Check each claim against its experiments, reported results, scope, and stated limitations.Know what the science actually supports before you trust the answer.
Press Enter ↵ to start review
Explore by Goal
"The biology of mind bridges the sciences – concerned with the natural world – and the humanities – concerned with the meaning of human experience."
- Eric Kandel
Quick Explanation
Copied
What this paper adds (in 1 minute)
Introduces a multiplexed HDX-MS (mHDX-MS) pipeline to measure opening-energy distributions (ΔGopen) across 5,778 protein domains, enabling dataset-scale “protein energy landscape” learning.
Finds large hidden variation in conformational fluctuation landscapes even within the same fold.
Uses those correlations/models to design mutations that can stabilize locally unstable segments and increase “opening cooperativity,” while explicitly acknowledging that inferred ΔGopen values rely on assumptions (e.g., EX2 kinetics, back-exchange modeling, and pH-combination).
Long Explanation
Paper Review (visual-first, skeptical): Large-scale discovery, analysis, and design of protein energy landscapes
DOI: 10.1101/2025.03.20.644235 • Date (in manuscript): March 25, 2025 • Focus: multiplexed HDX-MS to map “opening” energy distributions at scale
1) Scope & headline results (what was actually measured)
Key dataset quantities (from the paper text)
Initial synthetic sequence pool: 15,715 domains; analyzed: 5,778 domains by mHDX-MS after processing attrition
Measurably stable subset used for many analyses: 3,590 domains (after removing low-stability cases)
Inferred observables: exchange-rate distributions (kHX) and approximate opening free-energy distributions (ΔGopen), plus a global stability proxy ΔGunfold
2) Visual: how the dataset is distributed across families
Counts are taken from the manuscript’s cooperativity-model parameter table (N per family) .
3) Visual: model predictability limits for opening cooperativity
The manuscript reports that best coefficients of determination for predicting family-normalized cooperativity on unseen data lie between 0.16 and 0.24, while predicting ΔGunfold is substantially easier (R² 0.40–0.53) .
4) How the method maps “energy landscape” to measurable quantities (core skepticism)
4.1 What is inferred vs what is directly observed
Directly measured: time-resolved isotopic envelopes (mass distributions) for many domains simultaneously using mHDX-MS across multiple D2O incubation timepoints at pH 6 and pH 9
Inferred: residue exchange-rate sets kHX via Bayesian inference under an EX2/independent-first-order kinetics model (and with additional modeling for back exchange and intrinsic chemical exchange rates)
Then converted: to approximate ΔGopen distributions using assumed relations between kHX and intrinsic kchem and an expected exchange behavior in the unfolded state
4.2 Assumptions that could bias the “landscape” reconstruction
EX2 regime independence and first-order kinetics; EX1-like profiles are filtered out
pH combination: it uses pH 6 and pH 9 data in one model, assuming that the pH shift does not change the underlying energy landscape (only modulates measurable kinetic range through kchem)
Back-exchange modeling: assumes a single overall back-exchange probability for all exchanged residues within each protein/timepoint
kchem approximation: ΔGopen depends on an approximate estimate of intrinsic chemical exchange rates; the paper derives a scaling factor from site-resolved NMR to reduce error vs using a single median
5) Method validation: does mHDX-MS reproduce known reference measurements?
Validation numbers shown are taken directly from the manuscript’s comparisons: kHX distribution RMSE (1.9-fold), ΔGopen RMSE (0.53 kcal/mol) across 13 domains vs HDX NMR, and global stability correlation R=0.78 between mHDX-MS and cDNA display proteolysis .
Skeptical interpretation
Agreement exists, but it is not “identity”: the paper reports systematic ΔG unfold shifts likely due to D2O stabilization and assay context differences, and uses these differences to bound interpretability
Key unknown: even if kHX and ΔGopen correlate with NMR/cDNA readouts, the mapping from inferred exchange to the exact molecular sub-states that “compose” ΔGopen remains non-unique because intact-domain MS does not resolve which residue rates correspond to which measured residue exchange events
6) What “opening cooperativity” means, operationally
Definition as implemented
They define an average opening free energy, ΔGavg, as the average of ΔGopen over exchangeable residues, with very fast residues set to a lower bound of 0 kcal/mol (so they contribute to lowering ΔGavg).
An empirical model predicts expected ΔGavg from ΔGunfold, hydrogen-bond donation fraction (fxn_hb), and net charge; the standardized residual defines a z-scored “opening cooperativity” metric.
Critical blind spot
Interpretational non-uniqueness: because ΔGavg is a coarse summary and residues’ exchange is not residue-resolved in intact MS, “cooperativity” is an operational proxy for distribution shape rather than a direct measurement of multi-state pathways; the paper itself acknowledges ΔGopen patterns for neighboring residues suggest (but do not definitively prove) secondary-structure-level openings
7) Visual: designed mutations improve cooperativity (but modestly)
What was done
They used family-specific models to propose double mutants predicted to increase cooperativity while preserving or increasing stability, then tested a subset of top-ranked designed mutants plus random double mutants in mHDX-MS libraries.
The excerpt explicitly states that mutations predicted to increase both properties were rare: only 6% and 4% of all possible double mutants for HHH_rd4_0518 and EEHEE_rd4_0871, respectively, and that designed variants typically increase cooperativity but can trade off stability. .
Skeptical reading
Low absolute predictive power implies design uncertainty: their cooperativity models have limited R², so designs are likely to succeed only probabilistically; the paper explicitly reports that ML results were not quantitatively accurate but improved exploration over random selection.
