Leveraging AlphaMissense to predict the functional effects of transcription errors in coding RNA

Authors

Mark T. Handley [1]

[1] marktwh@pm.me

Abstract

Coding RNA is transcribed from template DNA with high - but not perfect - fidelity. Stochastic (random) transcription errors affecting individual RNA molecules can introduce sequence changes that cause amino acid substitutions within the proteins they encode.

AlphaMissense (1) is a dataset published by scientists at Google DeepMind that scores the functional effects of all possible amino acid substitutions in humans. It is intended primarily to inform the interpretation of potentially pathogenic genomic variants.

Our project will repeatedly simulate probabilistic transcription errors across the 39 million nucleotides of the canonical human coding transcriptome in silico. We will then leverage the 216 million substitutions scored in AlphaMissense as a lookup table to predict the functional effects of transcriptome-wide errors on a per-transcript basis. Our hypothesis is that these data will distinguish error-sensitive transcripts that contribute to disease pathogenesis.

Our particular research focus is severe inosine triphosphate pyrophosphatase (ITPase) deficiency, in which pathological inosine misincorporation into RNA occurs at high frequency. This rare disease has been the subject of three of our recent peer-reviewed publications (2-4). The endpoint of the project will be functioning open-source software, the data it has produced, and a publication describing these. We anticipate publication either as a standalone work or contributory to a larger study for which microarray, RNAseq, long-read sequencing and proteomics' experiments are already complete.

The wider significance of our work includes known or readily-inferred relevance to heart disease, chemotherapy, and ageing. Moreover, the 'central dogma' of molecular biology is concerned with the flow of information from DNA to RNA to protein. The effects of 'noise' on this flow are of general interest.

Together with funding, we hope that ResearchHub will provide a forum through which open development can proceed. We hope to benefit from the critical assessment and suggestions of the community as we progress.

Introduction

Motivation

By their nature, ‘random’ transcription errors affect single RNA molecules and so they are difficult to detect in the laboratory. The summed functional effects of many different errors in many different RNAs are therefore difficult to know. However, such errors can be simulated with a computer and a newly-published dataset, AlphaMissense, provides the opportunity to predict their effects. It is now possible to produce these potentially meaningful novel data at a relatively modest cost.

Impact/Significance

Severe ITPase deficiency causes a life-limiting multi-system disorder (2,4). Its symptoms include encephalopathy, seizures, and dilated cardiomyopathy. This study may inform understanding of its molecular pathogenesis.*
Dilated cardiomyopathy in the general population is frequently idiopathic (it's unclear what has caused it). If this study suggests that transcription errors may contribute to dilated cardiomyopathy in the context of a rare disease, it follows that these may also contribute to some cases where the cause is unknown. This study may inform investigation of that possibility.*
Cytotoxic chemotherapeutic drugs such as doxorubicin can cause dilated cardiomyopathy (5). This study may inform investigation of the possibility that transcription errors contribute to chemotherapy-induced cardiotoxicity.*
Ageing and exposure to mutagens are associated with transcriptional changes (6-8). This study may inform ageing research through identification of specific factors sensitive to these changes.*

*Other potential mechanisms underlying or contributing to the symptoms of severe ITPase deficiency are reviewed in (4). Relevance to heart disease, cancer chemotherapy and ageing is grounded speculation. Speculation is tested by experimentation.

Replicative error-rate and error-thresholds have long been studied in the context of the origin of life, the transmission of genetic information from one generation to the next, and the informational content of generations. This study may inform future research into the impact of error on the transmission of genetic information within discrete biological systems. Speculatively, transcription error may be subject to molecular surveillance and regulation at multiple levels, such as that of translation.

Hypothesis

This study will test the hypothesis that the coding-sequences of some transcripts are more sensitive to ‘random’ coding errors than others. Specifically, we will repeatedly simulate transcriptome-wide inosine misincorporation in the absence of ITPase. By reference to the AlphaMissense dataset, we will then determine whether particular transcripts are likely to be reproducibly functionally impaired under these conditions.

