Eva-Maria Strauch; Steffen M Bernard; David La; Alan J Bohn; Peter S Lee; Caitlin E Anderson; Travis Nieusma; Carly A Holstein; Natalie K Garcia; Kathryn A Hooper; Rashmi Ravichandran; Jorgen W Nelson; William Sheffler; Jesse D Bloom; Kelly K Lee; Andrew B Ward; Paul Yager; Deborah H Fuller; Ian A Wilson; David Baker
2017-06-12 (online)
Nature Biotechnology (Nat. Biotechnol.). 35, 7, 667-671. doi:10.1038/nbt.3907
Uses computation to design an antibody for influenza A
Avanti Shrikumar; Peyton Greenside; Anshul Kundaje
2017-04-10 (online)
Decompose ouput predictions
machine-learning deep-learning
Alex Graves; Greg Wayne; Malcolm Reynolds; Tim Harley; Ivo Danihelka; Agnieszka Grabska-Barwińska; Sergio Gómez Colmenarejo; Edward Grefenstette; Tiago Ramalho; John Agapiou; Adrià Puigdomènech Badia; Karl Moritz Hermann; Yori Zwols; Georg Ostrovski; Adam Cain; Helen King; Christopher Summerfield; Phil Blunsom; Koray Kavukcuoglu; Demis Hassabis
2016-10-12 (online)
Nature (Nature). 538, 7626, 471-476. doi:10.1038/nature20101
Augment deep networks with an external memory (RAM) matrix.
Bart says: "TL;DR: This work follows a line of research that teaches deep-nets to learn algorithmic tasks (addition, sorting, multiplication, key-value look-up). This paper goes a bit further and teaches their network to do shortest-path finding in graphs and demonstrates on maps of the London underground. Cool demo with nice results, but the hype-machine has blown it out of proportion (check out the FT article for a breathless take claiming thinking computers are one step closer...)"
machine-learning deep-learning
Rafael Gómez-Bombarelli; David Duvenaud; José Miguel Hernández-Lobato; Jorge Aguilera-Iparraguirre; Timothy D. Hirzel; Ryan P. Adams; Alán Aspuru-Guzik
2016-10-07 (online)
The authors train an auto-encoder to provide a vector representation for small molecules. Small molecules are graphs with varying sizes, so they're hard to feed into neural nets (which require fixed-length bitvectors). By fusing together an encoder and decoder (and making the "middle" representation sufficiently small), they learn a vector representation.
The authors lean heavily on arxiv:1511.06349 (ref. 25) to autoencode SMILES strings.
They use a variational autoencoder (noisy) to avoid "dead zones" in latent space.
They optomize OLED properties as an example.
Show References| Num | Entry | Why |
|---|---|---|
| 25 | arxiv:1511.06349 |
machine-learning cheminformatics misc
Kathryn M. Hart; Chris M. W. Ho; Supratik Dutta; Michael L. Gross; Gregory R. Bowman
2016-10-06 (online)
Nature Communications (Nat. Commun.). 7, 12965. doi:10.1038/ncomms12965
Labmate summarizes:
They generated ensembles using MD, then docked to those ensembles, then re-weighted the docking scores based on the MSM. This gave a huge improvement in the predictive power of docking to predict affinity/potency. It turned an inverse relationship (when docking using xtal structures) into a highly correlated trend.
They confirmed their hypothesis about the protein flexibility by using a mass spec. method.
They identified a loop movement important in the anti-antibacterial activity of the enzyme that was different from one previously proposed/suspected.
They proposed mutants that would stabilize their proposed loop, and tested them experimentally.
The power of using the MSM to re-weight other analyses is also very encouraging to see yet again. Also note that they did all this with what looks like a pretty low amount of aggregate sampling (few microseconds per mutant).
misc
Stefano Martiniani; K. Julian Schrenk; Jacob D. Stevenson; David J. Wales; Daan Frenkel
2016-09-15 (online)
Physical Review E (Phys. Rev. E). 94, 3, doi:10.1103/PhysRevE.94.031301
Use multistate benett acceptance (MBAR) to find volumes in high dimensions.
misc
Nicholas Guttenberg; Martin Biehl; Ryota Kanai
2016-09-01 (online)
Somehow uses deep networks to extract slow modes from dynamical signals.
machine-learning deep-learning
Samuel R. Bowman; Luke Vilnis; Oriol Vinyals; Andrew M. Dai; Rafal Jozefowicz; Samy Bengio
2015-11-19 (online)
Advances in autoencoding text, used by 2016-aspuru-mol-feat.
misc
Yann Dauphin; Razvan Pascanu; Caglar Gulcehre; Kyunghyun Cho; Surya Ganguli; Yoshua Bengio
2014-06-10 (online)
Labmate summarizes:
This one is a really cool paper. One of those "we've all been doing it wrong" papers that could have a big impact. Their main conclusions are
1. When optimizing functions in high dimensional spaces, saddle points are a much bigger problem than local minima. There are far more of them, and the few local minima that do exist mostly have values only slightly worse than the global minimum.
2. Standard optimization methods deal really badly with saddle points (and hence work really badly in high dimensional spaces). First order methods like gradient descent start taking tiny steps, so they take a really long time to escape. Quasi-Newton methods are even worse. They just converge to the saddle point and never escape.
3. They describe a new approach that doesn't have these problems and goes right through saddle points without slowing down.
They do all this in the context of neural networks, but it likely applies just as well to other high dimensional optimization problems. Proteins, for example. When you use an algorithm like L-BFGS for energy minimization, it's probably converging to a saddle point, not a local minimum. It could be really interesting to try their method. Could we fold a protein to the native state just by a straightforward energy minimization?
Force field optimization is another case whether this approach could be really useful.
They also show that at a saddle point, there's a strong monotonic relationship between the error and the fraction of negative eigenvalues of the Hessian. Potentially that could be used as a way to measure how far you are from the global minimum. For example, when optimizing force field parameters, it would tell you whether your parameters are close to optimal, or whether there's still a lot of room to improve them further.
misc machine-learning deep-learning