Miniaturizing and Modifying Natural Proteins with Raygun

• p. 11 Extended Data Fig. 1 caption

  randomly selected positions in the 50 × 280 Raygun fixed-length embedding

• p. 3 Results, 'Proteins as probability distributions'

  a fixed 64,000-dimensional space (50 × 1,280)

• p. 11 Extended Data Fig. 1 caption

  five randomly selected positions in the 50 × 280 Raygun fixed-length embedding

• p. 11 Extended Data Fig. 1 legend

  50 × 280 Raygun fixed-length embedding

Judge's note: Read the full page image main-p011.png and the caption crops. The caption as typeset reads "50 × 280". The panel column headers on the same page are (5, 0), (15, 300), (25, 600), (35, 900) and (45, 1200), so the second index runs to 1,200, which is impossible in a 280-column representation. Main p.3 (crop C018-b) prints "a fixed 64,000-dimensional space (50 × 1,280)", and the Extended Data Fig. 4 caption (crop C024-b) prints "a final latent representation of dimensions 50 × 1280". A dropped "1," in the Extended Data Fig. 1 caption; the printed value contradicts the figure's own axes.

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 Results, 'Miniaturized fluorescent proteins', right column

  Raygun shortened eGFP by up to 25 amino acids (10.5%) and mCherry by 37 amino acids (15.6%), producing proteins with 199 and 206 amino acids, respectively

• p. 8 Discussion, right column

  Raygun shortened eGFP and mCherry to as little as 199 and 206 amino acids

• p. 6 Fig. 3, eGFP candidate XFP02 / mCherry candidate XFP08 labels

  Length: 213

• p. 6 Fig. 3, top row structure labels (XFP08)

  Length: 199

Judge's note: Read the Results sentence (crop C007-a), the Discussion sentence (crop C007-b) and Fig. 3 (crop C007-c/d) on the page images. Fig. 3 lengths are XFP01 219, XFP02 213, XFP03 224, XFP04 233 (eGFP) and XFP05 206, XFP06 219, XFP07 234, XFP08 199 (mCherry); no eGFP candidate is below 213. The same paragraph gives the templates as eGFP 238 aa and mCherry 236 aa, so the stated reductions give 213 and 199. Supplementary Note 3 (crop C007-g) independently states 'One candidate (199 aa, derived from mCherry)'. I searched for an alternative population or construct that could give 199 for eGFP and found none.

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 Results, 'Miniaturized TurboID proteins'

  TurboID (approximately 335 amino acids) is large enough to complicate fusion tagging

• p. 6 Results, 'Miniaturized TurboID proteins', right column

  TurboID-1 (317 amino acids, 1% reduction)

• p. 6 Results, right column

  TurboID-1 (317 amino acids, 1% reduction) and TurboID-5 (304 amino acids, 6% reduction)

• p. 5 Miniaturized TurboID

  TurboID (approximately 335

Only the authors can say what is right here.

Judge's note: Read both sentences on the page images (crops C008-d/e for 'approximately 335 amino acids' on p. 5, crop C008-c for the two variant percentages on p. 6). From 335 residues, 317 is a 5.4% reduction and 304 a 9.3% reduction (301, the length printed in Fig. 4a, gives 10.1%), not the stated 1% and 6%; the printed percentages correspond to a template of about 320-321 residues, which is also the lower end of the generation window '240 to 320' in Supplementary Methods A.6. I checked the article and supplement for a second, tagged construct length that could explain the gap and found none (the HA tag is mentioned only for expression). 'Approximately' cannot absorb 1% versus 5.4%, but only the authors can say which figure is right.

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 Results, 'Miniaturized TurboID proteins', last paragraph

  We generated 500,000 TurboID variants targeting moderate miniaturization (195–235 amino acids)

• p. 12 Supplementary Methods A.6, TurboID candidate selection process

  the model was used to generate 500,000 Tur- boID candidates between the ranges of 240 to 320

• p. 7 Fig. 4a, candidate lengths

  258

• p. 7 Fig. 4a, lengths of TurboID-1 to TurboID-10

  317

Judge's note: Read the main-text sentence on the page image (crops C009-a/e), Supplementary Methods A.6 on the supp-01 page image (crops C009-g/h) and the Fig. 4a length row on main p. 7 (crops C009-f, C013-f). All ten 'moderate objective' candidates lie between 258 and 317 aa, i.e. inside the supplement's 240-320 window and entirely outside the main text's 195-235 window; 195-235 is the range the preceding section gives for the eGFP/mCherry generation ('generating 70,000 candidates per template at lengths of 195-235 amino acids'), so it appears to have been carried over. The extreme objective (150-180 aa) is consistent with TurboID-11 at 165 aa. No alternative generation run with a 195-235 window is reported anywhere in 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. 20 Extended Data Fig. 9 caption, panel (a)

