Is there any research on the application of AI for drug design?

Question

Is there any research on the application of AI for drug design?

For example, you could train a deep learning model about current compounds, substances, structures, and their products and chemical reactions from the existing dataset (basically what produces what). Then you give the task to find how to create a gold or silver from the group of available substances. Then the algorithm will find the chemical reactions (successfully predicting a new one that wasn't in the dataset) and gives the results. Maybe the gold is not a good example, but the practical scenario would be the creation/design of drugs that are cheaper to create by using much simpler processes or synthesizing some substances for the first time for drug industries.

Was there any successful research attempting to achieve that using deep learning algorithms?

Answer 1

Yes, many people have worked on this sort of thing, due to its obvious industrial applications (most of the ones I'm familiar with are in the pharmaceutical industry). Here's a paper from 2013 that claims good results; following the trail of papers that cited it will likely give you more recent work.

Answer 2

Yes, there were successful attempts at predicting the interaction between molecules and biological proteins which have been used to identify potential treatments by using convolutional neural networks.

For example in 2015, the first deep learning neural network has been created for structure-based drug design which trains 3-dimensional representation of chemical interactions which works similar to how image recognition works (composing smaller features into larger, complex structures).^wiki

^{Study: AtomNet: A Deep Convolutional Neural Network for Bioactivity Prediction in Structure-based Drug Discovery}

---

Another approach is to use evolutionary artificial neural networks which can achieve great optimization results.

Furthermore, the paper from 2015 demonstrated heuristic chemical reaction optimization (CRO) which is inspired by the nature of chemical reactions (e.g. transforming the unstable substances into stable ones). Simulation results show that CRO outperforms many evolutionary algorithms by population-based metaheuristics mimicking the transition of molecules and their interactions in a chemical reaction.

Sample pseudocode algorithm for predicting synthesis given ω1, ω2 (from the above paper):

 1: for all Matrices and vectors m in ω′ do
 2:     for all Elements e in m do
 3:         Generate a real r between 0 and 1
 4:         if r > 0.5 then
 5:             e =counterpart in m1
 6:         else
 7:             e =counterpart in m2
 8:         end if
 9:     end for
10: end for

which is used to generate a new solution ω′ based on two given solutions ω1 and ω2.

Answer 3

No one mentioned Planning chemical syntheses with deep neural networks and symbolic AI (published in Nature - here's arxiv link). Very impressive application of deep reinforcement learning - they use Monte Carlo Tree Search with a policy network (a-la AlphaZero) to do chemical synthesis planning. Authors claim that double blind test shown that professional chemists cannot distinguish between human- and AI-generated synthesis pathways.

Speaking of Alpha* stuff - AlphaFold is a quite recent result in protein folding, which shown breakthrough-level performance compared to all the competition.