I got really confused while doing refinements like what I choose between Homogeneous, heterogeneous, or non uniform refinement. What should be the pipeline one should follow?
Hi @BhawnaMishra! I know there are a lot of types of refinement, let me see if I can clear them up for you! I think one useful way to think about the jobs you mentioned is whether they are trying to make a single, high-quality map or classify particles.
Single, high-quality map
As you know, our ultimate goal in single-particle analysis is producing a high-quality map of whatever 3D object created the particle images. We do this by taking an input 3D volume and determining the most likely pose of that volume in each particle image. Recall that a pose is the rotation and translation applied to the reference volume to make it match the particle images.
Both Homogeneous Refinement and Non-Uniform Refinement do this.
When you use these jobs, you ideally would provide:
- A starting volume that looks reasonably similar to the particle images. Usually, this would come from a previous refinement or from
Ab Initio Reconstruction. - A clean particle stack. The less junk (i.e., non-particle images) there is, the better the result of these jobs will be
- Optional: a mask. You can use a mask to focus on one region of the protein, and to exclude parts of the image that will hurt alignments. Masks are often used to exclude detergent micelles, for instance.
When the job finishes, you will get:
- The volume most likely to have produced the particle images, filtered by the GSFSC to help reduce overfitting.
- A particle stack with update pose estimates. These poses are used in downstream jobs like
3D Classificationand3D Variability Analysis.
The difference between Homogeneous and Non-Uniform Refinement
The goal and algorithm of these two jobs is very similar. Non-Uniform Refinement includes an additional step which accounts for local differences in data quality. More detail can be found in the Non-Uniform Refinement job page and the original paper. Typically, Non-Uniform Refinement performs better on flexible targets or targets with large disordered domains (e.g., micelles), but is a bit slower than Homogeneous Refinement.
Classifying particles
Because micrographs have very low signal-to-noise ratio, it is difficult to see and pick particles. In practice, this means that our initial particle stacks have many images of empty ice, or of things which are not our particle of interest. To remove images of empty ice or contaminants, we can use jobs like Heterogeneous Refinement to classify the particle images into one of several volumes. You can then inspect the volumes and determine if you want to use the particles that made those volumes for downstream jobs.
For example, say you run a Heterogeneous Refinement job with two classes. This means each particle will be aligned to both volumes. It will then be assigned to the class of the volume it matches better. If one of the volumes looks like a good map of your target and one looks noisy or otherwise low-quality, you can throw away the bad volume and thereby remove all the junk particles that were used to make it.
In this way, Heterogeneous Refinement is similar to Ab Initio Reconstruction, but Heterogeneous Refinement uses the same alignment algorithm as the other refinement jobs, and also includes final passes through the data. Generally, this is taken to mean that the classifications from a Heterogeneous Refinement are more reliable than those from Ab Initio Reconstruction.
Summary
If your goal is to get a single, high-quality map of your particle stack, you should try Homogeneous Refinement. If your target has large flexible or disordered regions, you might prefer Non-Uniform Refinement instead.
If your goal is to separate particles into several classes, either to detect heterogeneity or to remove junk, Heterogeneous Refinement is the job you should use.
3 Likes