Statistical Genomics Foundry
Paper
domain/ancestral-reconstruction

Hochberg & Thornton 2017 — Reconstructing Ancient Proteins to Understand the Causes of Structure and Function

1. Citation

Hochberg GKA, Thornton JW. 2017. “Reconstructing Ancient Proteins to Understand the Causes of Structure and Function.” Annual Review of Biophysics 46:247–269. DOI: 10.1146/annurev-biophys-070816-033631. Open-access copy read: NIH Public Access author manuscript on PubMed Central, PMC6141191 (https://pmc.ncbi.nlm.nih.gov/articles/PMC6141191/). Access date 2026-07-03. Pinned to vol 46 / 2017 (not the newer Annu Rev Biophys ASR/energy-landscapes review).

2. Access note

Full text was readable via the free NIHPA author manuscript on PMC (no paywall boundary encountered for the main text). Reading was of the author-manuscript version; figure captions and exact page-level typesetting of the published ARR version were not consulted. Underlying work is Annual Reviews all-rights-reserved (an NIH author manuscript on PMC does not carry a Creative Commons license), so §7 uses own-words paraphrase (see §7 heading).

3. Thesis (1 sentence)

Reconstructing ancestral proteins and characterizing them experimentally (“vertical” evolutionary analysis) reveals the causes of modern proteins’ structure and function more powerfully than comparing present-day homologs (“horizontal” analysis), chiefly because it introduces sequence changes into the historical background in which they actually occurred.

4. Problem & context

Structural biology wants to know which sequence features cause which structural/functional properties. The naive approach — swapping residues or comparing states between extant homologs (horizontal comparison) — frequently produces nonfunctional or misleading results because the effect of a residue depends on the rest of the sequence (epistasis / context-dependence), and because divergent extant homologs differ at very many sites, most of them irrelevant to the function of interest. Ancestral sequence reconstruction (ASR) is offered as the corrective: infer the sequences of extinct ancestors on a phylogeny, synthesize and characterize them, and study change along specific branches.

5. Method / approach (what the review describes)

  • Vertical vs. horizontal analysis. Horizontal analysis compares/swaps between contemporary homologs. Vertical analysis reconstructs successive ancestors along a family phylogeny and characterizes them, isolating a functional/structural shift to a specific branch and thereby shrinking the set of candidate function-changing substitutions.
  • ASR inference mechanics. Inputs are a multiple sequence alignment, an evolutionary model, and a phylogeny; maximum-likelihood or Bayesian methods infer ancestral states. Each site is reconstructed individually. The best point estimate at a site is the maximum a posteriori (MAP) state; the string of MAP states across all sites is the MAP ancestor, colloquially the “maximum likelihood (ML) ancestor.”
  • Handling uncertainty. Some sites are reconstructed ambiguously (two or more states with non-negligible posterior probability). The review frames ASR output as an ensemble of plausible ancestral sequences with the MAP sequence as the single best estimate, and recommends experimentally characterizing alternate reconstructions that are less probable than the ML sequence but still statistically plausible, to test the robustness of functional conclusions.
  • Tool named: automated web servers now exist; PhyloBot (Hanson-Smith & Johnson 2016) is the example given.

Note: the review is a synthesis, not a step-by-step protocol; it does not state exact default parameters, a specific substitution model to use, or numeric cutoffs.

