structural-biology/ai produced the result/Viruses 2022 · v2

AlphaFold2 used to predict the shapes of hepatitis E virus copying proteins

Researchers fed the hepatitis E virus replicase sequence to AlphaFold2, the machine-learning structure prediction tool, and used the predicted shapes to mark out five protein domains and locate where their substrates and metal ions sit.

spectrum · one line per step, placed by what the step does · bright lines used AI

Structure Prediction and Analysis of Hepatitis E Virus Non-Structural Proteins from the Replication and Transcription Machinery by AlphaFold2
Viruses, 2022

doi:10.3390/v14071537 · record aix-00149 v2 · checked 2026-10-09

ai-resultrole of AI

AI was for: Structure determination
Model family: Transformer
Checked by: Replication
Code: not reported

The finding the paper is about came from the AI.

read as

The science is explained before the AI appears. Switch to field specialist to go straight to the method.

Assumes the discipline and goes straight to the method.

The diagram, the record and what the paper did not report are identical in both modes. Only the framing changes — never the evidence.

Introduction by AIxSci · plain language

What this research was about

Hepatitis E virus carries its genetic information as RNA. To multiply inside a cell it must copy that RNA, and it does so using a set of proteins that are made as one long chain, called a polyprotein. In this virus the chain in question, pORF1, is 1708 amino acids long. Working out what such proteins look like matters because a protein's three-dimensional shape is what lets it grip RNA, bind small helper molecules and do chemistry. But shapes are normally obtained by slow laboratory methods such as X-ray crystallography, and for hepatitis E virus the proteins of the copying machinery had not been pinned down that way.

The researchers set out to obtain shapes for these non-structural proteins from sequence alone, then to work out which parts of the long chain form separate, self-contained units, what each unit resembles among proteins whose structures are already known, and whether the units touch one another.

Where AI came in

AlphaFold2 did all the structure prediction. It is a machine-learning program that takes an amino acid sequence and returns a predicted set of atomic coordinates, and here it was used off the shelf through DeepMind's public Colab notebook, without being given existing laboratory structures as templates. Because the notebook's graphics memory stretches to roughly 1400 residues, the 1708-residue chain was split into two overlapping pieces, residues 1 to 1250 and 1000 to 1708, and each was predicted separately. The program also reports a confidence value for every residue, and those values were plotted to tell compact folded regions from the floppy linkers between them.

In place of experimental structure determination, the predictions became the raw material for ordinary structural analysis done by hand and by other software. The five separated units were matched against known structures with the Dali server, giving the labels a capping enzyme, a zinc-binding region, a macro domain, a helicase and an RNA-copying enzyme. Superimposing the predictions on matched laboratory structures that held ADP-ribose, an ATP look-alike, RNA or zinc placed those partners in the predicted sites. AlphaFold2 was run again on an nsP1 pair and on nsP2 to nsP5 together, which showed no contacts, and on a SARS-CoV-2 complex as a check against its known structure.

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 ran AlphaFold2 on the 1708-residue pORF1 replicase polyprotein of hepatitis E virus genotype 3, splitting it into two overlapping segments (residues 1-1250 and 1000-1708) to fit the notebook's memory limit. From the per-residue confidence scores they separated five well-predicted domains, assigned residue boundaries, and identified structural homologs in the PDB with Dali, describing them as a methyltransferase-like capping protein, a zinc-binding domain, a macro domain, a helicase and an RNA-dependent RNA polymerase. Superimposing the predictions onto experimental homologs bound to ADP-ribose, an ATP analogue, RNA or Zn2+ placed those substrates and cofactors in the predicted sites. Predictions of the polyprotein segments and of nsP2 to nsP5 submitted together showed no contacts between the domains, while an nsP1 dodecamer built by symmetry docking and an AlphaFold2 nsP1 dimer were mutually compatible.

