Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 104 of the 106 references it cites were found in a registry or an index, and 2 are not checked yet; 21 of the 39 sources its claims cite were checked for support.

Scaffolding protein functional sites using deep learning

Open access

Published version of a preprint

Abstract

The binding and catalytic functions of proteins are generally mediated by a small number of functional residues held in place by the overall protein structure. Here, we describe deep learning approaches for scaffolding such functional sites without needing to prespecify the fold or secondary structure of the scaffold. The first approach, “constrained hallucination,” optimizes sequences such that their predicted structures contain the desired functional site. The second approach, “inpainting,” starts from the functional site and fills in additional sequence and structure to create a viable protein scaffold in a single forward pass through a specifically trained RoseTTAFold network. We use these two methods to design candidate immunogens, receptor traps, metalloproteins, enzymes, and protein-binding proteins and validate the designs using a combination of in silico and experimental tests.

Verified 2026-10-05 · 2 references not checked yet

Overview

CoChat Summary

TL;DR

The paper introduces two deep learning methods, constrained hallucination and inpainting, for scaffolding functional sites into de novo protein structures without pre-specifying scaffold fold or secondary structure. It validates designs in silico and experimentally across immunogens, receptor traps, metalloproteins, enzymes, and protein binders.

Key findings

  • AlphaFold recapitulated designed structures with overall and motif RMSD typically <2 Å and <1 Å, respectively, and pLDDT >80; inpainting was much faster (1–10 s/design on an NVIDIA RTX2080 GPU) than hallucination (5–20 min/design).
  • For RSV-F site V, 37 hallucinated scaffolds were expressed in E. coli and three bound hRSV90 with Kd’s of 0.9–1.3 μM; four inpainted designs bound by yeast display but were poorly expressed in E. coli.
  • For the E. coli bacterioferritin di-iron site, 96 inpainted designs were tested: 76 had soluble expression, at least 8 showed a Co2+-binding spectroscopic shift, and 3 had CD spectra consistent with the designed fold and were stabilized by metal binding; mutation of metal-binding residues abolished binding.
  • For PD-L1, of 31 designs tested, pdl1_inp_1 bound PD-L1 with KD = 326 nM, compared with HAC PD-1 KD = 110 pM and WT PD-1 KD = 3.9 μM; carbonic anhydrase II active-site designs were <100 residues versus the 261-residue native protein.

Methods

  • Constrained hallucination optimized sequences with a composite loss combining a hallucination loss and a motif reconstruction loss using trRosetta and later RosettaFold; inpainting used RFjoint, trained on masked/inpainting sequence and structure recovery, to complete missing sequence and structure in a single forward pass.
  • In silico validation used AlphaFold; experimental validation for selected designs included E. coli expression, yeast display, binding assays (Kd, BLI), circular dichroism, metal/terbium fluorescence, and mutagenesis.
  • Functional motifs were derived from native proteins including RSV-F sites II and V, hACE2, bacterioferritin di-iron, EF-hand, carbonic anhydrase II, Δ5-3-ketosteroid isomerase, PD-1/PD-L1, p53/Mdm2, and TrkA-NGF.

Limitations

  • Stated in the paper: AlphaFold has limited ability to predict protein stability or mutational effects; hallucination is compute-intensive; inpainting usually requires larger input motifs and underperforms hallucination when the missing region is large; some inpainted RSV-F site V designs were poorly expressed in E. coli; di-iron hallucination designs were not tested due to buried polar residues; and side-chain placement for ketosteroid isomerase was difficult, requiring a two-stage approach.
  • Evident from the text: many designs were validated only in silico (e.g., receptor traps, enzyme active sites, some Mdm2 binders), and experimental hit rates were often low (e.g., 3/37 RSV-F site V binders, 3/96 di-iron designs with CD/stabilization, 1/31 PD-L1 binders); inpainting also generated less structural diversity than hallucination.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1126/science.abn2100