6. Key claims / findings (atomic)

  • Horizontal comparison often yields nonfunctional proteins because of epistasis — the phenotypic effect of a mutation depends on the genetic state at other sites.
  • A horizontal swap fails when a state from one homolog is incompatible with the receiving homolog’s sequence background — either because permissive residues needed to tolerate the state are absent, or because restrictive residues that block it are present.
  • Definitions: permissive mutations let a protein tolerate substitutions that would otherwise be nonfunctional; restrictive mutations make a protein unable to tolerate substitutions it otherwise would have tolerated.
  • Vertical (ancestral) analysis avoids epistatic incompatibilities by introducing function-changing substitutions into the historical background in which they arose; it thereby both overcomes and reveals epistasis.
  • Worked example (glucocorticoid/mineralocorticoid receptor family): restrictive substitutions sterically clash with the ancestral helix conformation, explaining why swaps between extant proteins failed.
  • The MAP/ML ancestor is a point estimate; ASR is better treated as an ensemble of plausible sequences. Some sites are inferred ambiguously.
  • Robustness finding (attributed to a study of several protein families): the qualitative functions of reconstructed proteins appear quite robust to sequence uncertainty, though the precise quantitative parameter values are not always robust. The stated reason: ambiguously reconstructed sites/states tend to be weakly constrained by function, whereas strongly-constrained states are reconstructed with confidence.
  • Critical-stance recommendation: users should neither uncritically accept the ML reconstruction as true nor dismiss a surprising inference as mere statistical/phylogenetic bias; sensitivity to sources of uncertainty and bias should be examined experimentally where possible.
  • Evolutionary/vertical analysis also explains structural features that are not functionally optimal (i.e., historical contingency, not optimization, can account for present structure).

7. Load-bearing statements — OWN-WORDS PARAPHRASE (restrictive license; functional strings excepted)

  • Epistasis is defined as context-dependence: a mutation’s phenotypic effect depends on the state at other sites. (Section: “Evolutionary Analysis Overcomes and Reveals Epistasis.”)
  • A horizontal swap produces a nonfunctional protein when a state (or set of states) is incompatible with the background it is placed into — because required permissive residues are absent or blocking restrictive residues are present. (Same section.)
  • ASR should be understood as supplying an ensemble of plausible ancestral sequences, with the MAP sequence as the single best estimate — not a single definitive reconstruction. (Section: “The Future of Retracing the Past for Structural Biology.”)
  • Robustness of functional inferences to sequence uncertainty is a particularly important part of the process, done by experimentally characterizing alternate reconstructions that are less likely than the ML sequence but still statistically plausible. (Same section.)
  • Across several protein families, qualitative functions of reconstructed proteins were quite robust to reconstruction uncertainty — but not always the precise quantitative parameter values — because ambiguous sites are weakly functionally constrained while strongly constrained states are reconstructed confidently. (Same section.)
  • Functional strings preserved verbatim (facts, not protected expression): MAP (maximum a posteriori), ML (maximum likelihood), permissive / restrictive (substitution types), PhyloBot (tool name).

8. Stated scope, assumptions, limitations (the source’s own caveats)

  • ASR gives point estimates with uncertainty; the MAP sequence is one estimate among an ensemble, and some sites are genuinely ambiguous.
  • Functional conclusions must be checked for robustness against that uncertainty; quantitative values in particular may not be robust.
  • The authors caution against two opposite errors: over-trusting the ML reconstruction and reflexively dismissing surprising results as artifacts.
  • The review is a conceptual synthesis of the field’s strategy, not a bench protocol; it does not supply parameter defaults or decision thresholds.

9. Failure modes / invalidity patterns (referee-relevant)

  • Epistatic incompatibility in horizontal swaps — transplanting a residue/state between extant homologs yields a nonfunctional protein when permissive residues are missing or restrictive residues are present in the recipient background. Symptom: loss of function after a swap that “should” have transferred function. The named remedy is vertical (ancestral) analysis.
  • Over-trusting the single ML/MAP ancestor — treating the per-site MAP string as the definitive ancestor rather than one estimate from an ensemble. Detector/remedy the review names: characterize alternate, less-likely-but-plausible reconstructions and check whether the functional conclusion holds.
  • Mis-attributing surprises to statistical uncertainty — dismissing an unexpected ancestral property as a reconstruction artifact without testing. Remedy: experimentally probe sensitivity to uncertainty/bias.
  • Quantitative fragility — precise quantitative parameter values of reconstructed proteins may not be robust to sequence uncertainty even when qualitative function is; conclusions resting on exact numbers are the risky ones. [summarizer-inferred] that this is a distinct usage caution (the review states the robustness asymmetry but does not spell out the downstream caution).

