alphaPredict 1.3.0

alphaPredict: A predictor of AlphaFold2 confidence scores

alphaPredict: A predictor of AlphaFold2 confidence scores

Last updated October 2024

v1.3 Removes the python3.12 requirement.

alphaPredict is moving to metapredict

NOTE: support for feature development in alphaPredict will likely end after v1.3. I will still keep it around and will put small amounts of time into it. However, the plan is to integrate the functionality of alphaPredict directly into metapredict. The main reason for this is to make it easier to support torch functionality such as predicting using GPUs, which massively improves the speed of prediction.

alphaPredict uses a bidirectional recurrent neural network (BRNN) trained on the per residue pLDDT (predicted IDDT-Ca) confidence scores generated by AlphaFold2 (AF2). The confidence scores from 21 proteomes (about 363,000 total protein sequences) were used to train the BRNN behind alphaPredict. These confidence scores measure the local confidence that AlphaFold2 has in its predicted structure. The scores go from 0-100 where 0 represents low confidence and 100 represents high confidence. For more information, please see: Highly accurate protein structure prediction with AlphaFold https://doi.org/10.1038/s41586-021-03819-2. In describing these scores, the team states that regionds with pLDDT scores of less than 50 should not be interpreted except as possible disordered regions.

Which proteomes were used to generate the network used by alphaPredict?

The confidence scores (pLDDT) from the proteomes of Danio rerio, Candida albicans, Mus musculus, Escherichia coli, Drosophila melanogaster, Methanocaldococcus jannaschii, Plasmodium falciparum, Mycobacterium tuberculosis, Caenorhabditis elegans, Dictyostelium discoideum, Trypanosoma cruzi, Saccharomyces cerevisiae, Schizosaccharomyces pombe, Rattus norvegicus, Homo sapiens, Arabidopsis thaliana, Zea mays, Leishmania infantum, Staphylococcus aureus, Glycine max, Oryza sativa were used to generate the BRNN (V4 and up).

Why is alphaPredict useful?

alphaPredict allows for rapid generation of predicted AF2 residue-by-residue confidence scores of any protein of interest. This can be used for many applications such as generating a quick preview of which regions of your protein of interest AF2 might be able to predict with high confidence, or which regions of your protein might be disordered. AF2 is not (strictly speaking) a disorder predictor, and the confidence scores are not directly representative of protein disorder. Therefore, any conclusions drawn with regards to disorder from predicted AF2 confidence scores should be interpreted with care, but they may be able to provide an additional metric to assess the likelihood that any given protein region may be disordered.

How accurate is alphaPredict?

The current BRNN (V7) has on average an ~8% error per residue. This is the average error rate for the V7 network. We are currently waiting on a new computer to try to generate new networks using more data, so this will be the most accurate network for the forseeable future. If you choose to use other networks: V1 has an average per residue error rate of about 11.5% V2 has an average per residue error rate of about 9.5% V3 through V6 were all networks that were tweaked before finishing V7, so I don't have those numbers on hand. They have errors of between 8.5% and 11% per residue.

Installation:

alphaPredict is available through PyPI - to install simply run

$ pip install alphaPredict

Alternatively, you can get alphaPredict directly from GitHub.

To clone the GitHub repository and gain the ability to modify a local copy of the code, run

$ git clone https://github.com/ryanemenecker/alphaPredict.git
$ cd alphapredict
$ pip install .

This will install alphapredict locally.

Usage:

alphaPredict is usable from Python.

Using alphaPredict from Python

First import alphaPredict -

import alphaPredict as alpha

Once imported, you can begin to generate predicted confidence scores.

Predicting Confidence Scores

The alpha.predict() function will return a list of predicted confidence scores for each residue of the input sequence. The input sequence should be a string. Running -

alpha.predict("DSSPEAPAEPPKDVPHDWLYSYVFLTHHPADFLR")

would output -

[39.5097, 43.5166, 46.9381, 55.6352, 54.2278, 56.5101, 60.3866, 58.0785, 60.2979, 65.6772, 69.3595, 66.0048, 68.0264, 68.4496, 71.1201, 70.3302, 73.5393, 76.7108, 81.8086, 85.8871, 86.4789, 87.4088, 88.8859, 87.3609, 84.9879, 79.5814, 80.5888, 79.3752, 79.8667, 83.2751, 83.6576, 81.2429, 78.8213, 72.8758]

Graphing Confidence Scores

The alpha.graph() function will return a graph of predicted confidence scores for each residue of the input sequence. The input sequence should be a string. Running -

alpha.graph("MASNDYTQQATQSYGAYPTQPGQGYSQQSSQPYGQQSYSGYSQSTDTSGYGQSSYSSYGQSQNTGYGTQSTPQGYGSTGGYGSSQSSQSSYGQQSSYPGYGQQPAPSSTSGSYGSSSQSSSYGQPQSGSYSQQPSYGGQQQSYGQQQSYNPPQGYGQQNQYNSSSGGGGGGGGGGNYGQDQSSMSSGGGSGGGYGNQDQSGGGGSGGYGQQDRGGRGRGGSGGGGGGGGGGYNRSSGGYEPRGRGGGRGGRGGMGGSDRGGFNKFGGPRDQGSRHDSEQDNSDNNTIFVQGLGENVTIESVADYFKQIGIIKTNKKTGQPMINLYTDRETGKLKGEATVSFDDPPSAKAAIDWFDGKEFSGNPIKVSFATRRADFNRGGGNGRGGRGRGGPMGRGGYGGGGSGGGGRGGFPSGGGGGGGQQRAGDWKCPNPTCENMNFSWRNECNQCKAPKPDGPGGGPGGSHMGGNYGDDRRGGRGGYDRGGYRGRGGDRGGFRGGRGGGDRGGFGPGKMDSRGEHRQDRRERPY")

would generate -

Acknowledgements

We would like to thank the DeepMind team for developing AlphaFold.

We would like to thank Dan Griffith from the Holehouse Lab at Washington University School of Medicine for developing PARROT, which is the tool that was used to generate the BRNN behind alphaPredict. For more info (and if you want to generate machine-learning networks for predicting anything related to proteins) see: https://idptools-parrot.readthedocs.io/en/latest/.

Changes

V1.3 - October 2024

• Now PEP517 compliant. All the tomls.

V1.2 - October 2024

• Updated to get rid of deprecation warning for torch.load() function.

V1.01 - November 2023

• Changed prediction to get rid of deprecation warning for numpy 1.25 and later.

• Restricted maximum python version to less than 3.12.0