TL;DR
This paper compares experimental and computational methods for predicting whether a genetic variant in the PTEN protein causes disease. They found that experimental methods, especially those measuring protein function, are more accurate than computational predictions, and that many disease-causing variants lead to PTEN's instability and subsequent degradation.
What Problem Does This Paper Address?
With the decreasing cost of genome sequencing, we're generating vast amounts of genetic data. A major challenge is interpreting these genetic variations, particularly missense variants (where a single DNA change alters a protein's building block), to determine if they cause disease and, if so, how. This is crucial for diagnosing genetic disorders. The authors focus on the PTEN protein, a known tumor suppressor, as a model system to evaluate different approaches for classifying variants.
Methods
The study compares four methods for classifying PTEN variants:
- Phosphatase-MAVE: An experimental method that uses multiplexed assays to measure the effect of many PTEN variants on its lipid phosphatase activity.
- VAMP-seq: Another experimental method that uses massively parallel sequencing to assess variant abundance in cells, indirectly reflecting protein stability and degradation.
- Evolutionary Sequence Energy (Ẽ): A computational method that uses sequence conservation to predict variant effects.
- Rosetta (ΔΔG): A computational method that predicts changes in protein stability caused by variants.
They also used a two-dimensional analysis to integrate stability and "activity" (function or abundance) scores to visualize variant effects.
Datasets
- ClinVar: A public database of variants and their clinical significance was used to identify 87 pathogenic missense variants in PTEN.
- gnomAD: A large database of human genetic variation was used to identify common variants likely to be benign.
- The study focused on a subset of 71 variants for which all four methods provided data, including 42 known pathogenic variants and 29 from gnomAD.
Key Insights and Ideas
- Experimental vs. Computational Performance: Experimental methods generally outperform computational predictions in classifying variant effects. This suggests that direct measurement of protein behavior in a cellular context is currently more reliable than purely in-silico predictions.
- Function vs. Stability: Methods that assess a broader range of functional consequences (like enzyme activity or cellular abundance) are more effective than those solely focused on predicting protein stability. This highlights that disease mechanisms can be complex and not solely reducible to stability.
- Loss of Stability as a Major Driver: A significant portion of disease-causing PTEN variants (~60%) appear to lead to disease by destabilizing the protein, causing it to be degraded by the cell's quality control machinery. This suggests that protein instability is a key mechanism underlying PTEN-related diseases.
- Experimental Data is Key, but Predictions Offer a Starting Point: While experimental data provides better separation of pathogenic from benign variants, computational predictions can still offer a valuable and cost-effective starting point for variant assessment.
Key Results
- All four methods showed a clear difference in scores between gnomAD (likely benign) and pathogenic variants.
- Receiver Operating Characteristic (ROC) curve analysis showed that experimental methods had higher Area Under the Curve (AUC) values than computational methods.
- Phosphatase-MAVE (experimental, activity-focused) had the highest AUC (0.94).
- VAMP-seq (experimental, stability-focused) had an AUC of 0.81.
- Evolutionary Sequence Energy (computational, activity-focused) had an AUC of 0.86.
- Rosetta ΔΔG (computational, stability-focused) had the lowest AUC (0.73).
- The study found that approximately 60% of disease-causing PTEN variants lead to loss of function by destabilizing the protein.
Limitations and Future Work
- Caveats on Datasets: The authors acknowledge that pathogenic variants from ClinVar might encompass different disease severities (e.g., ASD vs. PHTS), and rare variants from gnomAD are used as a proxy for benign variants, meaning some might still contribute to disease.
- Generalizability: While PTEN is used as a model, the findings suggest that assessing protein stability might be a generalizable approach for identifying disease-causing variants in other genes.
- Future Directions: The authors hypothesize that experiments and prediction methods probing protein stability can be extended to find similar mechanisms in other genes. They also note that while experimental methods currently perform better, computational approaches are improving and offer a cost-effective alternative.
Summary updated