, such that comparing modern ML to what was being done just a decade ago is akin to the difference between, say, programming theory now and 100 years ago. And for the same reason - just as compute advances enabled development of computer science, so to have recent hardware advances unlocked a new level of machine learning.
But I'd like to point out that, even moreso than software development, very little of the grand breakthroughs we will soon see will be possible without multidisciplinary domain knowledge. It is very difficult to effectively apply ML without a solid technical understanding of the properties of the applied data space, which for real world applications are constrained by physical laws and represented and communicated best by mathematical descriptions. ML engineering is a generalist's game - and what we are going to find is that the most successful ML engineers come from broadly applicable, math heavy backgrounds - physics in particular, electrical engineering, to a lesser degree mathematics, etc - because ultimately training a neural network comes down to adequately sampling a problem space and curating data with an intuition which is most ideally developed by the study of mathematics. It is a very general view of the world which is difficult to communicate to someone who is not experienced with higher math.
The current wave of applied ML startups will see a high rate of failure - because ML is still being treated as an extension of programming, in the sense that you expect to be able to hire a bunch of pure developers to translate a specialist's knowledge into code. But this emerging field is different, the few startups that succeed in the applied ML space will be those that are able to find the rare domain experts who have picked up ML along with their math and science experience. There will effectively emerge two classes of ML engineers with substantially different levels of compensation - the coveted generalists who have cross-pollinated with math heavy disciplines, and the rest.
Re: Scientists use ML to find an antibiotic able to kill superbugs in mice
#43I'm not an expert in ML/Biology, but I wonder if doing so won't completely eliminate all of humankind's diseases, but shift the battle from one between humans and bacteria/viruses, to one where ML takes the place of humans by proxy (say we let this ML vs superbugs play out over centuries). I wonder to what direction evolutionary pressure in the face of ML would take bacteria. Perhaps a super smart bug.
Thats a very interesting thought. I am not sure evolution can compete with artificial evolution ( ML ). But we never know :). Fingers crossed that ML can always beat evolution.
Re: Scientists use ML to find an antibiotic able to kill superbugs in mice
#44This paper shows what can be done when you carefully run an ML program alongside a wet lab experimental program tailored to feed back into the ML program. The results end up far more interesting than some recent "ML aided drug discovery" papers in that they actually discovered a drug that functions very differently than known antibiotics (AB). Even though the structures that came out look AB-like, they work different…
Not just in drug discovery but in most interesting industries. Using ML as either a human/cyborg aid or ML+real world ground truth is a secret superpower that I'm surprised more people don't know about. I'm glad they don't.
One challenge here is seen in the types of articles researhers in comp sci have published over the years. I believe that there has been a massive decline in case studies and field trials of software and a rise in analytic and empirical algorithm development. To unlock the super power for people we will need more good hci and the wheel will have to turn in research practice once more.
Re: Scientists use ML to find an antibiotic able to kill superbugs in mice
#45I'm not an expert in ML/Biology, but I wonder if doing so won't completely eliminate all of humankind's diseases, but shift the battle from one between humans and bacteria/viruses, to one where ML takes the place of humans by proxy (say we let this ML vs superbugs play out over centuries). I wonder to what direction evolutionary pressure in the face of ML would take bacteria. Perhaps a super smart bug.
Thats a very interesting thought. I am not sure evolution can compete with artificial evolution ( ML ). But we never know :). Fingers crossed that ML can always beat evolution.
That's also a very interesting thought. The general narrative with general AI is that natural evolution in humans is too slow, we won't be able to catch up mentally with such an AI and this is a bad thing. But the flip side is that natural evolution that is damaging to us (bacteria and viruses) also won't be able to evolve fast enough.
Re: Scientists use ML to find an antibiotic able to kill superbugs in mice
#46Decades ago when I worked in a lab, "high-throughput screening" was definitely on the list of buzzwords. Given what I saw back then, I'm struggling to understand how there could possibly be "a library [..] of 6,111 molecules at various stages of investigation for human diseases" (a.k.a. "Drug Repurposing Hub") which hasn't already been partially or fully screened for interesting antibiotic activity. Could it be there…
For DD, it seems initial screening of as many phages, microbes and compounds as possible using highly-automated brute force might be plenty efficient to test their effectivenesses against every horrible, resistant and opportunistic pathogen for candidate identification. Maybe flying drones out to collect samples in as many random places (public places, restrooms, dirt and even more random places) as possible, generating more samples than a team of humans ever could.
There doesn't seem to be any "One True Way," but a holistic synthesis of collection, identification and selection methods.
