Processing AlphaFold2, RoseTTAFold and other predicted models

Author(s)

Purpose

Process model files produced by AlphaFold, RoseTTAFold and other prediction software, replacing information these programs put in the B-factor field with pseudo-B values and optionally breaking the model into compact domains.

Background

As described in AlphaFold and Phenix, structure prediction software is now capable of generating models that are highly accurate over some or all parts of the models. Importantly, these predictions often come with reliable residue-by-residue estimates of uncertainty. The process_predicted_model tool is designed to help you remove low-confidence residues and break the model into domains for docking or molecular replacement.

How process_predicted_model works:

The process_predicted_model tool uses estimates of uncertainty supplied by structure prediction tools in the B-value (atomic displacement parameters) field of a model to create new pseudo-B values, to remove uncertain parts of the model, and to break up the model into domains.

If you used the Phenix server to get an AlphaFold prediction (See phenix.predict_model), the uncertainty estimate will be pLDDT (see section below for details of this metric).

The B-value field in most predicted models represents one of three possible values:

An actual B-value (atomic displacement parameter)

An estimate of error in A (rmsd)

A confidence (LDDT) on a scale of either 0 to 1 or 0 to 100.

In process_predicted_model, error estimates in A or confidence values are first converted to B-values. Then residues with high B-values are removed. Then the remaining residues are grouped (optionally) into domains.

Conversion of error estimates to B-values

Error estimates in A are converted to B-values using the standard formula for the relationshiop between rms positional variation and B-values:

B = rmsd**2  * ((8 * (pi**2)) / 3.0)

Conversion of LDDT values to error estimates

LDDT values are LDDT values are first converted to a scale of 0 to 1. You can specify whether the LDDT values in your model are from 0 to 1 (fractional) or from 0 to 100. If you don't specify, a model with all LDDT values between 0 and 1 is assumed to contain fractional LDDT values.

Then LDDT values on a scale of 0 to 1 are converted to error estimates using an empirical formula from

::
Hiranuma, N., Park, H., Baek, M. et al. Improved protein structure
refinement guided by deep learning based accuracy estimation. Nat Commun 12, 1340 (2021). https://doi.org/10.1038/s41467-021-21511-x

This empirical formula is:

RMSD = 1.5 * exp(4*(0.7-LDDT))

Trimming away low-confidence regions from predicted models

Normally it is a good idea to remove low-confidence regions from a predicted model before using them as a starting point for experimental structure determination. For AlphaFold2 models, low-confidence corresponds approximately to an LDDT value of about 0.7 (on a scale of 0 to 1, or 70 on a scale of 0 to 100), or to an RMSD value of about 1.5, or to a B-value of about 60. For other types of models these values might vary, so you might need to experiment or use values that others have found useful.

After trimming low-confidence residues, you will usually be left with a model that has some complete parts of various sizes and some small pieces.

Splitting a trimmed model into domains

It can be helpful to group the pieces from your trimmed model into compact domains, or even to split some pieces into compact domains. The process_predicted_model tool allows you to choose a typical domain size, and if you want, a maximum number of domains, and then it will try to split your model into compact domains.

There are two methods available. One is based finding compact domains, the other is based on using the predicted alignment error matrix (AlphaFold2 only).

Finding domains from a low-resolution model representation

The method used is to calculate a low-resolution map based on the input model, then to identify large blobs corresponding to domains. All the residues in the structure are assigned to an initial domain.Then the residues are regrouped in order to have as few cases where small parts of the model are part of one domain but neighboring parts are part of another as possible.

When using this method, you can easily adjust the number of domains you get by adjusting the target domain size (in A). You can also just restrict the number using the maximum_domains keyword (less good).

Finding domains using the predicted alignment error matrix

This method analyzes the predicted alignment error matrix (PAE) provided by AlphaFold2 and finds groupings of residues that have small mutual alignment error. This often corresponds to domains.

When using this method you can adjust the number of domains by changing the value of pae_power (the exponent applied to pae before using it in finding domains). You can also just restrict the number using the maximum_domains keyword (less good).

The pLDDT confidence measure

AlphaFold provides residue-level estimates of model accuracy in the form of predicted values of the LDDT (Local Difference Distance Test, Mariani et al., 2013). These predicted values are referred to as pLDDT values (predicted Local Difference Distance Test). The LDDT measure is a number from zero to one that reflects the similarity of CA positions in two structures.

The LDDT is not a simple measure of accuracy, but rather a composite that is based on the CA distances between all pairs of CA atoms in one model and how similar these inter-atomic distances are to those in the other model. It is a local test, where only inter-atomic distances that are less than 15 Å are considered.

The LDDT is a composite that is generated from the fraction of CA-CA distances that are very accurate (less than 0.5 Å off), the fraction that are quite accurate (less than 1 Å off), moderately accurate (less than 2 Å off), and somewhat accurate (less than 4 Å off). If all CA-CA distances are within 0.5 Å, the resulting score will be 1, if all are more than 4 Å off, the score is zero. Note that the LDDT is related to RMSD, but not in a simple way.

The pLDDT values (predicted Local Difference Distance Test) provided by AlphaFold are estimates of the LDDT values. The pLDDT values are essentially unbiased (just as likely to be too low as too high), and reasonably accurate. The correlation between pLDDT values and actual LDDT values calculated using AlphaFold models and models in the PDB is about 0.7 - 0.75, which means that the values of pLDDT you get are useful indicators of model quality but also that sometimes AlphaFold will have high confidence in an incorrect prediction or low confidence in a correct prediction.

Examples

Standard run of process_predicted_model:

Running process_predicted_model is easy. From the command-line you can type:

phenix.process_predicted_model my_model.pdb b_value_field_is=lddt

This will convert the B-value field in my_model.pdb based from LDDT to B-values, trim residues with LDDT less than 0.7, and write out a new model with individual chains (separate chain ID values) corresponding to compact domains.

Possible Problems

Specific limitations and problems:

Literature

Additional information

List of all available keywords