TL;DR
This paper explores applications of AlphaFold beyond its primary function of predicting protein structures. It demonstrates that AlphaFold's internal representations can accurately predict protein stability changes from mutations, a task its confidence scores fail at. Furthermore, it introduces a novel framework for estimating protein "designability" – a measure of functional robustness – by combining AlphaFold with protein design methods.
What Problem Does This Paper Address?
While AlphaFold has revolutionized protein structure prediction for natural sequences, its utility for designed or engineered proteins remains less clear. These sequences can be unstable, and predicting their stability alongside their structure is crucial for their function. AlphaFold's confidence scores don't correlate well with experimentally measured stability changes, leaving a gap in understanding how well it generalizes to non-natural sequences and how to assess the robustness of designed proteins. The paper also tackles the challenge of quantifying "designability," a concept linked to protein robustness and disease association, which has been difficult to estimate for real proteins.
Methods
The researchers employed several key techniques:
- Feature Extraction from AlphaFold Representations: They hypothesized that AlphaFold's internal learned representations, not just its final confidence scores, might contain information about protein stability. They extracted these features from the "single representation" of AlphaFold models for wild-type and mutant sequences.
- Multilayer Perceptron (MLP) Regression: A simple MLP model was trained using the extracted AlphaFold features to predict experimentally measured stability changes (ΔΔG and ΔΔTm) of point mutations.
- Protein Design with ProDCoNN: They utilized their previously developed ProDCoNN method, a deep neural network for inverse protein folding (IPF), to design sequences predicted to fold into specific target structures.
- Sequential Monte Carlo (SMC): To estimate designability, they employed an SMC strategy. This method is used to sample sequences and estimate the "partition function" of foldable sequences for a given structure.
- AlphaFold for Foldability Assessment: AlphaFold was used to predict the structures of designed sequences. Sequences predicted to fold with an RMSD below 4Å to the target structure were considered "foldable."
Datasets
- Point Mutation Stability Data: The study used data from the FireProtDB database, containing experimentally measured Gibbs free energy changes (ΔΔG) and changes in melting temperatures (ΔΔTm) for point mutations. They curated datasets for training and validation, including both wild-type and mutant sequences.
- Protein Structures for Design: They selected 9 protein structures from the SCOPe database, representing various fold classes, for designing sequences using ProDCoNN.
- Real Protein Structures for Designability Estimation: Five proteins with lengths ranging from 89 to 166 residues were used to estimate designability. This included chain A of FLT3 ligand (PDB ID: 1ETE).
Key Insights and Ideas
- AlphaFold's Hidden Knowledge: The most striking insight is that AlphaFold's internal representations, not just its confidence scores, encode valuable information about protein stability. This suggests AlphaFold has learned deeper physical principles of protein folding than initially apparent.
- Generalizability Beyond Natural Sequences: The success in predicting stability changes for point mutations indicates AlphaFold's potential generalizability to engineered sequences, even those not well-represented in its training data.
- Quantifying Designability: The paper introduces a practical method to estimate "designability" for real protein structures. This metric is crucial for understanding protein functional robustness and has implications for disease association.
- Distinguishing Foldable vs. Non-Foldable Designed Sequences: AlphaFold can effectively distinguish between designed sequences that are likely to fold correctly and those that are not, providing a valuable filtering step in protein design.
- MSA vs. Foldable Sequences: Comparing sequence logos of foldable designed sequences with those from traditional multiple sequence alignments (MSAs) reveals differences, suggesting that conserved residues in MSAs might be for function rather than solely for structural stability.
Key Results
- Stability Prediction Accuracy: The MLP model using AlphaFold features achieved a Pearson's correlation coefficient of 0.58 for predicting ΔΔG, outperforming previous methods and significantly exceeding the correlation observed with AlphaFold's confidence scores (which were near zero).
- Designability Estimation: For the first time, the designability of a real protein structure (chain A of FLT3 ligand, 134 residues) was estimated to be 3.12 ± 2.14E85.
- ProDCoNN Validation: The study validated ProDCoNN's ability to design sequences that fold to target structures with RMSDs less than 4Å, with AlphaFold serving as a crucial tool for assessing foldability.
Limitations and Future Work
- Stability Prediction Nuances: While stability changes can be predicted, predicting exact stability remains challenging. The current model predicts binary outcomes (stable/unstable) or changes, not absolute stability.
- Dataset Size: The authors suggest that increasing the size of the point mutation dataset could further improve prediction accuracy.
- Feature Engineering: Incorporating "pair representation" from AlphaFold and information from neighboring residues of the mutated site could enhance stability prediction performance.
- Computational Mutagenesis: The paper proposes using AlphaFold for computational mutagenesis to identify "minimum folding elements" (MFEs) and better understand sequence-structure relationships.
- Exploring Designability: Future work could involve estimating designability for a wider range of proteins and exploring its correlation with functional properties and disease.
Summary updated