Aims

Generation of open-source software to simulate inosine misincorporation and predict its functional effects.
Generation of experimental data with these software.
Publication of research findings.

Materials/Methods

Study type

The main part of the study is an in silico experiment.

Existing data

The source data for this study are all publicly available.

This study comprises source data preparation, the use of source data to generate experimental data, the processing of this experimental data, and its subsequent interpretation through additional source data.

This application is made prior to analysis of the source data, and prior to the creation and analysis of the experimental data.

Explanation of existing data

The canonical protein-coding transcriptome will be used as source material into which inosine misincorporation will be simulated. ‘Canonical’ transcripts are a reference-set defined by Ensembl. Each is intended to provide a single representative example of a transcript produced from a specific gene/locus. We have chosen these as source material so that genes with lots of transcripts due to alternative splicing are not overrepresented in the experimental data generated. The current list of canonical protein-coding transcripts can be downloaded using BioMart here.
The coding-sequences of each canonical transcript can be obtained by cross-referencing the list to the transcriptome. We're using Ensembl release 111, which can be downloaded here.
To parameterise simulated inosine misincorporation, we will use data we have previously generated. We have used RNA digestion and mass spectrometry to quantify inosine levels in RNA in the absence of ITPase in (2). We have quantified the differing levels of inosine misincorporation in place of each standard nucleotide in (3).
To interpret the recoding - amino acid changes - produced by inosine misincorporation, we will initially assume that substituted ‘I’ is interpreted by the translation machinery as ‘G’ as this has been an assumption in the literature in the past. However, a recent paper (9) has challenged it. The data from this paper will be used to parameterise the probabilities of alternative amino acid changes in later versions of the software.
To predict the functional effects of transcript-recoding, we will use the AlphaMissense dataset. This can be downloaded here.

Data collection procedures

Preparation of source data will be carried out using R.

The set of canonical protein-coding transcripts must be as clean, up-to-date and complete as possible. However, only transcripts for proteins that have been scored in AlphaMissense should be included. AlphaMissense used old releases of the transcriptome and proteome. New canonical transcript identifiers will be mapped to the protein identifiers contained in AlphaMissense over several datasets. Respective transcript and protein lengths will be checked.
The AlphaMissense datasets contain proteins encoded by canonical and noncanonical transcripts. These will be filtered and merged so that there is a 1:1 correspondence between each transcript in the transcript set and each protein in the AlphaMissense set.

Generation of experimental data will be carried out using Rust. A number of experimental approaches are possible. The one we have chosen seems the most straightforward and intuitive, though it may also be the least computationally efficient.

We will approximate the probability of inosine misincorporation in place of each other nucleotide from the data available in the literature (2-3).
Each transcript is a vector of nucleotides, each with a defined position. We will iterate through each nucleotide of each transcript and use a random number generator (RNG) to determine whether it is substituted or not. For example, if there's a 1% chance of substitution and a RNG generating a number between 0 and 1, then a roll of between 0 and 0.01 means a substitution, and a roll of between 0.01 and 1 means no substitution.
In case of a substitution(s) in a codon, we will capture the codons-sequence (original and substituted), its position in the transcript, and the transcript identifier.
Each ‘sample’ will comprise a record of a single run-through of the whole transcriptome.

Sample size

At this point, we don't know how much time/computing power it will take to iterate-through the approximately 39,000,000 nucleotides of the transcriptome to generate a single sample. However, we estimate that between 1000 and 1,000,000 samples will be generated.

