Towards coevolution-aware ancestral sequence reconstruction

Technical Summary: Towards Coevolution-aware Ancestral Sequence Reconstruction

Problem Statement
Ancestral Sequence Reconstruction (ASR) is a standard tool for studying molecular evolution and protein function. However, conventional ASR methods rely on the assumption that amino acid sites evolve independently. This simplification neglects epistatic constraints—interactions between residues that are critical for maintaining protein structure, stability, and function. The authors argue that this site-independence leads to two primary issues:

1. Over-idealization: Maximum-a-posteriori (MAP) reconstructions often concentrate probability into a single sequence that may be "over-idealized," exhibiting unusually high stability or promiscuity compared to extant proteins.
2. Implausibility: Sampling directly from site-independent posterior distributions can generate sequences that are statistically plausible under the independent model but biologically implausible or non-functional because they violate residue-residue constraints.
   Furthermore, existing ASR methods are rarely benchmarked against "Ground Truth" (GT) data generated under realistic epistatic constraints, as experimental recovery of ancient sequences is limited to very recent timescales.

Methodology
The paper proposes a coevolution-aware ASR framework that integrates standard phylogenetic inference with Direct Coupling Analysis (DCA). The methodology consists of three main components:

1. Standard Site-Independent Reconstruction:
   The process begins with a standard Bayesian ASR (using the Yang algorithm) to infer the per-site posterior distribution of amino acids at the ancestral node. This yields a set of candidate ancestral sequences ($S_{anc}$) sampled from the site-independent posterior. Crucially, this step preserves the marginal statistics (site-wise frequencies) inferred from the phylogeny.

2. Coevolutionary Reshuffling (The Core Innovation):
   To enforce epistatic constraints without the computational intractability of exact inference on a phylogeny under a DCA prior, the authors introduce a reshuffling procedure:

   • Candidate sequences are iteratively optimized via random amino acid swaps at specific sites.
   • Swaps are accepted based on a Metropolis criterion using the DCA energy function ($E_{DCA}$), which includes site-specific fields ($h_i$) and pairwise couplings ($J_{ij}$).
   • This process minimizes the DCA energy of the sequences while strictly preserving the per-site amino acid frequencies derived from the initial site-independent reconstruction.
   • The result is a set of coevolution-aware candidates ($S^{DCA}_{anc}$) that are phylogenetically consistent (matching the site marginals) but structurally and functionally constrained by the DCA model.

3. Benchmarking Framework:
   To rigorously evaluate the method, the authors developed a forward-evolution simulator based on the DCA model.

   • Starting from known "Ground Truth" root sequences ($s_{GT}$) selected from extant data, sequences are evolved along a fixed phylogenetic tree using a discrete-time Markov chain that respects DCA energy constraints.
   • This simulator generates synthetic leaf data that retains both phylogenetic correlations and realistic epistatic constraints, allowing for a direct comparison between reconstructed ancestors and the known $s_{GT}$.
   • The study tests various evolutionary regimes by scaling branch lengths ($\mu_{gen}$) and selecting root sequences with varying mutability (quantified by Context-Dependent Entropy, CDE).

Key Results
The framework was applied to $\beta$-lactamase and DNA-binding domain families. Key findings include:

• Impact of Root Mutability: Reconstruction accuracy is highly dependent on the mutability of the ancestral sequence. For "weakly mutable" roots (highly constrained by epistasis), the coevolution-aware method significantly outperforms site-independent methods. For "highly mutable" roots, the advantage is less pronounced but the method still produces biologically plausible candidates.
• Hamming Distance vs. DCA Energy:
  • Standard MAP reconstructions often yield sequences with lower Hamming distances to the GT than simple consensus sequences but exhibit artificially low DCA energies, suggesting they are over-optimized for stability and may not reflect the true ancestral state.
  • Site-independent sampling produces sequences with high DCA energies, often falling outside the functional region of sequence space.
  • The coevolution-aware candidates ($S^{DCA}_{anc}$) achieve a balance: they maintain Hamming distances comparable to MAP estimates while exhibiting DCA energies consistent with natural extant families.
• Structural Plausibility: Candidates from the coevolution-aware set show superior structural properties compared to site-independent samples. They exhibit higher predicted local distance difference test (pLDDT) scores and lower Root Mean Square Deviation (RMSD) when compared to high-accuracy AlphaFold structures of the ground truth.
• Identification of Evolutionary Regimes: The authors identify an intermediate evolutionary timescale ($\mu_{gen} \in [1, 100]$) where phylogenetic inference provides non-trivial information beyond simple consensus, and where the benefits of incorporating coevolution are most detectable.

Significance and Claims
The paper claims to provide a practical route for ancestral reconstructions that better reflect the coupled nature of protein evolution. Its significance lies in:

• Bridging the Gap: It offers a computationally tractable solution that bridges the gap between single-sequence MAP reconstruction (which ignores epistasis) and unconstrained posterior sampling (which ignores structural constraints).
• Biological Realism: By enforcing DCA constraints, the method generates ensembles of candidate ancestors that are not only phylogenetically consistent but also statistically compatible with natural protein families and structurally viable.
• Uncertainty Quantification: The approach advocates for sampling an ensemble of plausible ancestors rather than relying on a single point estimate, providing a more realistic representation of ancestral uncertainty and diversity.
• Validation: The work demonstrates that the benefits of coevolution-aware reconstruction are only detectable when the underlying data (or simulation) reflects true epistatic constraints, highlighting the necessity of such models for deep-time evolutionary studies.

The authors conclude that while their method currently applies to single nodes independently, it represents a principled step toward reconstructing ancestors that are both evolutionarily coherent and biologically authentic, avoiding the systematic biases of traditional site-independent approaches.