  We evaluated 24 templates from the PFAM analysis spanning four SCOP classes (one protein excluded due to missing GO annotations). For 14 of 19 templates

• p. 19 Extended Data Fig. 8 caption

  For each PFAM domain, we obtained 5 representative proteins of diverse sizes

• p. 20 Extended Data Fig. 9a

  We evaluated 24 templates

• p. 5 Results

  14 out of 19 templates

Judge's note: Read main-p020.png in full: panel (a) contains exactly 19 scatter panels (rows of 5, 5, 5 and 4: P32246, P32250, P35376, P25100, P30989; A9A3D4, D7DUZ7, B2A1Z7, D1Z1H4, A9I4C3; Q24CC0, H8KX25, C0Z8L5, D7C2V6, Q7UG99; P52697, Q21KS1, Q03PD4, Q5FUT3). Read main-p019.png: Extended Data Fig. 8 lists exactly five UniProt IDs for each of the four SCOP classes (PF00001, PF10282, PF04794, PF01996) and states "For each PFAM domain, we obtained 5 representative proteins", fixing the pool at 20. A9NE84 (Fig. 8d) is the one template absent from Fig. 9a, consistent with 20 − 1 = 19, which is also the denominator in the caption and in the main text on p.5 ("for 14 out of 19 templates"). The figure's "24" is therefore inconsistent with its own arithmetic and with Fig. 8.

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 Algorithm 1, Forward operation of Raygun Reduction Layer

  Me, Se = GetMeanStd(Es, mws + 1)

• p. 5 Algorithm 1, definition of the end block

  Ee = E[n −gr · (mws + 1) :, :];

Judge's note: Read the algorithm on the page image supp-01-p005.png. The three slices are defined as Es = E[0 : gl·(mws+1), :], Em = E[gl·(mws+1) : gl·(mws+1)+mid·mws, :] and Ee = E[n − gr·(mws+1) :, :]; the next three lines print Ms,Ss = GetMeanStd(Es, mws+1), Mm,Sm = GetMeanStd(Em, mws) and Me,Se = GetMeanStd(Es, mws+1). The third call takes Es, so Ee is defined and never consumed, and the concatenations M = Concatenate(Ms, Mm, Me) and S = Concatenate(Ss, Sm, Se) would summarise the N-terminal block twice and omit the C-terminal block. That contradicts the text on supp-01 p.4, which says the first 22 and last 21 blocks are each given size 11 within one partition of the sequence, and it is asymmetric with the mirror-image Algorithm 3, which correctly uses three distinct slices for Ts, Tm and Te. GetMeanStd is a pure statistic of its argument, so re-using Es cannot be deliberate sharing. The evident fix is GetMeanStd(Ee, mws + 1).

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 Algorithm 2 GetMeanStd()

  Output: m, s ∈ Rk×1280

• p. 5 Algorithm 2 GetMeanStd(), loop

  for i ∈ 1, . . . , r do

• p. 5 Algorithm 2, loop body

  for i ∈ 1, . . . , r do m[i, :] = mean(X[ik : (i + 1)k, :])

Only the authors can say what is right here.

Judge's note: Read Algorithm 2 on the page image supp-01-p005.png. As printed the loop index runs to r while the declared output has k rows, and the slice width is k while the declared block structure (rk rows in k output blocks) has width r. The complementary routine, Algorithm 4 Fill(), uses the consistent convention: input X ∈ R^{k×1280}, output Y ∈ R^{kr×1280}, "for i ∈ 1, . . . , k do Y[ir : (i+1)r, :] = broadcast(X[i, :], r)". Checking against the call site in Algorithm 1, GetMeanStd(Es, mws+1) passes r = mws+1 with Es ∈ R^{gl·(mws+1)×1280}, so the required output Ms has gl rows, whereas the printed loop would produce mws+1 rows; the two readings coincide only when r = k. The pseudo-code is internally inconsistent as printed, but whether the intended fix is to swap the loop bound and slice width or to redeclare r and k is for the authors to say. I did not treat the 1-based start (i = 1 … with index ik) as a separate defect, since Algorithm 4 shares that convention.

