Large-Scale Discovery, Analysis and Design of Protein Energy Landscapes

• p. 3 Results, 'mHDX-MS measurements are accurate'

  Global stabilities (ΔGunfold) measured for 4,464 domains by mHDX-MS (in D2O) and cDNA display proteolysis (in H2O) were also strongly correlated (r = 0.78; Extended Data Fig. 3)

• p. 22 Extended Data Fig. 3 caption and panel A annotation

  for 4,464 domains, with an overall Pearson correlation coefficient (r) of 0.75

• p. 22 Extended Data Fig. 3, panel A annotation

  w/o Cold-Shock domain N = 4,327, r = 0.78

• p. 22 Extended Data Fig. 3 legend, panel (A)

  values obtained from both methods for 4,464 domains, with an overall Pearson correlation coefficient (r) of 0.75

Judge's note: Viewed p3 and p22. The panel A annotation reads "All domains N =4,464, r = 0.75" and "w/o Cold-Shock domain N = 4,327, r = 0.78". The legend says "for 4,464 domains, with an overall Pearson correlation coefficient (r) of 0.75 ... Excluding these domains (N = 4,327) improves the correlation to r = 0.78". The main text pairs the full-set N with the subset correlation and mentions no exclusion. No other passage explains it.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 8 Results, 'Predicting stability and cooperativity'

  models trained to predict ΔGunfold performed better (R2, 0.40–0.53)

• p. 28 Extended Data Fig. 9b, left plot (ΔGunfold, second-generation models)

  R2 score

• p. 8 'Predicting stability and cooperativity'

  performed better (R2, 0.40–0.53)

• p. 8 Results, machine learning prediction

  (R2, 0.40–0.53)

Only the authors can say what is right here.

Judge's note: Extended Data Fig. 9 on p. 28. Panel B (clustered cross-validation, the 'unseen data' models whose cooperativity range 0.16–0.24 the text cites and which matches): best ΔGunfold R2 per family is about 0.40 (ββαββ), 0.41 (αββα), 0.44 (ααα), 0.43 (LysM); the y axis tops at 0.5 and nothing approaches it. Panel A (random folds): maximum about 0.50 (ααα). The caption clips only values below zero. Searched the text for 0.53: the only other occurrences are unrelated (kcal/mol reproducibility values). The lower bound 0.40 matches panel B; the upper bound 0.53 has no source in the article. Values read by eye against 0.1 gridlines.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 20 Extended Data Fig. 1 caption, panel E

  (E) PASTA_0085 with a lower factor (2.43)

• p. 20 Extended Data Fig. 1, panel E (arrow annotation)

  2.46

• p. 20 ED Fig. 1 caption (E)

  PASTA_0085 with a lower factor (2.43)

Only the authors can say what is right here.

Judge's note: Rendered panel E from the PDF at 6x: the annotation under the arrow is clearly '2.46'. The legend on the same page prints '(2.43)'. The other legend value, 3.22 for panel F, matches its panel annotation, and panel D is annotated 3.04 against 'near 10³' in the legend, so the legend is meant to repeat the plotted merge factor. One of the two values for PASTA_0085 is wrong; the files do not show which.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 23 Extended Data Fig. 4A, ααα labels

  62% 33% 32%

• p. 23 Extended Data Fig. 4A, ββαββ labels

  56% 28% 27%

• p. 23 Extended Data Fig. 4A, PASTA labels

  58% 27% 20%

Only the authors can say what is right here.

Judge's note: Checked on a 6x render of panels A and B. The other families' labels match panel B closely (LysM 72/45/27, αββα 42/25/22, WW 67/36/12, SH3 48/24/1, βαββ 49/19/17, Cold-Shock 34/18/7), so the labels and panel B use the same denominator, as the caption says ('The same three categories are expressed as percentages relative to the initial DNA set'). For ααα, ββαββ and the PASTA green value the labels are 4 to 12 points below the bars, and the panel A counts agree with panel B rather than with the labels. Bar heights are read by eye (about ±2 points), which is far smaller than the gaps. The arithmetic entries use approximate counts read from panel A. Which of the labels or the bars is right needs the authors.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 3 Fig. 1g, in-panel annotation

  5/13 shown (Supplementary Fig. 7)

• p. 3 Fig. 1 legend, panel g

  g, mHDX-MS results are consistent with HDX NMR (Supplementary Fig. 5).

