structural-biology/ai produced the result/Molecular Cell 2025 · v2
Classifier sorts true from false AlphaFold predictions of human protein pairs
Researchers folded curated pairs of human proteins with AlphaFold-Multimer, then trained a random forest classifier, SPOC, to score which predicted pairings look real. SPOC was applied to 40,459 predictions among 286 genome maintenance proteins.
spectrum · one line per step, placed by what the step does · bright lines used AI
Predictomes, a classifier-curated database of AlphaFold-modeled protein-protein interactions
Molecular Cell, 2025
doi:10.1016/j.molcel.2025.01.034 · record aix-00151 v2 · checked 2026-10-09
- AI was for
- Structure determination, Classification
- Model family
- Random forest, Transformer, Protein language model
- Checked by
- Held-out946 tested, 541 worked
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
Proteins rarely work alone. They stick to one another in pairs and larger assemblies, and knowing which protein touches which is a large part of understanding how a cell works. Finding those contacts experimentally is slow, and working out the shape of a joined pair slower still. Software that predicts the three-dimensional shape of a protein pair from its sequence alone has made it possible to try vast numbers of candidate pairings cheaply. The difficulty is that such software will produce a structure for any two sequences handed to it, whether or not the two proteins ever meet in a cell. Separating the real pairings from the confident-looking mistakes is the bottleneck.
The authors set out to build a scoring system for that judgement, and a public collection of scored predictions. They assembled reference sets of protein pairs: some drawn from chemical crosslinking studies that tag proteins sitting close together, some from experimentally solved structures deposited after the prediction software's training cut-off, and some deliberately assembled at random to serve as likely negatives. They then focused on proteins involved in genome maintenance, the repair and copying machinery that keeps DNA intact.
Where AI came in
AlphaFold-Multimer, run through a local ColabFold installation, predicted a structure for each protein pair. This stands in for the laboratory work of determining a structure experimentally. Predictions that showed no plausible contact between the two proteins were filtered out before anything else happened.
The second use of machine learning was the scoring step. For each surviving pair the researchers extracted numbers describing the predicted contact surface, such as confidence measures and counts of salt bridges and hydrogen bonds, alongside information from outside the structure: whether the two genes are switched on together, whether cells need both at once, how often the pair has been reported before, and where in the cell each protein is predicted to sit, the last from a separate sequence model. A random forest, a method that combines many simple decision trees, learned from these features to give each pair a single score. It replaces the hand-designed confidence numbers that the folding software reports.
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 folded curated sets of human protein pairs with AlphaFold-Multimer and trained a random forest classifier, SPOC, on structural features of the predicted interfaces together with omics features such as co-expression, CRISPR co-dependency and BioGRID interaction counts, in order to separate true from false predicted protein-protein interactions. On held-out test sets SPOC reached an ROC AUC of 0.96, compared with 0.92 for a classifier using structural features alone, and at a 128:1 negative-to-positive ratio it recalled 50.4% of positive pairs at 5% false discovery rate versus 11.3% for pDockQ2. SPOC was applied to 40,459 predictions covering nearly all binary combinations of 286 human genome maintenance proteins, of which 11,523 met the contact criteria and 1,151 scored above 0.33; these are browsable at predictomes.org, which also scores user-supplied predictions.
How AI was used
AlphaFold-Multimer v3 weights (models 1, 2 and 4, three recycles, templates enabled) were run through a local ColabFold v1.5.2 installation with MMseqs2-generated MSAs to predict structures for binary protein pairs drawn from randomly sampled UniProt proteins, non-contacting subunits of known complexes, crosslinked pairs mined from 20 XLMS studies, randomised XLMS pairs, and PDB heterodimers deposited after the training cutoff. Predictions were filtered to contact-positive interfaces using distance, pLDDT and PAE criteria, and the five reference sets were homology-reduced before being split 75/25 into training and test partitions. For each pair, numeric features were extracted from the predicted interface (PAE, pLDDT, avg_models, salt bridges, hydrogen bonds) and from external sources including DEPMAP co-dependency vectors, coexpressDB co-expression, BioGRID interaction counts, BioGRID-ORCS CRISPR hit vectors, precomputed T5 per-protein embeddings and subcellular localisation probabilities predicted with a locally installed DeepLoc 2.0; features were combined across both proteins so protein identity was not learnable. Random forests were fitted with scikit-learn using a hyperparameter grid search over three-fold splits together with iterative pruning of features whose GINI importance fell below 0.01, yielding a structural classifier and SPOC. The trained classifier was then run over the genome maintenance matrix, the ranking experiments and three proteome-wide screens, with scores thresholded for inclusion in the web portal.
The shape of the work
Structural · the record, drawn
no AI
Assemble protein pair sets for training, testing and screening
Obtaining raw data, whether by measurement, download or retrieval.
We compiled 8,685 unique binary human protein pairs based on cross-linked peptides from 20 XLMS studieswhere the paper describes this · verbatim
AI
Predict pair structures with AlphaFold-Multimer
Running a trained model over new data to predict, classify or score. The AI stood in for physical experiment.
We used a locally installed version of ColabFold v 1.5.2 to run AF-M.where the paper describes this · verbatim
no AI
Filter to contact-positive pairs and reduce homology
Cleaning, filtering, normalising or labelling data already obtained.
Only these “contact positive” (“C+”) pairs are subject to downstream analysis.where the paper describes this · verbatim
AI
Build structural and omics feature vectors for each pair
Encoding data into features, descriptors, embeddings or graphs. The AI stood in for manual curation.
we utilized predictions from the DeepLoc 2.0 protein sequence transformer modelwhere the paper describes this · verbatim
AI
Train random forest classifiers with feature pruning
Fitting model parameters, including fine-tuning an existing model. The AI stood in for conventional algorithm.
We used the RandomForestClassifier package from the scikit-learn to train our random forest (RF) models.where the paper describes this · verbatim
AI
Score genome maintenance matrix and proteome-wide screens with SPOC
Running a trained model over new data to predict, classify or score. The AI stood in for conventional algorithm.
Having developed SPOC, we used it to score all possible pairwise interactions within a biological pathway.where the paper describes this · verbatim
no AI
Threshold scored pairs and publish the predictome
Reducing a candidate set by filtering or ranking, in a single pass.
Across all GM pairs, 1,151 (2.8%) had SPOC scores > 0.33where the paper describes this · verbatim
no AI
Evaluate classifiers on held-out sets and ranking experiments
Testing outputs against ground truth. Its result feeds back into an earlier step.
we evaluated its performance on the 25% testing data held backwhere 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 reported resource is a set of AlphaFold-Multimer structure predictions plus a trained classifier that scores them; the paper's central outputs are model outputs
we used machine learning on curated datasets to train a Structure Prediction and Omics informed Classifier (SPOC)where the paper describes this · verbatim
541 (57.2%) had DockQ > 0.23, the CAPRI cutoff for acceptable model qualitywhere the paper describes this · verbatim
The genome maintenance dataset is available for interactive browsing online at predictomes.org.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.
- Version of SPOC (Structure Prediction and Omics informed Classifier)Which version of the model was used is not stated.
- Version of Structural classifierWhich version of the model was used is not stated.
- Version of SPOCMatchedWhich version of the model was used is not stated.
About this article
Record aix-00151, 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