Rationale for using neural network particle pickers

Hi everyone, I wonder what’s the community’s opinion on when to use deep learning based methods for particle picking vs template-based picking?

To me, the rationale is the following: if there are a lot of different conformations in your sample (that you don’t necessarily know), than deep learning methods can be good since they can distinguish particles from non-particles based on their vague features.

However, if you know the conformational ensemble of your sample, or if it’s really rigid and inflexible, than there are no advantages in using NN-based methods over template-based picking.

Is it correct, or there are other advantages that I may not be aware of?

In my opinion, template picking is the best, and blob picking with optimized parameters is also very good.

In the former case, even if you don’t know the conformations perfectly, after the 15 - 20Å low pass, most particles are going to already be picked well with more limited sets of views/conformations. In the latter case, the goal will be to overpick anyway, which is pretty easy to achieve. In my experience, most people with complaints about blob picking turn out not to have tested very many parameters, and I am able to help them fairly easily. If you set the picker to process only ~10 micrographs, it will complete almost instantly, and around 30-50 steps of parameter optimization can be completed in under 30 minutes. Looking at the “red dot” images (“show from top”) is sufficient.

I have never seen deep network picking to work so much better so as to justify the greater amount of user input and processing time required. On the other hand, the more old school signals-based neural net method in EMAN2 is highly effective, and I have used it for initial picking in some super difficult cases, e.g. very heterogeneous RNA-only particles. This method makes radial averages of the training particles and uses a simple convolutional network to classify them. Only ~100 examples of foreground (particles), background, and bad picks are required and it’s very efficient. The main downsides are having to repeat CTF estimation via an EMAN2 project and then import a bunch of .box files to Relion, but it’s not that big of a deal.

PS Once I had structures I used template picking again for those RNAs.

1 Like

We have found the neural network pickers very useful when trying to selectively pick a low percentage component of a heterogeneous mixture (e.g. in a native extract) - in these cases we have had much better luck with Topaz etc than with either template or blob based approaches.


Another situation in which neural net pickers work better is on micrographs containing a lot of high-contrast features that are not particles (frost contamination, carbon edge, etc.). Neural net pickers are very good at avoiding these features, while template matching and blob pickers tend to either pick a lot of these, or miss a lot of good particles if you tweak their settings to make them avoid the things you don’t want to pick.

Some will say “make better grids” without these high-contrast contaminants, but sometimes you have to make do with the micrographs you already have, so it’s good to have tools for that.


@Guillaume I find cryoSPARC’s power threshold approach works really well for those.

1 Like