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. 6 Results, 'Miniaturized TurboID proteins', right column

  TurboID-5 (304 amino acids, 6% reduction)

• p. 7 Fig. 4a, TurboID-5 length row

  301

• p. 13 Supplementary Table 2, TurboID-5 sequence

  MKDNTVPLLLILLANGFFSGELGETLGSRAAINKHLTLRDGVDVFFVGKGYSPEPIPLNAKQILGQLDGGSVAVLPVVDSTNYYLLDRIG

• p. 7 Fig. 4a, TurboID-1 length row

  317

Judge's note: Read the Results sentence on the page image (crop C013-a/C008-c) and the Fig. 4a length row on main p. 7 (crops C013-f, C009-f), which gives 317, 314, 310, 305, 301, 297, 294, 289, 271, 258 and 165. The Supplementary Table 2 TurboID-5 entry (crops C013-c, C013-e; supp-01 p. 13) is three 90-character lines plus the 31-character line 'FGISGIDKGGALLEQGVIKPWMGEISLRAEK' = 301 residues, while TurboID-1 is 90+90+90+47 = 317, matching both the figure and the text. So the two documents use the same measure and the text value 304 is the outlier; nothing in the article or supplement describes a tagged 304-residue construct.

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. 10 Supplementary Table 1, row xfp01

  VQLDDHQQTPIGGGVLLPDNHYLSTQSALSKDPNEKRDHMVLLEFVTAAGITLGTDELYK

• p. 6 Fig. 3, XFP01 label

  Length: 219

Only the authors can say what is right here.

Judge's note: Viewed supp-01 p010 (full page and crop C037-a) and main Fig. 3 crop C037-b. Table 1 is set in a monospace font with wrapped 80-character lines; xfp01 line 1 (MVSKGEELFT...YDDKMQDFFK) and line 2 (SCPEGVQERT...VNMIRHNEDG) are each 80 residues counted in blocks of ten, and line 3 is 60, giving 220. Fig. 3 prints "Length: 219" for XFP01. I counted the other seven entries the same way and every one matches its figure label exactly (xfp02 80+80+53=213, xfp03 80+80+64=224, xfp04 80+80+73=233, xfp05 80+80+46=206, xfp06 80+80+59=219, xfp07 80+80+74=234, xfp08 80+80+39=199), so the one-residue gap is specific to XFP01. Nothing on either page explains a different length convention. Only the authors can say whether the figure label or the tabulated sequence is at fault.

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. 7 Fig. 4d, sequence printed under the EGF-Raygun-1 structure

  NNLYKRCAQSHDGYCLHDGVCMYIEALD KYACNCVVGYIGERCQYRDLKWWWREES

• p. 14 Supplementary Table 4

  EGF-Raygun-1 YSAVRRCERSHDGYCLHDGVCMYIEALDKYACNCVVGYIGERCQYRDLKWIAGLII

• p. 8 Results, 'EGFR binders via magnification', right column

  EGF-Raygun-1 (70.7% identity) and EGF-Raygun-2 (76.8% identity)

• p. 7 Fig. 4d, sequence printed under the EGF-Raygun-2 structure

  YSAVRRCERSHDGYCLHDGVCMYIEALDKYACNCVVGYIGERCQYRDLKWIAGLII

Judge's note: Read Fig. 4d on the page image (crops C015-a and C015-d, showing the sequence printed under each labelled structure) and Supplementary Table 4 on the supp-01 page image (crop C015-b). The two published assignments contradict each other. The internal evidence favours the supplementary table: the Results (crop C015-c) state EGF-Raygun-1 has the lower identity to wild type (70.7%) and EGF-Raygun-2 the higher (76.8%), and against the wild-type EGF template given in Supplementary Table 3 (Q6QBS2, NSDSECPLSHDGYCLHDGVCMYIEALDKYACNCVVGYIGERCQYRDLKWWELR, 53 aa) the NNLYKR... candidate retains more wild-type residues outside the shared 41-residue core (it keeps the initial N and a W in the C-terminal extension) than the YSAVRR... candidate does, so NNLYKR... is the more similar of the two and must be the 76.8% candidate, i.e. EGF-Raygun-2. I also checked the separate complaint that 70.7% is not a multiple of 1/56: 41/58 = 70.7%, attainable with a gapped 58-position alignment, so that part is not a defect.

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 Supplementary Methods A.1, second paragraph (decoder)

  All the projected outputs coming from the Reduction layer, after passing through the T-Blocks, are then aggregated in the final projection layer.