Processing and Analysis Plan

Individual entries - original and substituted codons - in the sample data will be in the form ‘codon-position-codon’ together with transcript- and run-identifiers. Entries in the prepared AlphaMissense dataset will be in the form aa-position-aa together with a transcript identifier.
To transform codons to amino acids under the assumption that ‘I’ is interpreted as ‘G’ just requires the genetic code as a lookup table.
To transform codons to amino acids where ‘I’ is variably interpreted as in (9) first requires a RNG to transform the I-containing codon in to a conventional codon, and then the genetic code lookup table to transform that into an amino acid.
With codon-position-codon changes transformed into aa-position-aa changes, corresponding AlphaMissense scores can be looked-up and collated per-transcript.
Further changes including stop codons introduced before the last exon, and loss of start codons, can also be considered.
Mean total score per transcript per run ( $\pm$ SE), proportion of runs where a transcript-score is over a given threshold, and FDR-corrected comparisons between different transcript-score distributions are appropriate ‘first-pass’ analyses.

Budget

Costs

The costs associated with this study will be $5000. In order of priority, funds will be utilized as follows:

Computation. I am in the process of arranging HPC access to complete the experimental data generation and analysis component of this project. However, in case the HPC is unavailable or inappropriate for this purpose, I will use a cloud-provider instead. Funds will be retained in case this is necessary. Additionally, it would be convenient to obtain a subscription to postit cloud for the preparation of the source data. A lot of RAM is required to even open the AlphaMissense datasets in R.
Bribery. I will reward critical assessment, suggestions and contributions to the project. I may offer bounties, for example “How do I make a large SQLite database and then use it as a lookup table?”.
Support. You will pay me for my time.

Other

Please note that I'm an experienced geneticist but a novice programmer. Whilst I'm confident I'll complete this project, and I'm not expecting anyone else to do it for me, I want it to be as good as possible. Your intellectual contribution could be as important as your financial one. I'm interested in Open Science, so this is an Open Science project. As an experimentalist, I'm going to try it like this:

1. Do a science project in the open.

2. Consider what worked and what didn't.

3. Make adjustments.

4. Repeat.

This is my test project.

References

Cheng J, Novati G, Pan J, Bycroft C, Žemgulytė A, Applebaum T, et al. Accurate proteome-wide missense variant effect prediction with AlphaMissense. Science. 2023 Sep 22;381(6664):eadg7492.
Handley MT, Reddy K, Wills J, Rosser E, Kamath A, Halachev M, et al. ITPase deficiency causes a Martsolf-like syndrome with a lethal infantile dilated cardiomyopathy. Plagnol V, editor. PLoS Genet. 2019 Mar 11;15(3):e1007605.
Schroader JH, Jones LA, Meng R, Shorrock HK, Richardson JI, Shaughnessy SM, et al. Disease-associated inosine misincorporation into RNA hinders translation. Nucleic Acids Res. 2022 Sep 9;50(16):9306–18.
Schroader JH, Handley MT, Reddy K. Inosine triphosphate pyrophosphatase: A guardian of the cellular nucleotide pool and potential mediator of RNA function. WIREs RNA. 2023 Sep;14(5):e1790.
Singal PK, Iliskovic N. Doxorubicin-Induced Cardiomyopathy. N Engl J Med. 1998 Sep 24;339(13):900–5.
Vermulst M, Denney AS, Lang MJ, Hung CW, Moore S, Moseley MA, et al. Transcription errors induce proteotoxic stress and shorten cellular lifespan. Nat Commun. 2015 Aug 25;6(1):8065.
Fritsch C, Gout JF, Haroon S, Towheed A, Chung C, LaGosh J, et al. Genome-wide surveillance of transcription errors in response to genotoxic stress. Proc Natl Acad Sci USA. 2021 Jan 5;118(1):e2004077118.
Debès C, Papadakis A, Grönke S, Karalay Ö, Tain LS, Mizi A, et al. Ageing-associated changes in transcriptional elongation influence longevity. Nature. 2023 Apr 27;616(7958):814–21.
Licht K, Hartl M, Amman F, Anrather D, Janisiw MP, Jantsch MF. Inosine induces context-dependent recoding and translational stalling. Nucleic Acids Research. 2019 Jan 10;47(1):3–14.

Funding

Funding

Leveraging AlphaMissense to predict the functional effects of transcription errors in coding RNA

Funding Progress

Nonprofit Recipient

Funders

DOI