• p. 3 Results, 'mHDX-MS measurements are accurate'

  all 13 domains are shown in Supplementary Fig. 5

• p. 3 Fig. 1 caption, g

  consistent with HDX NMR (Supplementary Fig. 5)

Judge's note: Viewed the Fig. 1g crop and the p3 page. The panel annotation says Supplementary Fig. 7. The legend, the Results text and the Methods ("13 proteins in multiple HDX NMR conditions (Supplementary Fig. 5)") all say Supplementary Fig. 5. Supplementary Fig. 7 is cited on p4 for the ΔGavg versus ΔGunfold relationship and the per-family model fits. The supplement is not in the files, so I could not open either figure, but the pointers contradict each other within the article and the panel label is the odd one out.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 31 Reporting Summary, Life sciences study design, 'Replication'

  The reproducibility of the mHDX-MS analysis was checked in Figure 1G, Figure S5 and Figure S23

• p. 3 Fig. 1f legend

  f, Reproducibility across Nrep libraries.

• p. 3 Fig. 1g legend

  g, mHDX-MS results are consistent with HDX NMR (Supplementary Fig. 5).

Only the authors can say what is right here.

Judge's note: The sentence is on the page 31 image. The Fig. 1 legend text reads 'f, Reproducibility across Nrep libraries. Inset: the MAD (Supplementary Fig. 4)' and 'g, mHDX-MS results are consistent with HDX NMR (Supplementary Fig. 5)'; the Methods section on reproducibility cites Supplementary Fig. 14. The highest supplementary figure cited in the article is 14. The summary's 'Figure 1G' and 'Figure S5' agree with each other (both the NMR comparison), so the authors may be using 'reproducibility' loosely for validation, or following an earlier numbering. The supplement is not among the files, so 'Figure S23' cannot be checked; the Supplementary Information would settle it. The Fig. 1 legend was checked in the text layer, not on the page 3 image.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 5 Fig. 2d, right scatter plot (green circle at ΔGunfold ≈ 5.5)

  LysM_3314

• p. 5 Fig. 2 legend, panel d

  Twelve proteins are highlighted and shown in the same colour in each plot.

Only the authors can say what is right here.

Judge's note: Viewed the Fig. 2d crop and the p5 page. There are twelve highlighted circles but eleven distinct names: EEHEE_rd4_0642, HHH_rd3_0062, LysM_3314 (green), 5FWB, LysM_3314 (red), PASTA_0056, EEHEE_rd4_0871, HHH_rd4_0557, HHH_rd4_0518, LysM_0873, LysM_1380 and 2CVR. Fig. 2a–c show LysM_3314 in green at ΔGunfold ≈ 6, so the red label is the wrong one. The true identity of the red point cannot be determined from the files.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 16 Methods, 'Deriving normalized cooperativity and family-normalized cooperativity'

  log(c) ~ normal(–0.3, 2) with c = exp(log(c)) and d ~ normal(–0.3, 2) with d = exp(log(d))

Judge's note: Page 16 image, right column, shows exactly 'log(c) ~ normal(-0.3, 2) with c = exp(log(c)) and d ~ normal(-0.3, 2) with d = exp(log(d)))'. As printed the d clause is circular (a normal prior on d followed by d = exp(log(d))); the parallel construction and identical hyperparameters make 'log(d) ~ normal(-0.3, 2)' the evident intended text. Notation error in the model specification; rated S2 rather than S1 because the intended meaning is recoverable from the c clause.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 17 Methods, Feature extraction

  ESM2 (650 million parameters)102, Unirep103 and SaProt104

• p. 17 Methods, Feature extraction

  we averaged the representations from layers 33–36 and flattened the result into an array of 2,560 features

Only the authors can say what is right here.

Judge's note: Rests on outside knowledge: the published ESM2 architectures (650M model = 33 layers, 1,280-dimensional embeddings; 3B model = 36 layers, 2,560-dimensional embeddings). The page 17 crop shows the text as quoted. Inside the paper, the first-generation models used both ESM2 sizes ('650 million and 3 billion parameters', page 18; Extended Data Fig. 9A lists ESM2_650M and ESM2_3B) while the second-generation panel (Fig. 9B) is labelled ESM2_650M only, so either the model size or the layer/dimension description is wrong. Only the authors can say which.

Second check: another model agreed · checked by hand

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 21 Extended Data Fig. 2 caption, panel A

  the observed isotopic distribution widths (black) are compared with fitted reference distributions assuming EX2 kinetics (red)

• p. 20 Extended Data Fig. 2, panel A (in-panel labels)

  Observed distribution

Judge's note: Viewed the page image for Extended Data Fig. 2 (PDF page 21; the cluster's second location cites page 20, but the figure is on page 21). In panel A 'Observed distribution', 'Observed width' and the numerator 'Observed width' are blue text with blue curves; the EX2 fit is red. The only black elements are the FWHM bracket, the arrow, 'EX1 detected' and axes. The colour key in the caption is wrong for the observed curve.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 4 Results, 'Quantifying opening cooperativity'

  the second is strongly correlated with normalized cooperativity (r = 0.61–0.80; Extended Data Fig. 6)

• p. 25 Extended Data Fig. 6 caption

  the second principal component (PC2) is strongly correlated with family-normalized cooperativity (r = 0.61-0.80; panel B)

• p. 25 Extended Data Fig. 6, panel B (y-axis label of the PC2 plots)

  Family-normalized cooperativity

• p. 25 Extended Data Fig. 6 legend

  PC2) is strongly correlated with family-normalized cooperativity (r = 0.61-0.80; panel B)

Judge's note: Viewed p4 and p25. The panel B y-axes are labelled "Family-normalized cooperativity", with PC2 r values of 0.61, 0.80, 0.73 and 0.73. The Methods define normalized and family-normalized cooperativity as two distinct metrics and say the PCA scores were correlated with family-normalized cooperativity. The main text drops "family-". The last sentence of the Extended Data Fig. 6 legend also says "normalized cooperativity".

