3 ms·

> "People who can code in biology, medicine, government, sociology, physics, history, and mathematics are respected and can do amazing things to advance those d

by amirmc 15y ago

> "People who can code in biology, medicine, government, sociology, physics, history, and mathematics are respected and can do amazing things to advance those disciplines" True. I have a theoretical physics friend who's now applying a ton of machine-learning techniques to biophysics stuff (in an academic setting). I've no idea exactly what he's doing other than he's always snowed under and people keep coming back to him for more.

sdfjkl 15y ago
My favorite example of this is OsiriX[1], a GPL'd medical imaging (DICOM) viewer written by a radiologist who couldn't find existing software that did what he wanted. It's now an established part of the DICOM world. [1] http://www.osirix-viewer.com/Snapshots.html http://www.osirix-viewer.com/Snapshots.html
- samuel 15y ago
  I have known people who started as programmers and became highly valued domain experts, so it works in the other way, too.
zedshaw 15y ago
Now that is interesting. I'm trying to imagine how you would apply machine learning to biology. Related to DNA sequences probably?
- crntaylor 15y ago
  Here's just one possibility. Say you want to do genetic testing for susceptibility to a particular trait, for example heart disease. The human genome is vast, and no human could pick out significant patterns relating to heart disease even with years of experience and specialist training. However, this is precisely the task at which machine learning algorithms excel. You split your data into two sets: people who suffered heart disease by time X in their lives, and people who didn't. You then apply a machine learning technique (some kind of classifier - maybe a logistic regression, support vector machine or a decision tree) to the data set, and it picks out features (i.e. particular gene sequences) which are most strongly predictive of heart disease. If you've chosen the right technique, you can not only predict a binary yes/no response, but you can predict a probability between 0% and 100%. Now you can do genetic testing on an individual who hasn't suffered heart disease yet - with a suitable portion of their genome, you feed it into your classifier and it spits out the probability of them developing heart disease by stage X in their life. Of course, the inputs don't have to be genomes (and quite commonly aren't). You might consider someone's age, calorie intake, level of exercise, social class etc. all as valid inputs to this kind of algorithm.

by amirmc 15y ago

sdfjkl 15y ago

My favorite example of this is OsiriX[1], a GPL'd medical imaging (DICOM) viewer written by a radiologist who couldn't find existing software that did what he wanted. It's now an established part of the DICOM world. [1] http://www.osirix-viewer.com/Snapshots.html http://www.osirix-viewer.com/Snapshots.html

samuel 15y ago
I have known people who started as programmers and became highly valued domain experts, so it works in the other way, too.

zedshaw 15y ago

Now that is interesting. I'm trying to imagine how you would apply machine learning to biology. Related to DNA sequences probably?

crntaylor 15y ago
Here's just one possibility. Say you want to do genetic testing for susceptibility to a particular trait, for example heart disease. The human genome is vast, and no human could pick out significant patterns relating to heart disease even with years of experience and specialist training. However, this is precisely the task at which machine learning algorithms excel. You split your data into two sets: people who suffered heart disease by time X in their lives, and people who didn't. You then apply a machine learning technique (some kind of classifier - maybe a logistic regression, support vector machine or a decision tree) to the data set, and it picks out features (i.e. particular gene sequences) which are most strongly predictive of heart disease. If you've chosen the right technique, you can not only predict a binary yes/no response, but you can predict a probability between 0% and 100%. Now you can do genetic testing on an individual who hasn't suffered heart disease yet - with a suitable portion of their genome, you feed it into your classifier and it spits out the probability of them developing heart disease by stage X in their life. Of course, the inputs don't have to be genomes (and quite commonly aren't). You might consider someone's age, calorie intake, level of exercise, social class etc. all as valid inputs to this kind of algorithm.