#compchem #machinelearning people, what's the current best out-of-the-box GNN prediction library for small molecules? trying to use chemprop but finding it impossible
towards_entropy retweeted
New paper: an end-to-end approach for learning to simulate coarse-grained molecular dynamics!
- Multi-scale GNN
- complex systems: polymers and batteries
- very long stable rollouts
- orders of magnitude faster than classical force fields!
arxiv.org/abs/2204.10348
🧵1/n
towards_entropy retweeted
Today we release MRL, an open source python package for #drugdiscovery built on @RDKit_org, @PyTorch and #nbdev. MRL combines generative models, reinforcement learning and #cheminformatics to accelerate drug design. Stay tuned for more!
github.com/DarkMatterAI/mrl
Are there any code samples from the SOM networks from the Deep Learning/DDR1 Kinase paper (nature.com/articles/s41587-0…)? The SOM networks aren't included in the author's shared code
What's the current status of the Linearity hypothesis for adversarial examples? Trying to reconcile some contradictory information.
However, there's this from Deepmind which presents contradictory information arxiv.org/pdf/1811.09716.pdf
The authors find that robustness via adversarial training is achieved by decreasing the curvature of the loss landscape, effectively making the network more linear
1
2
Has anyone followed up on this? Is linearity at the logit level good or bad? If it's good, what explains the Church window plots from
pdfs.semanticscholar.org/b5e…
Mean squared error is derived from maximum likelihood assuming p(y|x) is Gaussian. Are there derivations for loss functions following the same procedure using a different assumed distribution? Specifically something fat tailed, like a Levy distribution
Following up on some work by @vijaypande's group. Retrosynthesis reaction prediction top-1 accuracy boosted to 62% (previously 54.6%) using transformers and SMILES based data augmentation
github.com/kheyer/Retrosynth…
2
Are there any publications on neural architecture search for LSTM/GRU cell variants? Or really any recent exploration on recurrent cell design
I'm aware of Greff (arxiv.org/pdf/1503.04069.pdf) and Jozefowicz (proceedings.mlr.press/v37/jo…) on the topic, but both papers are from 2015. It feels like the tools/compute available have evolved enough since then that a more expansive investigation should be possible
When the paper says "details can be found in the Supplementary Information" but there is no supplementary information







