Articles · Research
Texas A&M screened 630,003 protein pairs with AlphaFold-Multimer
Texas A&M scientists ran an AI structure tool across every pairing of 1123 curated mitochondrial proteins and kept 2,895 likely partners out of 630,003.
Read the story
About 1,100 proteins sit inside human mitochondria, the energy centers of the cell, and mutations in nearly 30%% of them have been linked to human disease.
The team works in Texas A&M's Department of Biochemistry and Biophysics in College Station. It turned AlphaFold-Multimer, a tool built to predict the shapes of protein complexes, into a classifier for protein interactions. It screened 630,003 protein pairs and kept 2,895 as likely interactions.
How a pair makes the list
AlphaFold-Multimer is a neural network trained to predict the structures of complexes made of more than one protein chain. Google DeepMind created it, the release says. For this screen the authors built a curated list of 1123 proteins and asked the tool about every pair among them.
A pair counts as likely when its mean ipTM score reaches 0.5. The authors say that cutoff gives 67%% recall at a false discovery rate of about 15%% on their test set. At that cutoff the screen predicts 2895 interactions involving 1004 proteins. The paper also states 85%% precision for direct interactions.
What the list turned up
More than half of the likely interactions had never been reported before. Checked against existing interaction databases, 57%% were in none of them. The screen found at least one predicted partner for 85 uncharacterized proteins.
The team ran the same analysis on 11 diverse eukaryotes and found interactions among the human hits that held across evolution. Lab work guided by the predictions pointed to interactions that form the coenzyme Q metabolon and mapped the pathway that delivers copper to cytochrome c oxidase.
About two years of laboratory experiments confirmed several of the predicted partnerships in yeast and human cells. The team posted the predictions as a free resource called MitoMatch, which is made up of the predicted interactions and is online at mitomatch.web.app.
What the screen can't say
The release does not say how many predicted pairs were tested in the lab, only that several were confirmed. The paper says the method can't give the binding affinity of an interacting pair. It adds that many predicted interactions may occur only inside a larger complex rather than as true two protein pairs.
The release cautions that cures for diseases linked to mitochondria may still be far off. The computing came from Texas A&M High Performance Research Computing, which the release names among the study's supporters with The Welch Foundation.
Sources for this story
- AgriLife Today, Artificial intelligence points researchers toward new disease targets hidden in mitochondria
- Nature Communications, The predicted interactome of the human mitochondrial proteome (full text via Europe PMC)
- Nature Communications, The predicted interactome of the human mitochondrial proteome (abstract via Europe PMC)
Claim-by-claim verification · 30 claims
Roughly 1,100 proteins are known to sit inside human mitochondria, the energy centers of the cell, according to the Texas A&M AgriLife release.
about 1,100 proteins exist within mitochondria
For the screen itself the authors built a curated list of 1123 proteins.
generated a curated list of 1123 proteins
The team screened 630,003 protein pairs and assembled a compendium of 2,895 interactions, some already known and some new.
By screening 630,003 protein pairs, we create a compendium of 2,895 previously known and newly observed interactions
Of the 630,003 possible pairings, the screen flagged 2,895 as likely interactions.
Out of the 630,003 possible protein pairs, the team cataloged 2,895 likely interactions.
AlphaFold-Multimer is a neural network trained to predict the structures of protein complexes made of more than one protein chain.
AlphaFold-Multimer (AFM) is a neural network model trained to predict the structures of multimeric protein complexes
Google DeepMind created AlphaFold-Multimer, according to the Texas A&M AgriLife release.
Google DeepMind created and unveiled AFM
The Texas A&M team turned a structure prediction tool into a classifier for protein interactions across the mitochondrial proteome.
we repurpose the protein structure prediction algorithm AlphaFold Multimer (AFM) as a classification model to predict protein-protein interactions
The authors picked a mean ipTM score of 0.5 as the line between interacting and non-interacting pairs. On their test set that cutoff gives 67% recall at a false discovery rate of about 15%.
we decided to use 0.5 as a cutoff to separate interacting from non-interacting pairs, which gives 67% recall at a false discovery rate of about 15%
At the 0.5 mean ipTM cutoff, the screen predicts 2895 interactions involving 1004 proteins.
Based on our cutoff of 0.5 mean ipTM score, we predict 2895 interactions (0.46%) involving 1004 proteins
The paper says the full set of predictions carries an apparent false discovery rate of 15%.
at an apparent FDR of 15%
The authors say AFM can predict direct protein interactions in the mitochondrial proteome with 85% precision.
protein-protein interactions in the mitochondrial proteome with 85% precision
On a precision-recall benchmark, using the mean ipTM score as the classifier recovered ~40% of all true positives at 90% precision.
~40% of all true positives at 90% precision
More than half of the likely interactions had never been reported before, the release says.
More than half of them had never been reported before.
Checked against existing interaction databases, 57% of the predicted interactions were missing from all of them, so the paper counts them as previously unreported.
the remaining 57% were not present in these databases and thus represent previously unreported interactions
The predictions gave 85 previously uncharacterized proteins their first probable interaction partners.
gave 85 previously uncharacterized proteins their first probable partners
The team ran the same AFM analysis on 11 diverse eukaryotes and found interactions among the human hits that are conserved through evolution.
Extending the AFM-based analysis to 11 diverse eukaryotes identifies evolutionarily conserved interactions among human hits
Lab work guided by the predictions points to protein interactions that make up the coenzyme Q metabolon.
nominate protein interactions that form the coenzyme Q metabolon
The same experiments map the pathway that delivers copper inside mitochondria to cytochrome c oxidase.
metabolon and define the mitochondrial copper delivery pathway to cytochrome c oxidase
Roughly two years of lab experiments confirmed several of the predicted partnerships in yeast and human cells.
About two years of laboratory experiments followed, confirming several of the predicted partnerships in yeast and human cells
The team has posted its predictions online for free use under the name MitoMatch.
made its predictions available as a free online resource, MitoMatch
MitoMatch is made up of the 2895 predicted interactions.
These 2895 interactions constitute MitoMatch
Mutations in nearly 30% of mitochondrial proteins have been tied to human disease.
mutations in nearly 30% of mitochondrial proteins have been linked to human diseases
The release cautions that cures for diseases linked to mitochondria may still be a long way off.
Cures for mitochondria-linked diseases may still be far into the future
The paper states a limitation of the approach. It can't give the binding affinity of an interacting pair.
a limitation of this approach is that it cannot provide binding affinities of the interaction pair
The authors note that many predicted interactions may happen inside a larger complex and may not exist as true two-protein pairs.
in the context of a larger complex and may not exist as true heterodimers
The computing behind the work came from Texas A&M High Performance Research Computing.
advanced computing resources provided by the Texas A&M High Performance Research Computing
The release names Texas A&M High Performance Research Computing and The Welch Foundation among the study's supporters.
Texas A&M High Performance Research Computing and The Welch Foundation
The Texas A&M authors are in the Department of Biochemistry and Biophysics at Texas A&M University in College Station.
Department of Biochemistry and Biophysics, Texas A&M University, College Station, TX, USA
MitoMatch is online at mitomatch.web.app.
Our dataset, which we have named MitoMatch ( mitomatch.web.app ), consists of 2895
The screen found at least one interacting partner for 85 uncharacterized proteins.
we identified at least one interacting partner for 85 uncharacterized proteins