structural-biology/ai produced the result/Cancer Genetics 2025 · v2
Software models of two cancer proteins used to rank mutations by likely effect
Researchers built computer-predicted structures of the proteins PRMT5 and RUVBL1 in humans and two yeasts, and scored cancer mutations in them. AlphaFold 3 supplied the structures; a classifier called CHASM Plus rated each mutation as a likely driver or a bystander.
spectrum · one line per step, placed by what the step does · bright lines used AI
In silico protein structural analysis of PRMT5 and RUVBL1 mutations arising in human cancers
Cancer Genetics, 2025
doi:10.1016/j.cancergen.2025.01.002 · record aix-00143 v2 · checked 2026-10-09
- AI was for
- Structure determination, Classification
- Model family
- Transformer, Diffusion model, Random forest
- Checked by
- None stated
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
Proteins work by folding into particular shapes, and a mutation that swaps one amino acid for another can matter a great deal or not at all, depending on where in that shape it lands. Tumours carry many mutations, but only some of them help the cancer grow. The rest are passengers, picked up along the way. Telling the two apart is hard, and it helps to know the protein's shape. Shapes are normally worked out by experiment, which is slow, and for many proteins no experimental structure exists at all. That is the case for the yeast counterparts of the two proteins studied here.
The two proteins are PRMT5 and RUVBL1. The researchers set out to compare their shapes across humans and two yeast species, to see how far the yeast versions resemble the human ones, and to work through mutations in both proteins recorded in a public catalogue of cancer mutations, asking which were likely to be drivers and where in the structure they sat. No laboratory experiments were done; the authors note that the predictions would need experimental testing.
Where AI came in
Two trained tools were used as they came, without retraining. AlphaFold 3 predicts a protein's three-dimensional shape from its amino acid sequence. It was given the PRMT5 sequences from human and from the two yeasts, and the RUVBL1 and RUVBL2 sequences, and returned full-length models, including models of the ring-shaped assembly the RUVBL proteins form. For the yeast proteins, which have no experimental structures, this stood in for the laboratory work that would otherwise have produced them. The models were then superposed in ordinary structural software to measure how closely the shapes matched, and the same overlays were later used to find which yeast positions correspond to the mutated human ones.
The second tool, CHASM Plus, was reached through a public web server and handed the single amino acid swaps taken from the COSMIC mutation catalogue. For each swap it returns a probability that the change is a driver rather than a passenger, which the researchers judged against a cutoff of 0.05. The remaining steps, including a stability calculation and the identification of residues sitting at the contact surfaces between proteins, used conventional software rather than learned 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
AlphaFold 3 models of PRMT5 and of the RUVBL1-RUVBL2 hexamer were built for human, S. pombe and S. cerevisiae, then aligned in PyMOL; global alignment of human PRMT5 with S. pombe Skb1 gave an RMSD of 4.64 Å against 9.65 Å for S. cerevisiae Hsl7, and 1.04 Å against 1.23 Å when outlier atoms were rejected. PRMT5 and RUVBL1 mutations drawn from COSMIC version 100 were then scored with the CHASM Plus driver classifier and the CUPSAT stability predictor: ten of eighteen RUVBL1 substitutions co-occurring with PRMT5 mutations had CHASM p-values below 0.05, while none of the PRMT5 mutations did. Interface analysis of the TIP60 cryo-EM structure placed 10 of the 15 co-occurring RUVBL1 mutated residues at the RUVBL1-RUVBL2 interface. The authors state that the predictions would need experimental validation.
How AI was used
Two learned tools were used off the shelf. AlphaFold 3 was run on PRMT5 sequences from human, S. pombe and S. cerevisiae and on RUVBL1-RUVBL2 to produce full-length structure models where no experimental structures exist, and those models were superposed in PyMOL with the align command, both as global alignments with cycles set to 0 and with outlier atoms rejected, to collect RMSD values; the same alignments were later read off to find S. pombe residues analogous to the human mutated positions. Mutation records for PRMT5 and RUVBL1 were downloaded from COSMIC and the single amino acid substitutions were submitted through the OPEN CRAVAT server to CHASM Plus, which returns a probability that a substitution is a driver rather than a passenger, with 0.05 taken as the significance cutoff. Separately, and without a learned model, mutated residues were mapped onto the experimental TIP60 complex structure (PDB 8XVG) and PRMT5 structure (PDB 4GQB) in PyMOL, CUPSAT was used to compute ΔΔG values for each mutation, and PDBePISA was used to identify which mutated residues sit in protein-protein interfaces.
The shape of the work
Structural · the record, drawn
no AI
Download cancer mutation records
Obtaining raw data, whether by measurement, download or retrieval.
Files with COSMIC mutations for PRMT5 and RUVBL1 were downloaded on September 10, 2024.where the paper describes this · verbatim
AI
Predict structures of human and yeast homologs
Running a trained model over new data to predict, classify or score. The AI stood in for physical experiment.
The S. pombe and S. cerevisiae PRMT5 homologs do not have experimental structures available, thus AlphaFold was used to generate models.where the paper describes this · verbatim
no AI
Align predicted structures and record RMSD
Extracting understanding from model behaviour.
Alignments were performed in PyMOL using the align command with cycles set to 0 for global alignmentswhere the paper describes this · verbatim
AI
Classify substitutions as driver or passenger
Running a trained model over new data to predict, classify or score.
We used the CHASM Plus tool within this website which generates probability valueswhere the paper describes this · verbatim
no AI
Map mutated residues onto experimental structures
Cleaning, filtering, normalising or labelling data already obtained.
Mutated residues were mapped onto the respective structures using PyMOL version 3.0.3.where the paper describes this · verbatim
no AI
Predict stability change and locate interface residues
Extracting understanding from model behaviour.
Structure files were uploaded to the CUPSAT web server to determine ΔΔG values and overall stability of each analyzed mutation.where the paper describes this · verbatim
no AI
Check conservation of mutated residues in S. pombe
Extracting understanding from model behaviour.
The PyMOL alignments in Fig. 1 were used to identify the analogous residues in S. pombe to the human residueswhere 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 two main claims — structural conservation of PRMT5/RUVBL1 between S. pombe and human, and that RUVBL1 mutations are more likely drivers than PRMT5 mutations — rest directly on AlphaFold-generated models and on CHASM Plus driver probabilities; no experimental work was performed.
AlphaFold 3 models were generated for the human, S. pombe, and S. cerevisiae protein structureswhere the paper describes this · verbatim
the predictions made here would need to be validated by experimentationwhere the paper describes this · verbatim
All data are available on public repositories COSMIC.where 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 CHASM PlusWhich version of the model was used is not stated.
- What step 4 replacedThe paper gives no basis for what the AI stood in for.
About this article
Record aix-00143, 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