• p. 15 Extended Data Fig. 4 caption

  each containing two Decoder TransformerLayer (TD) blocks and a Repetition layer

Judge's note: Read supp-01 p.4 as a page image. The paragraph opens "Similarly in the Raygun decoder, the T-Blocks are also usually arranged in pairs before and after the Repetition operation" and ends "We use 21 T-Blocks in the Raygun decoder: 10 before the Repetition operation and 11 after"; the length-transforming layer in the decoder is therefore the Repetition layer, and the Reduction layer belongs to the encoder (previous paragraph: "In the Raygun encoder, a T-Block output is usually the input to another T-Block and the Reduction layer"). Extended Data Fig. 4's caption (crop C024-b) confirms: the decoder cascades each contain "two Decoder TransformerLayer (TD) blocks and a Repetition layer". The sentence names the encoder's layer inside the decoder description; the label is wrong, the architecture described around it is not.

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 Results, 'Structure preservation across sizes' (continued), left column

  with retention from 37% (α + β) to 63% (α/β)

• p. 13 Extended Data Fig. 3d bar chart

  PF10282

• p. 19 Extended Data Fig. 8 panel titles

  b Beta-propeller PF10282 (SCOP: β-only)

• p. 5 Functional preservation across Pfam

  retention from 37% (α + β) to 63% (α/β)

Judge's note: Read the Results sentence on the page image (crop C010-d), Extended Data Fig. 3d on main p. 13 and the Extended Data Fig. 8 panel titles on main p. 19. Fig. 8 assigns the classes explicitly: (a) Rhodopsin PF00001 alpha-only, (b) Beta-propeller PF10282 beta-only, (c) YdjC-like PF04794 alpha/beta, (d) F420 ligase PF01996 alpha + beta. In Fig. 3d the Raygun bars rise left to right PF01996 (~38%), PF00001 (~44%), PF04794 (~57%), PF10282 (~63%), whose mean matches the quoted 50.65% average. The lower bound, 37% for alpha + beta, is correctly attributed; the class label on the 63% value is not. Bar heights read from the page image are approximate, but the ordering of the two highest bars is unambiguous.

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. 9 Reference 20

  Denoising diffusion implicit models. In International Conference on Learning Representations https://openreview.net/pdf?id=St1giarCHLP (ICLR, 2023)

Only the authors can say what is right here.

Judge's note: Rests on outside knowledge: the year in which Song, Meng & Ermon's "Denoising diffusion implicit models" appeared at ICLR, which is not stated anywhere in the files. I read reference 20 on the page image of main p.9 (crop C017-a) and it does print "(ICLR, 2023)" exactly as quoted; the OpenReview id St1giarCHLP is also printed. Only the authors or an external check of the proceedings can settle the correct year, so this is an author query rather than an outright correction.

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. 9 Supplementary Methods A.3.3

  In PFAM-based experiments (Section 2.3 of the main paper), we used HMMER

Only the authors can say what is right here.

Judge's note: Read supp-01 p009 (crop C032-a and full page) and main p001. The published Nature article uses unnumbered headings (Article/abstract, running headings) and has no numbered sections, so "Section 2.3 of the main paper" cannot be resolved; the supplement's own sections are lettered (A.1, A.2, A.3.3), so the pointer is not to the supplement either. Only the authors can say which article subsection was intended.

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. 11 Supplementary Methods, CLEAN benchmarking paragraph

  We obtained a total of 170 human PIK proteins from UniProt, out of which 13 templates were selected for benchmarking.

• p. 20 Extended Data Fig. 9 caption, panel (c)

  miniaturized variants of 13 human protein tyrosine kinase templates (EC 2.7.10.1)

• p. 20 Extended Data Fig. 9 caption, panel (c)

  13 human protein tyrosine kinase templates (EC 2.7.10.1)

• p. 11 A.3.5 CLEAN benchmarking

  170 human PIK proteins from UniProt

Judge's note: Viewed supp-01 p011 (full page, crops C038-a/d/e) and main p020 with the Extended Data Fig. 9 caption (crops C038-b/c/f). The same sentence states the family is receptor protein-tyrosine kinases (EC 2.7.10.1), and the figure caption calls the same 13 templates "13 human protein tyrosine kinase templates (EC 2.7.10.1)". A grep of the whole extract shows "PIK" occurs exactly once and is never defined; "PTK" appears only inside reference [65]'s title. The abbreviation in the supplement is inconsistent with the family named in the same sentence and in the article.