Mechanistic ambiguity: they propose plausible interaction-level explanations (hydrophobic interactions, backbone dipole effects, etc.) for specific top-performing mutants, but intact-domain HDX does not guarantee that the inferred “opening cooperativity” mechanistically maps uniquely to those residue-level explanations.
8) Reproducibility & transparency checklist (what you’d want before trusting the learned rules)
Code availability: the manuscript states the analyses can be found at three public GitHub repositories: mhdx_pipeline, hdxrate_pipeline, and mhdx_analysis.
Raw-ish evidence: it claims that raw MS data, extracted signals per protein, and Bayesian fits are available in the data availability section / datasets.
Model assumptions are explicitly enumerated, which is crucial for reproducibility; however, correctness still depends on whether these assumptions hold for every family and timepoint.
Practical reproducibility risks to watch
Multi-step automated pipelines: reproducibility depends on thresholds (idotp, q-value, PO score cutoffs) and on whether re-running yields identical path choices under numerical/implementation differences .
Parameter identifiability: because intact-domain MS does not map inferred rates to specific residues, reproducing the exact ΔGopen distributions involves the same identifiability constraints and modeling approximations. .
Scale + systematic acquisition: the paper attempts to measure opening-energy distributions for thousands of small domains under largely identical experimental conditions via mHDX-MS with automated inference.
Validation against orthogonal readouts using NMR (kHX, ΔGopen) and cDNA display proteolysis (ΔGunfold).
Explicit modeling assumptions and known limitations: the paper does not hide that ΔGopen is approximate and depends on EX2 kinetics, back-exchange simplifications, and intrinsic exchange-rate approximations.
Uncertainties / where the claims may be fragile
Mapping “landscape” to mechanistic sub-states: ΔGopen distributions compress multiple possible conformational excitations into an inferred opening-energy distribution; different microscopic ensembles could produce similar macroscopic ΔGopen under EX2-like fitting and kchem approximations.
Predictive ceiling: cooperativity prediction R² is low-to-moderate (0.16–0.24), implying that “learned rules” only partially capture determinants; further data and better measurement resolution likely change learned features.
Context dependence: measurements are in D2O and under specific pH/temperature conditions; energy scales and kinetics are likely not directly transferable across biochemical contexts without re-calibration.
What would most disprove the central “energy landscape learning” usefulness?
Demonstrating that ΔGopen-derived cooperativity metrics do not generalize when tested on new domains outside the trained topology families, or that re-running under slightly altered experimental conditions yields inconsistent ΔGopen/cooperativity rankings beyond the paper’s reported reproducibility bounds.
The novelty claim is dataset-scale multiplexed HDX-MS that infers ΔGopen distributions across thousands of small domains and then uses those inferred landscape variables for both correlational modeling and mutation design; this combination of throughput + landscape-level inference is unusually ambitious in the protein energy-landscape literature described in the manuscript.
Scientific Quality
80%
Scientific quality is strong on transparency, validation against orthogonal measurements, explicit assumptions, and automation/reproducibility claims; however, key outputs (ΔGopen) are inherently model-dependent approximations (EX2 kinetics, pH-combination invariance, back-exchange simplifications, and kchem estimation), and downstream cooperativity models show limited predictive power (R² ~0.16–0.24).
Study Generality
70%
Generality is supported by including multiple fold families and both designed and natural domains, but the learned rules are demonstrated primarily on small domains and rely on approximations specific to mHDX-MS; generalization to larger proteins, membrane proteins, multi-domain contexts, and cellular environments is not fully established.
Study Usefulness
80%
High usefulness as an experimental+computational framework: provides a scalable route to excited-state energetics and a dataset enabling ML modeling of conformational fluctuation landscapes; practical usefulness is tempered by the modest predictability of fine-grained cooperativity (suggesting continued model/measurement development).
Study Reproducibility
80%
Good reproducibility signals: public analysis code repositories (three GitHub repos) and claims of dataset availability including raw MS data, extracted signals, and Bayesian fits; remaining reproducibility risk comes from automated thresholding/path selection and model assumptions that must be reimplemented identically.
Explanatory Depth
70%
The paper provides a mechanistic-to-phenomenological bridge: defines an opening-cooperativity metric tied to a residual after stability/hydrogen-bond/charge expectations, and correlates features (e.g., proline content, compactness, charge in helix turns) with cooperativity; but feature correlations are modest and causal inference is limited by identifiability and by correlations/possible confounds in the feature set.
It will parse the paper’s reported family counts and cooperativity prediction ranges, then generate publication-style summary tables and calibration-style Plotly charts for dataset coverage and predictive ceilings.
Get emailed when your analysis is done!
We'll email you the results when your analysis is finished.
Hypothesis Graveyard
A simplistic claim that cooperativity is almost entirely determined by global stability (ΔGunfold) is inconsistent with the paper’s explicit statement that ΔGavg increases sub-linearly with ΔGunfold and that the cooperativity metric is constructed to be nearly orthogonal to stability across the dataset.
A strong claim that computed AlphaFold confidence (pLDDT) alone explains cooperativity should be disfavored because the paper reports no single feature dominates correlations and reports modest maximum PCC values and limited cooperativity prediction R², even when including language model embeddings and interpretable structural/sequence features.
Explore by Goal
Quick Explanation
Long Explanation
Top Data Sources
Ask a Follow-Up
Key Insight
Keep Exploring
Analysis 
Hypothesis Graveyard
Potential Experiments
Discussion