structural-biology/ai produced the result/Communications Chemistry 2025 · v2

Repeating-coil proteins designed from random sequences using AlphaFold2 in an evolution loop

Researchers built a design pipeline in which AlphaFold2 and a solenoid-classifying network scored randomly generated repeat sequences inside a genetic algorithm. Of 41 designs made in the laboratory, several alpha-solenoids behaved as modelled, and one was solved by X-ray crystallography.

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

Designing novel solenoid proteins with in silico evolution
Communications Chemistry, 2025

doi:10.1038/s42004-025-01817-3 · record aix-00145 v2 · checked 2026-10-09

ai-resultrole of AI

AI was for: Candidate generation, Structure determination, Classification
Model family: Transformer, Protein language model, Graph neural network, Diffusion model
Checked by: Experimental41 tested
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

Proteins are chains of amino acids that fold into shapes, and the shape decides what the protein can do. One family of shapes is the solenoid: the chain repeats a short motif over and over, and each copy stacks on the last to build a long coil or spiral, rather like a spring or a rolled-up ribbon. Solenoids come in flavours depending on which local structures the repeats use — helices, flat strands, or a mix. Designing one from scratch is awkward because the same repeat has to stack cleanly on itself many times over, and the open ends of the stack are exposed and prone to sticking to other copies.

The researchers set out to generate solenoid proteins without starting from any natural example, beginning instead from random repeating sequences, and then to make a selection of the resulting designs in the laboratory and see whether they folded as intended.

Where AI came in

AI supplied the search. Each candidate sequence was fed to AlphaFold2, a structure-prediction model, which proposed the shape the sequence would fold into along with its own confidence in that shape. A second network, SOLeNNoID, judged how solenoid-like the predicted shape was and of which class. The two scores were combined into a single cost, and a genetic algorithm mutated and selected sequences against it over successive generations. In place of building and testing in the laboratory, the loop used the models' predictions as the test.

Further models then took over other steps. ProteinMPNN rewrote the sequences for the accepted backbones, and AlphaFold2 and ESMFold re-predicted them as a filter. ESMFold was also used to probe how robust a design was when parts of its sequence were hidden. For the strand-type solenoids, ProteinGenerator designed caps for the two exposed ends. Every protein taken to the bench existed first only as the output of these models.

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 built an in silico evolution platform that couples a solenoid classifier network, SOLeNNoID, with AlphaFold2 as an oracle inside a genetic algorithm, starting from random repeat sequences. It produced 18 alpha-, 94 beta- and 8 alphabeta-solenoid designable backbones, which were sequence-redesigned with ProteinMPNN; of 24 designs taken to the bench, five alpha-solenoids gave SEC and circular dichroism behaviour consistent with their models, and one, A_1_1, was solved by X-ray crystallography at 2.8 A with a 0.89 A Calpha RMSD to the design model. All beta-solenoid designs in that round failed to give stable folded protein. After terminal caps were designed with ProteinGenerator and an ESMFold in silico melting filter was applied, 17 capped designs were tested and two showed beta-strand circular dichroism spectra and retained secondary structure up to 90 C.

How AI was used

Sequences were initialised as a random repeat of length L concatenated N times, then optimised in a hallucination loop: ColabFold-run AlphaFold2 in single-sequence mode with one recycle predicted each candidate's structure, the SOLeNNoID discriminator assigned per-residue alpha-, beta-, alphabeta- or non-solenoid probabilities, and the two scores were normalised to 0-1 and combined with equal weight into a single cost. A genetic algorithm with Wright-Fisher selection and a per-position mutation probability of 1/L updated a steady-state population until threshold criteria were met, with runs terminated after 30 generations. Accepted backbones were sequence-redesigned with ProteinMPNN at temperature 0.2 with cysteines disallowed, repredicted with AlphaFold2 and ESMFold, and kept as designable on RMSD, pAE and pLDDT thresholds. Designs were compared to the PDB with Foldseek and to the nr database with MMseqs, and ESMFold was used for in silico melting, masking 30-99% of input positions in 5% increments with 64 models per level to fit a sigmoidal Succ50 robustness value. For beta-solenoids, ProteinGenerator built N- and then C-terminal capping regions on the evolved backbones under an HXX secondary-structure pattern with cysteines excluded, with ProteinMPNN sequences and the same structural filters plus a Succ50 cutoff and Foldseek clustering used to select constructs for synthesis.