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. 14 Extended Data Fig. 3 caption, panel (e)

  across noise parameters from 0.1 to 2.0

• p. 13 Extended Data Fig. 3e(i) and 3e(ii) x axes

  Noise level (σ)

• p. 14 Extended Data Fig. 3e legend

  The only notable difference occurs at very low noise (0.1)

• p. 14 Extended Data Fig. 3e legend

  noise parameters from 0.1 to 2.0

Judge's note: Read the full figure page main-p013.png: panels e(i) 'Same length generation' and e(ii) 'Length reduction by 20%' each show six grouped bars at noise levels 0.1, 0.5, 1, 1.5, 2 and 3, with the axis labelled 'Noise level (σ)'. The caption on main-p014.png (crop C020-a) restricts the range to "from 0.1 to 2.0", so the stated range does not cover the plotted data. The second reviewer's additional complaint about "The only notable difference occurs at very low noise (0.1)" I do not confirm: at 0.1 the unfinetuned/fine-tuned bars differ by roughly 0.11 vs 0.015, by far the largest relative gap in the panel, so that sentence is a defensible reading of the data. Confirmed only on the axis-range mismatch.

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. 13

  Figure R6 (Figure A.16 in manuscript)

Only the authors can say what is right here.

Judge's note: Viewed supp-03 p012 and p013 full page images. Page 12 points the reader to Figure A.14 for Figure R6; page 13's caption for that very figure (PTM/iPTM bar charts plus three spCas9-DNA/gRNA complex renderings) says Figure A.16. Both cannot be right; only the authors can say which manuscript figure number is correct.

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 Supplementary Methods A.1.1, Reduction Layer

  Dividing the variable length embeddings into non-contiguous blocks is tricky

• p. 3 Results, 'Proteins as probability distributions'

  we partition each protein’s embedding into K = 50 contiguous stretches of residues (blocks)

• p. 4 Supplementary Methods A.1.1, Reduction Layer

  Dividing the variable length embeddings into non-contiguous blocks is tricky if the sequence length n is not an exact multiple of the fixed-length dimension K.

Judge's note: Read supp-01 p.4 as a page image (supp-01-p004.png) and crop C023-c: the sentence prints "non-contiguous". The worked example immediately following partitions a 543-residue sequence into the first 22 blocks of size 11, the middle 7 of size 10 and the last 21 of size 11, i.e. a contiguous, non-overlapping cover (22+7+21 = 50 blocks). Algorithm 1 on supp-01 p.5 likewise takes consecutive slices E[0 : gl(mws+1)], E[gl(mws+1) : gl(mws+1)+mid·mws] and E[n − gr(mws+1) :]. The main text on p.3 states "we partition each protein's embedding into K = 50 contiguous stretches of residues (blocks)". The supplement's wording is the opposite of what is described and implemented; "non-overlapping" appears to be meant. Wording defect only; the method is clear from the example and pseudo-code.

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. 7 Supplementary Methods A.2, loss list item 3

  suppose we generated a new protein p′ of length n′ using Raygun. Then, this loss is designed to ensure that the fixed-length embedding of p is close to that of p′

• p. 7 Supplementary Methods A.2, paragraph defining the losses and equation (4)

  Let p′′ be another Raygun sequence obtained from p of length n′ < n

Only the authors can say what is right here.

Judge's note: Read supp-01 p007 crop C031-a and the full page image. Bullet 3 assigns p′ length n′ and says L_rp compares the fixed-length embedding of p with that of p′. The paragraph immediately below redefines p′ as the reconstructed (length-n) Raygun sequence used in L_ce = CrossEntropy(p,p′) (eq. 2), and names the shorter length-n′ sequence p′′; eq. (4) reads L_rp = ||RGUN_p^(50) − RGUN_p′′^(50)||_2. The same symbol therefore denotes two objects and the bullet does not match eq. (4). No disambiguating statement appears elsewhere on the page.

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. 11

  showed that 14 of 19 templates achieved PS > 10

Only the authors can say what is right here.

Judge's note: Viewed supp-03 p011 full page (crop C043-a) and main p020. Figure R4 on supp-03 p011 contains 19 template panels (5+5+5+4), matching the "19" in the text, and the identical figure appears as Extended Data Fig. 9a in the article with the same 19 panels (P32246, P32250, P35376, P25100, P30989, A9A3D4, D7DUZ7, B2A1Z7, D1Z1H4, A9I4C3, Q24CC0, H8KK25, C0Z8L5, D7C2V6, Q7UG99, P52697, Q21KS1, Q03PD4, Q5FUT3). I searched for an explanation: the Extended Data Fig. 9 caption adds "(one protein excluded due to missing GO annotations)", which would give 23, not 19, and supp-03 gives no exclusion at all; the article Methods and supp-01 contain no further ProTrek template accounting. The gap between the stated starting set and the reported denominator is therefore left unexplained in both documents.

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