TL;DR
This paper introduces ETDBN, a novel generative model that jointly simulates protein sequence and structure evolution. By modeling dihedral angles as a diffusion process on a torus and coupling it with amino acid changes, ETDBN captures both gradual and abrupt structural shifts, offering a more realistic view of evolution and enabling new insights into sequence-structure relationships.
What Problem Does This Paper Address?
Understanding protein evolution traditionally relies on analyzing amino acid sequences. However, proteins' three-dimensional structures are crucial for their function, and their evolution is intrinsically linked to sequence changes. Previous models often struggled to integrate structural information effectively, relying on heuristic methods or fixed alignments that ignore uncertainty. This paper tackles the challenge of building a probabilistic model that can simultaneously capture the evolutionary dynamics of both protein sequences and their local 3D structures, specifically focusing on the local dependencies between aligned sites in homologous proteins. The goal is to create a more realistic and interpretable model of evolution that can also be used for tasks like predicting protein structures.
Methods
The core of the ETDBN model is a generative, evolutionary model that operates on pairs of homologous proteins. It treats the evolutionary trajectory of protein structures as a random walk in dihedral angle space.
Key methodological innovations include:
- Novel Angular Diffusion Process: Dihedral angles (phi and psi), which describe local protein backbone conformation, are modeled using a new Wrapped Normal (WN) diffusion process on a two-dimensional torus. This is crucial because dihedral angles are periodic, unlike standard diffusion processes.
- Coupled Sequence and Structure Evolution: The model explicitly links amino acid changes to changes in dihedral angles. This allows for modeling both "smooth" conformational changes and "catastrophic" conformational jumps, conditioned on specific amino acid mutations.
- Dynamic Bayesian Network (DBN) / Hidden Markov Model (HMM) Architecture: The model is structured as a DBN, visualized as an HMM. Each "hidden node" in the HMM represents an aligned site in a protein pair and encodes an evolutionary state. These states, in turn, specify distributions over evolutionary "regimes" that govern the specific processes of amino acid, dihedral angle, and secondary structure evolution.
- Evolutionary Regimes: The model uses two types of evolutionary regimes (constant and jump) to capture different modes of evolution. Constant evolution implies gradual changes within a regime, while jump evolution allows for abrupt shifts.
- Statistical Alignment: The model incorporates a modified TKF92 alignment HMM to account for insertion and deletion (indel) events, thus handling alignment uncertainty.
Datasets
- Training Data: 1200 protein pairs (2400 proteins) from the HOMSTRAD database were used for training.
- Test Data: 38 protein pairs (76 proteins) from the HOMSTRAD database were used for evaluation.
- PDB Coordinates: Protein Data Bank (PDB) coordinates were used to compute dihedral angles.
- DSSP: The DSSP algorithm was used to annotate secondary structure.
The paper does not introduce new benchmark datasets but utilizes existing, well-established resources like HOMSTRAD.
Key Insights and Ideas
- Dihedral angles hold significant evolutionary information: The model demonstrates that dihedral angle data alone can provide evolutionary time estimates comparable to those derived from amino acid sequences, suggesting that structural changes are as informative as sequence changes over evolutionary timescales.
- Coupling sequence and structure is key to realism: The ability to model "smooth" and "catastrophic" structural changes conditioned on amino acid mutations provides a more biologically plausible representation of protein evolution than models that treat these processes independently.
- Evolutionary regimes offer interpretable insights: The concept of distinct evolutionary regimes (constant vs. jump) and their association with specific amino acid changes (e.g., a motif involving glycine and asparagine linked to jump evolution) provides a powerful lens for understanding sequence-structure co-evolution.
- Generative nature enables powerful predictions: The model's generative capability allows it to predict missing structural data (dihedral angles) from various combinations of observed data (amino acid sequences, homologous structures, etc.), highlighting its utility in homology modeling and structure prediction.
Key Results
- Consistent Evolutionary Time Estimates: Estimates of evolutionary time derived from amino acid sequences and dihedral angles were found to be consistent, with a paired t-test failing to reject the null hypothesis of no difference.
- Predictive Accuracy: The model achieved good predictive accuracy for dihedral angles, with accuracy increasing substantially as more data types (amino acid sequences, homologous structures) were incorporated. For instance, using both amino acid sequences and homologous dihedral angles (combination 5) yielded better predictive accuracy than using only amino acid sequences or secondary structures.
- Alignment Accuracy: ETDBN demonstrated comparable or superior alignment accuracy to other methods when evaluated on simulated data and real data, particularly when incorporating multiple data types.
- Identification of Evolutionary Motifs: The model successfully identified an interpretable evolutionary motif associated with jump evolution, involving specific amino acid exchanges (e.g., glutamate to glycine) that correlate with significant changes in dihedral angles.
Limitations and Future Work
- Local Model: The current model focuses on local dependencies and does not capture global protein features like radius of gyration.
- WN Diffusion Limitations: While the Wrapped Normal diffusion captures angular drift well, it is less adept at modeling "angular shift" (large changes in dihedral angles), and the jump model only allows for jumps between two regimes per hidden state. A more realistic model would allow for arbitrary switches between regimes.
- Pairwise Model: The implemented model is pairwise. Generalizing it to a phylogeny with more than two sequences is theoretically possible but computationally challenging, especially for continuous dihedral angle states.
- Computational Cost: Inference scales O(|pa||pb|h²) when the alignment is unobserved, which can be computationally intensive for long sequences. Fixing the alignment reduces this to O(mh²), but this ignores alignment uncertainty.
- Future Work: The authors suggest further research into identifying more evolutionary motifs, improving homology modeling predictions, and developing more realistic models that can capture arbitrary switches between evolutionary regimes. Extending the model to phylogenetic trees and exploring more efficient algorithms for marginalizing continuous ancestral states are also key future directions.
Summary updated