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 4 Quantifying opening cooperativity, last sentences

  We did not observe significant differences in normalized cooperativity

• p. 5 Fig. 2g right panel and caption

  family-normalized cooperativity (right)

• p. 4 Results, 'Profiling energy landscapes across families'

  We did not observe significant differences in normalized cooperativity (mean differences of 0.0 ± 0.2 s.d. and 0.2 ± 0.5 s.d.

Only the authors can say what is right here.

Judge's note: Viewed p4 and Fig. 2g on p5. The right panel's y-axis and the legend both say family-normalized cooperativity, and the preceding sentence introduces the family-normalized metric for intrafamily comparisons. The Methods define the two metrics as distinct. The printed differences cannot be recomputed from the files, so only the authors can say which metric the 0.0 and 0.2 s.d. values refer to; most likely "family-" was dropped.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 12 Methods, 'Library design' (first sentence)

  The initial set of 15,715 domain sequences was organized into five batches and further divided into 18 libraries (mix 1–4, libraries 1 and 4; libraries 7–15; and mutants 2–4)

• p. 12 Methods, 'Library design' (last sentence)

  Oligo libraries encoding the original 15,715 sequences were purchased from Agilent Technologies, while the 280 designed mutations were sourced from Twist Bioscience.

• p. 23 Extended Data Fig. 4d,e (library labels)

  Lib15

• p. 12 Library design

  sourced from Twist Bioscience

Only the authors can say what is right here.

Judge's note: Read the full page 12 image. The first sentence counts 4 + 2 + 9 + 3 = 18 libraries and lists '(5) mutant libraries' as one of the five batches of the initial 15,715 set; the last sentence of the paragraph treats the 280 mutants as separate from the 'original 15,715'. Extended Data Fig. 4d,e (page 23 image) plots 15 library labels for the initial set: Mix1-4, Lib01, Lib04, Lib07-Lib15, with no mutant libraries. No statement elsewhere says the 280 mutants are part of the 15,715. Likely loose wording (18 libraries in total, 15 carrying the initial set), so wording-level; only the authors can say whether 15,715 includes the mutants. The supplier-name spelling variation (Bioscience / Biosciences) is visible on the page and is cosmetic. The heading point (reviewer note 3) is handled under C033.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 20 Extended Data Fig. 1 caption, panels D-F

  (D) HHH_rd4_0943.pdb_G with a merge factor near 10³. (E) PASTA_0085 with a lower factor (2.43), and (F) HHH_rd2_0070.pdb_GG with a higher factor (3.22).

• p. 20 Extended Data Fig. 1, panel D annotation

  3.04 (Merge factor log10 scale)

• p. 24 Extended Data Fig. 5 caption

  A merge factor of 3 (i.e. a 1,000-fold difference in kHX) is expected

Judge's note: Viewed the page image and a 4x render of panels C-F. Panel D prints 3.04 on a log10 scale, panel E prints 2.46, panel F prints 3.22. The caption gives D as a fold-change (10³) while E and F are given as log10 values and compared as lower/higher, so the sentence is only coherent if D is read as about 3. Extended Data Fig. 5 caption confirms the merge factor is defined on a log10 scale ("A merge factor of 3 (i.e. a 1,000-fold difference in kHX)"). Additional point not raised by the reviewer: the caption's 2.43 for panel E disagrees with the 2.46 printed in the panel.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 20 Extended Data Fig. 1 legend, panel I

  with the credible interval (5–95%) shown as error bars

• p. 20 Extended Data Fig. 1 legend, last sentence

  Error bars indicate 95% credible intervals extracted from the posterior distributions.

• p. 20 Extended Data Fig. 1 legend, panel I

  credible interval (5–95%) shown as error bars

• p. 20 ED Fig. 1 caption, last sentence

  Error bars indicate 95% credible intervals

Only the authors can say what is right here.

Judge's note: Both sentences are on the page image in the panel I part of the legend. A 5th-95th percentile interval is a 90% interval, so the two descriptions differ unless '5–95%' is loose shorthand. The Fig. 1e legend says 95% credible intervals. No statement elsewhere in the article resolves which was plotted; only the authors can say.

Second check: another model agreed · checked by hand

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send