Re: Scientists use ML to find an antibiotic able to kill superbugs in mice
#47Decades ago when I worked in a lab, "high-throughput screening" was definitely on the list of buzzwords. Given what I saw back then, I'm struggling to understand how there could possibly be "a library [..] of 6,111 molecules at various stages of investigation for human diseases" (a.k.a. "Drug Repurposing Hub") which hasn't already been partially or fully screened for interesting antibiotic activity. Could it be there…
For DD, it seems initial screening of as many phages, microbes and compounds as possible using highly-automated brute force might be plenty efficient to test their effectivenesses against every horrible, resistant and opportunistic pathogen for candidate identification. Maybe flying drones out to collect samples in as many random places (public places, restrooms, dirt and even more random places) as possible, generat…
> highly-automated brute force might be plenty efficient
Hasn't this been going on in one form or another for many decades?
When I was in this field (20+ years ago) I got to visit labs at Glaxo Wellcome, SmithKline Beecham, Zeneca and so on.
Even back then they were proudly showing off lab robots which allowed them to run large-scale screening experiments.
Not sure any of this stuff is quite as revolutionary as it looks.
Re: Scientists use ML to find an antibiotic able to kill superbugs in mice
#48> That is an especially pressing challenge in the development of new antibiotics, because a lack of economic incentives has caused pharmaceutical companies to pull back from the search for badly needed treatments. Each year in the U.S., drug-resistant bacteria and fungi cause more than 2.8 million infections and 35,000 deaths, with more than a third of fatalities attributable to C. diff, according to the the Centers…
A couple of good recent article to understand this complex problem (USA perspective): https://endpts.com/can-we-make-the-antibiotic-market-great-a... and https://endpts.com/biopharma-has-abandoned-antibiotic-develo...
Reasons are mainly: No incentive to develop antibiotics from a legal perspective (FDA), as insurance companies prefer to reimburse the cheap and generic, still working mostly "well enough" for now.
Insurers pay for in-patient antibiotics as part of a lump sum to hospitals known as a Diagnosis Related Group (DRG). Using a cheap antibiotic increases hospital profit margins, while using an expensive new drug could mean that a hospital might lose money by treating a given patient. As a result, hospitals are incentivized to use cheaper antibiotics whenever possible. This puts significant pricing pressure on new antibiotics, which are one of the only type of medicines paid for like this.
Re: Scientists use ML to find an antibiotic able to kill superbugs in mice
#49Earlier quoted context omitted.
That's probably because Bill & Melinda issued their grants a few years back. You can see a list of all of the high-tech outhouse makers here [0]. Many of the designs are out in the field for long-term testing. [0] https://stepsforsanitation.org/innovation-center/
Or maybe it's because you don't see articles about open defecation posted here all that much. I can find two. One from two years ago and one from six years ago. https://hn.algolia.com/?q=open+defecation If you search on toilet , there's a lot more articles that come up, but at first glance, most don't appear to be about solving open defecation, though there is one on the front page of the search about the Gates found…
There's one more. https://news.ycombinator.com/item?id=20730146 from six months ago is titled "California’s Biggest Cities Confront a ‘Defecation Crisis’" and deals with "excrement on the sidewalks of San Francisco".
It has 159 comments, probably because of how it affects "first world wealthy people".
Re: Scientists use ML to find an antibiotic able to kill superbugs in mice
#50Earlier quoted context omitted.
For DD, it seems initial screening of as many phages, microbes and compounds as possible using highly-automated brute force might be plenty efficient to test their effectivenesses against every horrible, resistant and opportunistic pathogen for candidate identification. Maybe flying drones out to collect samples in as many random places (public places, restrooms, dirt and even more random places) as possible, generat…
> highly-automated brute force might be plenty efficient Hasn't this been going on in one form or another for many decades? When I was in this field (20+ years ago) I got to visit labs at Glaxo Wellcome, SmithKline Beecham, Zeneca and so on. Even back then they were proudly showing off lab robots which allowed them to run large-scale screening experiments. Not sure any of this stuff is quite as revolutionary as it lo…
I did a stint at GSK about 13 years ago. The large scale mechanised experiments also produced big databases of compounds and their properties and mechanisms to make these accessible to researchers and build bigger clusters of systems to run models on. There was a lot of talk of ML but implementation was nowhere near the scale we see nowadays. I think what has changed significantly is the scale, second to that the methods including algorithms and methods to feed back real data into it. It's an evolutionary improvement but I sense it's one of those things where one lucky step in the evolution could make a big difference in productivity.