structural-biology/ai produced the result/Bioinformatics Advances 2022 · v2
Structure predictions place an uncharacterised human protein in the BRICHOS family
Researchers compared AlphaFold-predicted shapes of known human BRICHOS proteins, distilled a shared core, and searched the predicted human proteome with it. The search picked out a little-studied protein called Out at First, or OAF.
spectrum · one line per step, placed by what the step does · bright lines used AI
OAF: a new member of the BRICHOS family
Bioinformatics Advances, 2022
doi:10.1093/bioadv/vbac087 · record aix-00158 v2 · checked 2026-10-09
- AI was for
- Structure determination, Property prediction
- Model family
- Transformer
- Checked by
- None stated
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
Proteins fold into particular three-dimensional shapes, and that shape often says more about what a protein does than its chemical sequence alone. Families of related proteins tend to share a fold even when their sequences have drifted far apart. One such family is defined by the BRICHOS domain, a region found in several otherwise different human proteins. Sorting proteins into families like this is hard when experimental structures are scarce: working out a protein's shape in the laboratory, by X-ray crystallography or similar methods, is slow and often fails. For the BRICHOS domain, only a single experimental structure exists, that of the proSP-C domain.
The researchers set out to ask whether any other human protein shares the BRICHOS fold without having been recognised as a relative. They worked out what the family's structures have in common, used that common core as a search pattern across human proteins, and then tested whether the best match was genuinely related rather than merely similar in outline. They also looked at what the candidate's properties might be, including whether part of it is prone to forming amyloid, the sticky, ordered protein clumps associated with several diseases.
Where AI came in
AlphaFold, a system that predicts a protein's three-dimensional shape from its sequence, supplied nearly all the structures this analysis ran on. Because only one BRICHOS domain has been solved experimentally, the predicted models stood in for laboratory structure determination. The authors inspected AlphaFold models of the human BRICHOS proteins alongside the proSP-C X-ray structure to mark out a shared domain core, then used one core per subfamily to search AlphaFold models of the whole human proteome. That search returned OAF. The search and alignment tools themselves, Dali and HHpred, are conventional methods rather than machine learning.
Two further ready-made predictors were applied to sequences. AMYPred-FRL, described in the paper only as a machine learning approach, scored how likely stretches of protein are to form amyloid-like structures; it gave OAF's predicted mature polypeptide a 97% probability, the highest among the BRICHOS polypeptides examined, standing in for a laboratory assay the authors say is now warranted. PolyPhen2 scored a reported T171I substitution, a single-letter change in the protein's sequence, as probably damaging. No model was trained or adjusted for this work, and the paper reports no model versions, parameters or computing details.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
The authors compared AlphaFold-predicted structures of the ten known human BRICHOS domain-containing proteins with the single existing X-ray structure of the proSP-C BRICHOS domain, defined a conserved structural core, and used that core to search the AlphaFold human proteome with Dali. The search returned the uncharacterized Out at First (OAF) protein, and HHpred profile-to-profile comparison found statistically significant sequence similarity between OAF and BRICHOS subfamily profiles, with the OAF profile matching the Gastrokine, ITM and proSP-C profiles at E-values of 1.6 x 10-4, 0.008 and 0.015. The authors report that OAF carries five features shared across the family, including a putative proprotein convertase cleavage site and a predicted mature polypeptide of 70 residues (residues 204-273) held by four conserved disulphide bridges. A machine learning predictor, AMYPred-FRL, assigned that polypeptide a 97% probability of forming amyloid-like structures, the highest among the BRICHOS mature polypeptides examined.
How AI was used
AlphaFold models supplied the structures the analysis ran on, since only one BRICHOS domain structure has been determined experimentally: models of the human BRICHOS proteins were inspected alongside the proSP-C X-ray structure to locate mature polypeptides and to delimit a conserved domain core, and one core structure per subfamily was then used as a Dali query against AlphaFold models of the whole human proteome. Dali itself is a conventional structural alignment method, as is the HHpred profile-to-profile search that was then run against the Pfam profile database, using profiles built for each BRICHOS subfamily, to test whether the structural similarity reflected homology. Two further pre-existing predictors were applied to sequences: AMYPred-FRL, described only as a machine learning approach, to score the amyloid-forming propensity of the wild-type OAF mature polypeptide and of other BRICHOS mature polypeptides, and PolyPhen2 to score the reported de novo T171I substitution. No model was trained or fine-tuned in this study, and no model versions, parameters or compute are reported.
The shape of the work
Structural · the record, drawn
AI
Obtain AlphaFold models of human BRICHOS proteins
Obtaining raw data, whether by measurement, download or retrieval. The AI stood in for physical experiment.
We started by inspecting all AlphaFold-predicted structures of human BRICHOS domain proteinswhere the paper describes this · verbatim
no AI
Define conserved BRICHOS structural core as query
Cleaning, filtering, normalising or labelling data already obtained.
Structural comparison of BRICHOS AlphaFold models thus defined a conserved structural core for the BRICHOS domainwhere the paper describes this · verbatim
no AI
Structure search against AlphaFold human proteome
Reducing a candidate set by filtering or ranking, in a single pass.
Subsequent structural searches with Dali were undertaken against the AlphaFold human proteome, using human BRICHOS domain cores as query structureswhere the paper describes this · verbatim
no AI
Test sequence homology with profile comparison
Testing outputs against ground truth.
we performed a sequence conservation analysis using the HHpred profile-to-profile comparison toolwhere the paper describes this · verbatim
AI
Predict amyloid propensity of mature polypeptides
Running a trained model over new data to predict, classify or score. The AI stood in for physical experiment.
we applied a machine learning approach, AMYPred-FRL, to its wild-type sequencewhere the paper describes this · verbatim
AI
Score pathogenicity of the T171I variant
Running a trained model over new data to predict, classify or score.
Its mutation to isoleucine is predicted by PolyPhen2 to be probably damaging (score: 0.992; sensitivity: 0.70; specificity: 0.97)where the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The paper's central claim — that OAF belongs to the BRICHOS family — rests on AlphaFold-predicted structures, since only one experimental BRICHOS structure exists; the amyloid-propensity claim rests on a machine learning predictor.
Using structural comparison of coevolution-based AlphaFold models and sequence conservation, we identified the Out at First (OAF) proteinwhere the paper describes this · verbatim
Supplementary data are available at Bioinformatics Advances onlinewhere the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- CodeWhether the code is available is not stated.
- Trained model weightsWhether the trained model is available is not stated.
- ComputeThe hardware or time used is not stated.
- ValidationNo validation of the AI is described.
- Version of AlphaFoldWhich version of the model was used is not stated.
- Version of AMYPred-FRLWhich version of the model was used is not stated.
- Version of PolyPhen2Which version of the model was used is not stated.
- What step 6 replacedThe paper gives no basis for what the AI stood in for.
About this article
Record aix-00158, version 2, checked by a person on 2026-10-09. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error