Discovering functionally important sites in proteins

Matteo Cagiada, Sandro Bottaro, Søren Lindemose, Signe M. Schenstrøm, Amelie Stein, Rasmus Hartmann‐Petersen, Kresten Lindorff‐Larsen

TL;DR

This paper introduces a new machine learning approach to identify functionally important amino acid residues in proteins. By combining evolutionary sequence conservation with biophysical models of protein stability, their method can distinguish between residues critical for function and those important for structural integrity, a long-standing challenge. This tool has broad applications, from understanding disease mechanisms to protein engineering.

What Problem Does This Paper Address?

Proteins are the workhorses of biology, and understanding which parts of a protein are crucial for its function is fundamental to biology, medicine, and biotechnology. Identifying these "functional sites" is challenging because many residues are conserved during evolution for reasons of structural stability rather than direct functional involvement. This makes it difficult to pinpoint residues critical for activities like catalysis, binding, or regulation. Existing methods often rely solely on sequence conservation, which can be misleading, or require extensive experimental characterization. The authors aim to bridge this gap by developing a computational method that can reliably identify functionally important sites by disentangling their roles in stability versus function.

Methods

The core of this work is a supervised machine learning model, specifically a gradient boosting classifier. The model is trained using a combination of features derived from protein sequences and structures:

  • Sequence Conservation: Evolutionary information scores (ΔΔE) calculated using GEMME, which estimate the evolutionary "distance" of a variant from the wild-type sequence.
  • Biophysical Stability: Predicted changes in thermodynamic protein stability (ΔΔG) calculated using Rosetta.
  • Physicochemical Properties: Hydrophobicity of amino acids.
  • Structural Context: Weighted contact number (WCN) and averaged neighbor scores for stability and conservation.

The model is trained on data from multiplexed assays of variant effects (MAVEs) that measure both protein abundance and functional activity. Variants are classified into four categories: WT-like, total loss, "stable but inactive" (SBI), and low abundance/high activity. The key innovation is using the SBI class (high abundance, low activity) to identify residues directly involved in function, as opposed to those whose loss of function is due to instability.

Datasets

The training dataset consists of paired MAVE data for three proteins: NUDT15, PTEN, and CYP2C9, totaling 9945 variants at 923 positions. These MAVEs provided both abundance and functional readouts.

For validation, the authors used:

  • Independent MAVE datasets: Data from GRB2 SH3 domain and PSD95 PDZ3 domain (MAVEs probing abundance and binding).
  • High-throughput experiment: Data on alkaline phosphatase (PafA) catalytic efficiency.
  • Previously published models: Comparison with existing functional site prediction methods on 10 enzymes.
  • Diverse protein sets: 20 enzymes from the Mechanism and Catalytic Site Atlas (M-CSA) and 5 non-enzymes from the Protein-Protein Interaction Affinity Database 2.0.
  • Prospective study: Experimental validation of predicted missense variants in HPRT1, a protein linked to Lesch-Nyhan disease.

Key Insights and Ideas

  • Disentangling Stability and Function: The central insight is that by requiring variants to maintain wild-type-like abundance (i.e., be "stable but inactive" or SBI), the model can effectively isolate residues critical for function from those important for structural integrity. This is a crucial step beyond simple sequence conservation.
  • Broad Applicability: The model can identify a wide range of functional sites, not just active sites, but also regulatory sites, binding interfaces, and allosteric sites, suggesting a more comprehensive view of protein functionality.
  • "Second-Shell" Residues: The method can identify residues that are not directly in the active site but play crucial roles in supporting catalysis or binding, often referred to as "second-shell" residues.
  • Mechanism of Disease Variants: The model can provide mechanistic insights into how missense variants cause disease, distinguishing between loss of function due to instability versus direct functional impairment.
  • Oligomeric vs. Monomeric Structures: Using dimer or oligomer structures for stability calculations can better capture the functional importance of residues at protein-protein interfaces, distinguishing roles in subunit interaction from intrinsic stability.

Key Results

  • The trained model achieves an average validation accuracy of 58% and a Matthews correlation coefficient of 0.57.
  • On the training data, the model correctly classifies 95% of variants, with 90% accuracy for identifying SBI variants.
  • For the training proteins, the model correctly identifies 116 out of 127 experimentally determined functional residues.
  • Validation on independent datasets showed good agreement with experimental data, with accuracies around 72% for variant classification and 68% for SBI variants.
  • The model outperforms previous methods for functional site prediction, achieving a true positive rate of 0.71 compared to 0.40 and 0.60 for earlier approaches.
  • In a prospective study on HPRT1 disease variants, the model correctly predicted the functional impact of 76% of variants, including all SBI variants.
  • The authors developed a faster stability prediction method (RaSP) that, when integrated, maintains comparable performance to the Rosetta-based model.

Limitations and Future Work

  • "Total-Loss" Positions: The model may miss residues that are important for both function and stability. In such cases, substitutions might lead to both loss of function and loss of abundance, and these are classified as "total-loss" rather than "functional." The authors note that approximately 20% of active site residues fall into this category.
  • Data Dependency: The model's performance is dependent on the availability and quality of MAVE data.
  • Computational Cost: While improved by RaSP, stability calculations can still be computationally intensive.

Future work could involve further improving the model with more diverse MAVE datasets, exploring its application to protein engineering tasks, and developing more refined methods for distinguishing between stability-related and function-related conservation. The authors also highlight the potential for using their model to guide the development of mechanism-based therapies, such as identifying variants that could be rescued by pharmacological chaperones.

Summary updated

Abstract

Proteins play important roles in biology, biotechnology and pharmacology, and missense variants are a common cause of disease. Discovering functionally important sites in proteins is a central but difficult problem because of the lack of large, systematic data sets. Sequence conservation can highlight residues that are functionally important but is often convoluted with a signal for preserving structural stability. We here present a machine learning method to predict functional sites by combining statistical models for protein sequences with biophysical models of stability. We train the model using multiplexed experimental data on variant effects and validate it broadly. We show how the model can be used to discover active sites, as well as regulatory and binding sites. We illustrate the utility of the model by prospective prediction and subsequent experimental validation on the functional consequences of missense variants in HPRT1 which may cause Lesch-Nyhan syndrome, and pinpoint the molecular mechanisms by which they cause disease.
BibTeX
@inproceedings{cagiada2023discovering,
  author = {Matteo Cagiada and Sandro Bottaro and Søren Lindemose and Signe M. Schenstrøm and Amelie Stein and Rasmus Hartmann‐Petersen and Kresten Lindorff‐Larsen},
  title = {{Discovering functionally important sites in proteins}},
  booktitle = {Nature Communications},
  year = {2023},
  doi = {10.1038/s41467-023-39909-0},
}
Figure