Enhancing protein structure prediction accuracy by prioritizing important residues using protein language models

Qinghua Cui, Yu Liu, B. S. KangbioRxiv 2025 · DOI

TL;DR

Researchers have developed "i-Fold," a new protein structure prediction model that improves upon AlphaFold2 by incorporating residue-specific importance scores from protein language models. This approach allows the model to focus computational resources on critical amino acid residues, leading to more accurate predictions, especially for challenging protein types, without significantly increasing computational cost.

What Problem Does This Paper Address?

Predicting the 3D structure of proteins from their amino acid sequences is a fundamental but incredibly difficult problem in computational biology. While AlphaFold2 has revolutionized this field, achieving remarkable accuracy, it still exhibits systematic errors and struggles with certain types of proteins, such as those with limited evolutionary information (orphan proteins) or complex structures like ribosomal and membrane proteins. The researchers hypothesize that not all amino acids are equally important for a protein's structure and function, and that by prioritizing the more critical residues, prediction accuracy can be enhanced.

Methods

The core innovation of this work is the development of i-Fold (importanceFold), which builds upon the AlphaFold2 architecture. The key modification is the integration of Residue Importance Scores (RIS) derived from the ESM protein language model. These RIS values are used as dynamic positional weights during the training process.

Here's a breakdown of the approach:

  • Protein Language Model (ESM-2): This model is used to generate an amino acid probability matrix for each protein sequence.
  • Information Entropy: From this matrix, information entropy is calculated for each residue to quantify its "mutation potential." The rationale is that conserved sites with lower mutation potential are more likely to be functionally critical and structurally important.
  • Residue Importance Score (RIS): Based on the information entropy, a RIS is computed, assigning higher weights to residues with lower mutation potential.
  • Weighted Loss Function (FAPE<sub>RIS</sub>): The standard Frame-Aligned Point Error (FAPE) loss function used in AlphaFold2 is modified by weighting it with the calculated RIS. This means that errors in predicting the positions of high-RIS residues incur a larger penalty during training.
  • Dynamic Weighting: This weighting is applied dynamically during structure prediction, guiding the model's computational attention towards functionally critical residues and regions.

Datasets

  • Training Data: The researchers utilized a large dataset of 185,158 protein structure files and their corresponding Multiple Sequence Alignments (MSAs) originally used for training AlphaFold2.
  • Benchmark Test Set: A random sample of 3,599 sequences from the training data was selected for evaluation.
  • Independent Test Set: A completely independent set of 167 recently released protein structures from the Protein Data Bank (PDB) was curated for further validation. MSAs for this set were generated using JackHMMER and HHBlits.

Key Insights and Ideas

  • Differential Residue Importance: The fundamental idea that individual amino acids contribute differently to protein structure and function is key. By quantifying this importance, we can guide prediction models more effectively.
  • Leveraging Protein Language Models for Structure Prediction: Protein language models like ESM-2, trained on vast amounts of sequence data, capture rich evolutionary and functional information that can be directly applied to improve structure prediction, even without explicit structural input.
  • RIS as a Proxy for Functional Criticality: The inverse relationship between mutation potential (information entropy) and residue importance is a clever way to identify critical residues. Conserved residues are often crucial for structural integrity and function.
  • Targeted Computational Attention: The weighted loss function effectively directs the model's "attention" to the most important parts of the protein, analogous to how a human expert might focus on key features.
  • Complementary to MSA: The RIS approach can compensate for limitations in traditional Multiple Sequence Alignments (MSAs), particularly for proteins with sparse evolutionary data (low-homology sequences).

Key Results

  • Improved Accuracy: i-Fold achieved a statistically significant reduction in median Root Mean Square Deviation (r.m.s.d.) compared to AlphaFold2 on the benchmark test set (1.482 Å vs. 1.794 Å).
  • Higher Success Rate: The prediction success rate (defined as r.m.s.d. ≤ 2.0 Å) increased from 55.1% for AlphaFold2 to 62.7% for i-Fold (a 7.6% absolute improvement).
  • Performance on Challenging Proteins: i-Fold showed particular improvements for proteins that are typically difficult for AlphaFold2, including ribosomal proteins, membrane proteins, and orphan proteins.
  • Validation on Independent Set: These improvements were replicated on a completely independent test set, with i-Fold achieving a success rate of 49.7% compared to AlphaFold2's 43.7% (a 6.0% absolute improvement).
  • Computational Cost: The retraining of AlphaFold2 with the RIS-weighted loss function did not substantially increase computational cost, with training times remaining comparable.
  • Gene Ontology Enrichment: Proteins better predicted by i-Fold were enriched in terms related to translation, ribosomes, and membrane-associated biological processes, suggesting its strength in these areas.

Limitations and Future Work

  • Limitations: The paper acknowledges that while i-Fold improves upon AlphaFold2, there's still room for enhancement. The study primarily focuses on single-chain protein prediction.
  • Future Work:
    • Multi-dimensional Importance: Future developments could incorporate additional physicochemical factors (e.g., hydrophobic gradients, electrostatic complementarity) alongside evolutionary conservation to create a more comprehensive importance scoring system.
  • Complexes: Extending the importance-weighted framework to protein-protein and protein-ligand interactions is a promising direction, given the growing importance of biomolecular complexes.
  • Adaptability to Other Tools: The concept of residue importance weighting could be adapted to enhance other AI-based protein structure prediction tools beyond AlphaFold2.

Summary updated

Abstract

Abstract Accurate prediction of protein tertiary structures from amino acid sequences remains a fundamental challenge in computational biology. Although AlphaFold2 represents a major advance, systematic discrepancies persist between its predictions and experimentally determined structures. Given that individual residues contribute differentially to protein function, we hypothesized that incorporating residue-specific importance metrics could improve prediction accuracy. Here, we develop i -Fold ( importance Fold), an enhanced neural architecture enhances the AlphaFold2 architecture by integrating protein language model ESM-derived residue importance scores (RIS) as dynamic positional weights during training. Our approach dynamically weights amino acids using RIS during structure prediction, thereby directing computational attention toward functionally critical residues and regions. Evaluation on a benchmark test set of 3,559 protein structures reveals that i -Fold significantly improves accuracy (reduction in r.m.s.d., p = 0) and achieves a higher prediction success rate (7.6% improvement: 55.1% → 62.7%). Notably, i -Fold demonstrates particular improvements for targets that are typically challenging for AlphaFold2, including ribosomal proteins, membrane proteins, and orphan proteins. Consistent results were obtained on a completely independent test set of 167 recently released protein structures, where i -Fold again exhibited a higher prediction success rate (6.0% improvement: 43.7% → 49.7%) compared to AlphaFold2. Our findings indicate that explicit integration of RIS can advance the state-of-the-art in protein structure prediction, producing more accurate and generalizable models without substantially increasing computational cost.
BibTeX
@misc{cui2025enhancing,
  author = {Qinghua Cui and Yu Liu and B. S. Kang},
  title = {{Enhancing protein structure prediction accuracy by prioritizing important residues using protein language models}},
  year = {2025},
  doi = {10.1101/2025.09.28.679101},
  howpublished = {bioRxiv},
}
Figure