2023-06-05 (Mon)


Jana2020 と PETS2 は自由なライセンスでないしソースコードも非公開なので避けていたが、多重散乱を考慮した精密化ができる唯一のソフトなので、一度は体験して実力を評価しておかないと MicroED 従事者として共同研究者に対して不誠実だろうと思い、インストールしてみた。これから時間のあるときに tutorial を少しずつやっていく予定。

悲報: 小学生のときから 25 年以上 Linux 使ってるとりさん、ディレクトリ自身の permission を知りたいときの ls -ld XXX を知らなかった。あるいは過去にも調べたことがあるかもしれないが忘れていた。ずっと、1 つ上のディレクトリで ls -l | grep XXX してた。

Neural-network based CryoEM density modification

I agree with tweets by Sjors: 1, 2, 3, 4, while I strongly disagree with 5 and 6.

If "artifact problems" mean streaky maps, I would not call it an artifact. It simply reflects the lack of information along the direction. A neural network might be able to make it more isotropic, but it does not add any "real" information. It is just guessing.

A more intuitive example: a 5 Å map looks dull but we don't call it artifact. A neural network might make it like a 3 Å map but all the added details come from prior information stored in the network. It is a guess. It might be true, but is not supported by the experiment.

As Sjors wrote, it is fine to use such networks to help your map interpretation. But you must not forget that experimental information in the poor direction is absent when you discuss e.g. side chain interactions with a ligand, even if the modified map shows good densities. The situation is similar (albeit bit better) to discussion based on AlphaFold models in that it is probably correct but not backed up by experimental observations.

I won't refine atomic models against maps modified by neural networks or evaluate the resolution of modified maps.