The shape of the work

Structural · the record, drawn

AI stepNo AI↤ what the AI stood in for

1Preparation
no AI

Initialise random repeat sequences

Cleaning, filtering, normalising or labelling data already obtained.

The input repeat sequence of length L is repeated N times to make the full NxL sequencewhere the paper describes this · verbatim

in the paper

2Inference
AI

Predict structure and score solenoid class

Running a trained model over new data to predict, classify or score.

The platform is guided by a solenoid discrimination network, SOLeNNoID, which assigns per-residue solenoid class scores.where the paper describes this · verbatim

in the paper

3Optimisation
no AI

Evolve sequence pool with genetic algorithm

Iterative search over a space. Its result feeds back into an earlier step.

The sequence pool is updated using a genetic algorithm.where the paper describes this · verbatim

in the paper

4Generation
AI

Redesign sequences and test designability

Producing candidate objects that did not previously exist.

The sequences from the in silico evolution loop were re-designed using ProteinMPNN.where the paper describes this · verbatim

in the paper

5Experiment
no AI

Express and characterise first design round

Physical execution, by hand or by robot.

We selected 24 solenoid designs for experimental characterisationwhere the paper describes this · verbatim

in the paper

6Validation
AI

Test in silico metrics against experimental outcomes

Testing outputs against ground truth.

we conducted a retrospective analysis of solenoid designs, derived from the in silico evolution pipeline, that had previously been experimentally testedwhere the paper describes this · verbatim

in the paper

7Generation
AI

Design terminal caps for beta-solenoids and filter

Producing candidate objects that did not previously exist.

used as scaffolds to design N- and C-terminal capping regions using ProteinGeneratorwhere the paper describes this · verbatim

in the paper

8Experiment
no AI

Express and characterise capped beta-solenoids

Physical execution, by hand or by robot.

We selected 17 capped β-solenoid designs for experimental characterisationwhere 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

All designed backbones and sequences were produced by the AF2-oracle hallucination loop, the discriminator network, ProteinMPNN and ProteinGenerator; the proteins characterised exist only as outputs of those models.

~What the AI was for

Candidate generation Structure determination Classificationour reading

~Model families

Transformer Protein language model Graph neural network Diffusion modelour reading

~How it was taught

Zero-shotSupervisedour reading

~Models named

AlphaFold2 (run via ColabFold, single-sequence mode) · Off the shelfSOLeNNoID solenoid discriminator · Off the shelfProteinMPNN · Off the shelfESMFold · Off the shelfProteinGenerator · Off the shelfour reading

+How results were checked

Experimental41 testedin the paper

We experimentally characterise 41 solenoid designswhere the paper describes this · verbatim

−Code · weights · data

code not reportedweights not reporteddata not reportednot reported

−Compute

not reportednot reported

What this paper did not report

Technical · absence is published deliberately

Reported as not stated — 13 items

CodeWhether the code is available is not stated.
Trained model weightsWhether the trained model is available is not stated.
DataWhether the data are available is not stated.
ComputeThe hardware or time used is not stated.
Version of AlphaFold2 (run via ColabFold, single-sequence mode)Which version of the model was used is not stated.
Version of SOLeNNoID solenoid discriminatorWhich version of the model was used is not stated.
Version of ProteinMPNNWhich version of the model was used is not stated.
Version of ESMFoldWhich version of the model was used is not stated.
Version of ProteinGeneratorWhich version of the model was used is not stated.
What step 2 replacedThe paper gives no basis for what the AI stood in for.
What step 4 replacedThe paper gives no basis for what the AI stood in for.
What step 6 replacedThe paper gives no basis for what the AI stood in for.
What step 7 replacedThe paper gives no basis for what the AI stood in for.

About this article

Record aix-00145, 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