5
$\begingroup$

I have recently read many papers where neural networks (NNs) are trained to predict chemical properties (starting from the structure of small chemical compounds) for compounds rather close to the ones with which they were trained. In some cases the predictions are rather good. But I see that apart from the predictive capability, people use NNs as "black boxes" in the sense that they do not tell you or give you chemical insight about why that compound has that value for that predicted property. At least that is my impression now. So my question here is, do you know cases where useful chemical information can be extracted from NN predictions?

$\endgroup$
8
  • $\begingroup$ What do you consider "useful"? Usually the prediction is the interesting bit. The ANN is an abstraction, i.e., it contains/is the chemical information gleaned from the training set. $\endgroup$ Mar 1, 2013 at 15:37
  • $\begingroup$ That is not true. Let's say a "perfect" ANN that could predict whether a certain compound will have some value for some property. But, can this ANN explain why? $\endgroup$ Mar 1, 2013 at 15:39
  • $\begingroup$ Newtonian mechanics explains nothing. They are merely a set of rules that predict outcomes. The "why" is in the determinism of those rules. I have an answer for you below, but I fear you really have to state clearly what kind of information you want to take away and at what point you are willing to stop asking why. $\endgroup$ Mar 1, 2013 at 17:01
  • $\begingroup$ I think this is sort of interesting, but it's bordering on being off-topic. I think you can ask a question like "How does my ANN tell pictures of roses and tulips apart?" and the answer would be similar. I'll leave it for now, but if you could expand on how the chemistry ANNs would be different, that would be great. $\endgroup$
    – jonsca
    Mar 2, 2013 at 5:33
  • $\begingroup$ NN and other qsar methods usually works on top of some kind of 'descriptors'. It is possible to find relevant descriptors with NN and then track, what produced them from molecule. The area is not well-developed to my knowledge. $\endgroup$
    – permeakra
    Mar 2, 2013 at 10:00

2 Answers 2

4
$\begingroup$

I did some work on interpreting neural network QSAR models - I won't claim that they explain everything and has a number of limitations (linearizes network connections, applicable only to feed forward networks). But maybe it will be useful

Interpreting Computational Neural Network QSAR Models:  A Measure of Descriptor Importance

Interpreting Computational Neural Network Quantitative Structure−Activity Relationship Models:  A Detailed Interpretation of the Weights and Biases

Of course, if the descriptors going into the network are opaque (as in abstract mathematical properties such as many topological descriptors) then I don't think any interpretation method is going to be very helpful.

$\endgroup$
4
$\begingroup$

Since you are interested in Artificial Neural Networks that also explain their decisions/predictions you may be interested in transparent neural networks, although they have not been used to my knowledge on chemical information.

Other interesting approaches include inductive logic programming which have among other things been applied to predicting binding of proteins to hexose

$\endgroup$
3
  • $\begingroup$ Whilst relevant to neural networking in general, this answer has nothing to do with applications of neural networking to chemistry, per se. $\endgroup$ Mar 2, 2013 at 8:21
  • $\begingroup$ Deathbreat's answer is good. He does not provide examples but gives me insights; I did not know about ILP. So I read about it, and tried to find papers related to ILP and my question so I have found this (biomedcentral.com/1471-2105/13/162), which is directly related to my question. You should not be that strict. $\endgroup$ Mar 2, 2013 at 10:04
  • $\begingroup$ @RichardTerret: The question does not require specificity to chemistry. I'll incorporate the paper flow found in order to make the chemistry link. $\endgroup$ Mar 4, 2013 at 18:32

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.