How AI was used

AlphaFold2 was used off the shelf, through the public DeepMind Colab notebook that does not use PDB templates, to predict structures from sequence for the HEV-3 Kernow-C1 pORF1 polyprotein (GenBank HQ389543). The sequence was split into two overlapping segments because of the notebook's GPU memory limit, and each segment was predicted separately; per-residue pLDDT values stored in the PDB B-factor column were plotted to distinguish folded domains from linkers. Predicted structures were edited in Coot to separate individual non-structural proteins and fix their boundaries, searched against the PDB with the Dali server to retrieve structural homologs, and rendered and superimposed in ChimeraX so that ligands, ions and nucleic acids present in homolog structures could be positioned in the predictions. AlphaFold2 was additionally run on an nsP1 dimer and on nsP2 to nsP5 submitted together to test for inter-domain contacts, and on the SARS-CoV-2 nsp7/nsp8/nsp12 complex as a control; the nsP1 dodecameric ring was built not by AlphaFold2 but by geometry-based symmetry docking with SymmDock.

The shape of the work

Structural · the record, drawn

AI stepNo AI↤ what the AI stood in for

1Preparation
no AI

Define polyprotein segments for prediction

Cleaning, filtering, normalising or labelling data already obtained.

The HEV-3 polyprotein pORF1 of 1708 residues was split into two overlapping segments for AF2 structure predictionswhere the paper describes this · verbatim

in the paper

2Inference
AI

Predict segment structures with AlphaFold2

Running a trained model over new data to predict, classify or score. The AI stood in for physical experiment.

we used AF2 with the replicase encoded by the polyprotein pORF1 of the human-infecting HEV-3where the paper describes this · verbatim

in the paper

3Preparation
no AI

Separate individual nsPs and set domain boundaries

Cleaning, filtering, normalising or labelling data already obtained.

The structures were edited with Coot to separate the individual nsPs.where the paper describes this · verbatim

in the paper

4Validation
no AI

Retrieve and compare experimental structural homologs

Testing outputs against ground truth.

Related structures retrieval was performed with Dali.where the paper describes this · verbatim

in the paper

5Interpretation
no AI

Locate substrate and cofactor sites by superimposition

Extracting understanding from model behaviour.

the superimposition of the predicted structures to the best Dali hits encompassing nucleic acids or ionswhere the paper describes this · verbatim

in the paper

6Simulation
no AI

Build nsP1 dodecamer by symmetry docking

Numerical or physics simulation, including where a learned surrogate replaces it.

we generated the HEV-3 nsP1 dodecamer with SymmDock, a server for the prediction of complexes with Cn symmetrywhere the paper describes this · verbatim

in the paper

7Inference
AI

Predict nsP assemblies and inter-domain contacts with AlphaFold2

Running a trained model over new data to predict, classify or score. The AI stood in for physical experiment.

we performed two structural predictions using nsP2 to nsP5 as inputswhere the paper describes this · verbatim

in the paper

8Validation
AI

Predict SARS-CoV-2 replication complex as a control

Testing outputs against ground truth. The AI stood in for physical experiment.

our AF2 structure predictions of the well-known SARS-CoV-2 replication complexwhere the paper describes this · verbatim

in the paper

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.

~Role of AI

AI produced the resultour reading

Every structure and domain boundary reported in the paper is an AlphaFold2 prediction; no experimental structure determination was performed here

~What the AI was for

Structure determinationour reading

~Model families

Transformerour reading

~How it was taught

Supervisedour reading

~Models named

AlphaFold2 · Off the shelfour reading

+How results were checked

Replicationin the paper

resulted in a protein complex comparable to that observed in experimental structures (Figure S2)where the paper describes this · verbatim

+Code · weights · data

code not reportedweights not reporteddata availablein the paper

Predicted structures coordinates (PDB format) will be available in the Supplementary Material.where the paper describes this · verbatim

+Compute

AlphaFold2 Colab notebook servers, whose GPU memory is stated to have an upper limit of ~1400 residuesin the paper

What this paper did not report

Technical · absence is published deliberately

Reported as not stated — 4 items

CodeWhether the code is available is not stated.
Trained model weightsWhether the trained model is available is not stated.
How many were testedThe paper gives no count of what was tested.
Version of AlphaFold2Which version of the model was used is not stated.

About this article

Record aix-00149, 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