10. What the source does NOT address (confident silences)

  • No numeric posterior-probability cutoff defining an “ambiguous” site. The review describes ambiguity qualitatively (“two or more states with non-negligible posterior probabilities”) but names no threshold (e.g., no 0.2 / 0.8 rule) and no fixed number of alternate constructs to build.
  • No claim of systematic stability/thermostability bias. The review does not assert (or rebut) that reconstructed ancestors are systematically biased toward greater thermostability or toward most-probable/optimal states. It emphasizes robustness of qualitative function, not directional bias.
  • No systematic indel/gap reconstruction protocol. Insertions/deletions come up only in passing (e.g., a malate dehydrogenase example); no indel-aware or gap-inference method is prescribed.
  • No explicit statement that the ML ancestor is a non-existent combined sequence. See §12 Q1 — the review treats sites independently and flags per-site ambiguity, but does not spell out that independently combining per-site MAP states could produce a state combination that never coexisted.

11. Open questions / ambiguities the source leaves unresolved

  • Exactly how many alternate reconstructions to test, and at which sites, to declare a conclusion “robust” — left to practitioner judgment.
  • What posterior-probability level makes a site “ambiguous” — unspecified.
  • Whether/when the per-site-independent MAP assembly is itself a hazard (vs. per-site ambiguity) — not addressed head-on.

12. Guidance answers

Q1 — Epistasis problem for ASR: is the per-site-independent ML ancestral sequence possibly an untested/never-existing combination of states, and does the review state this? The review supplies the pieces but does not make this exact claim. It states that sites are reconstructed individually, that the ML/MAP ancestor is the string of per-site MAP states, and that some sites are ambiguous — and it recommends testing alternate reconstructions. But it does not explicitly say that independently combining the highest-probability state at each site can yield a combined sequence that never coexisted or is untested due to epistasis. The review’s epistasis discussion is aimed at horizontal swaps between extant homologs (states incompatible with a foreign background), not at the internal coherence of the ML ancestor. So: the “single ML sequence can be a non-existent state combination” framing is not directly stated by this source — treat that specific proposition as not-sourced-here (must be backed elsewhere or flagged as the skill’s own inference).

Q2 — Operational strategy for reconstruction uncertainty (alternative constructs, testing, report a range; thresholds; how many constructs; how “ambiguous” is defined): The review recommends treating ASR output as an ensemble of plausible sequences (MAP = best estimate) and experimentally characterizing alternate reconstructions that are less likely than the ML sequence but still statistically plausible, to test robustness of functional inferences. It does not name a posterior-probability cutoff for “ambiguous,” does not specify how many alternate constructs to build, and does not prescribe reporting a numeric “functional range.” Ambiguity is defined only qualitatively (a site with two or more states at non-negligible posterior probability). → Any specific posterior cutoff for “ambiguous” is convention, not sourced here; flag it.

Q3 — Are reconstructed ancestors systematically biased (e.g., toward stability / most-probable states), and how robust are functional conclusions to statistical uncertainty? No systematic stability/thermostability bias is claimed or discussed. On robustness: across several protein families, qualitative functions were quite robust to reconstruction uncertainty, but precise quantitative parameter values were not always robust; the stated reason is that ambiguous sites are weakly constrained by function while strongly constrained states are reconstructed with confidence. The authors add a critical-use directive: don’t uncritically accept the ML reconstruction, and don’t reflexively dismiss surprises as bias — test sensitivity experimentally.

Q4 — Named strategies/tools for epistasis-aware or indel-aware reconstruction the review endorses: The endorsed epistasis-handling strategy is vertical/ancestral analysis itself — introducing substitutions into the historical background in which they occurred, which sidesteps the incompatibilities seen in horizontal swaps. The only software tool named is PhyloBot (automated ASR web server, Hanson-Smith & Johnson 2016). No epistasis-aware model or indel-aware reconstruction tool/method is specifically endorsed; indels are mentioned only illustratively with no dedicated protocol. (Common tools such as PAML, FastML, GRASP were not found named in the portion read.)