Cites

104 of 106 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1. Methods for protein function design (A) Applications of functional-site scaffolding. (B-C) Design methods. (B) Constrained hallucination. At each iteration, a sequence is passed to the trRosetta or RosettaFold neural network, which predicts 3D coordinates and residue-residue distances and orientations (Fig. S2) which are scored by a loss function that rewards certainty of the predicted structure along with motif recapitulation and other task-specific functions. (C) Missing information recovery (“Inpainting”). Partial sequence and/or structural information is input into a modified RosettaFold network (termed RFjoint), and complete sequence and structure are output. (D) Protein design challenges formulated as missing information recovery problems. (E) Joint RosettaFold (RFjoint) can simultaneously recover structure and sequence of a masked region of protein. 2KL8 was fed into RFjoint with a continuous (length 30) window of sequence and structure masked out, with the network tasked with predicting the missing region of protein. Outputs (inpainted region in gray) closely resemble the original protein (2KL8, left) and are confidently predicted by AlphaFold (pLDDT/Motif RMSD of models shown: 91.6/0.91, 92.0/0.69, 90.4/0.82 respectively). (F-G) Motif scaffolding benchmarking data comparing RFjoint with constrained hallucination. A set of 28 de novo designed proteins, published since RosettaFold was trained, were used. For each protein, 20 random masks of length 30 were generated, and RFjoint and hallucination were tasked with filling in the missing sequence and structure to “scaffold” the unmasked “Motif”. For this mask length, RFjoint typically modestly outperforms hallucination, both in terms of the RMSD of the unmasked protein (the “motif”) to the original structure (F), and in AlphaFold confidence (pLDDT in the replaced region) (G). Circles: Average of 20 outputs for each of the benchmarking proteins. Triangle: 2KL8. Colors in all panels: native functional motif (orange); hallucinated/inpainted scaffold (gray); constrained motif (purple); binding partner (blue); non-masked region (green); masked region (light gray, dotted lines).
Figure 1. Methods for protein function design (A) Applications of functional-site scaffolding. (B-C) Design methods. (B) Constrained hallucination. At each iteration, a sequence is passed to the trRosetta or RosettaFold neural network, which predicts 3D coordinates and residue-residue distances and orientations (Fig. S2) which are scored by a loss function that rewards certainty of the predicted structure along with motif recapitulation and other task-specific functions. (C) Missing information recovery (“Inpainting”). Partial sequence and/or structural information is input into a modified RosettaFold network (termed RFjoint), and complete sequence and structure are output. (D) Protein design challenges formulated as missing information recovery problems. (E) Joint RosettaFold (RFjoint) can simultaneously recover structure and sequence of a masked region of protein. 2KL8 was fed into RFjoint with a continuous (length 30) window of sequence and structure masked out, with the network tasked with predicting the missing region of protein. Outputs (inpainted region in gray) closely resemble the original protein (2KL8, left) and are confidently predicted by AlphaFold (pLDDT/Motif RMSD of models shown: 91.6/0.91, 92.0/0.69, 90.4/0.82 respectively). (F-G) Motif scaffolding benchmarking data comparing RFjoint with constrained hallucination. A set of 28 de novo designed proteins, published since RosettaFold was trained, were used. For each protein, 20 random masks of length 30 were generated, and RFjoint and hallucination were tasked with filling in the missing sequence and structure to “scaffold” the unmasked “Motif”. For this mask length, RFjoint typically modestly outperforms hallucination, both in terms of the RMSD of the unmasked protein (the “motif”) to the original structure (F), and in AlphaFold confidence (pLDDT in the replaced region) (G). Circles: Average of 20 outputs for each of the benchmarking proteins. Triangle: 2KL8. Colors in all panels: native functional motif (orange); hallucinated/inpainted scaffold (gray); constrained motif (purple); binding partner (blue); non-masked region (green); masked region (light gray, dotted lines).CC BY · PubMed Central
Figure 2. Design of epitope scaffolds and receptor traps. (A) Design of proteins scaffolding immunogenic epitopes on RSV protein F (site II: PDB 3IXT chain P residues 254–277; site V: 5TPN chain A residues 163–181). Comparisons of the RF hallucinated models to AF2 structure predictions from the design sequence are in Fig. S9; here because of space constraints we show only the AF2 model; the two are very close in all cases. Here and in the following figures, we assess the extent of success in designing sequences which fold to structures harboring the desired motif through two metrics computed on the AF2 predictions: prediction confidence (AF pLDDT), and the accuracy of recapitulation of the original scaffolded motif (motif RMSD AF versus native). For RSV-F designs, these metrics are rsvf_ii_141 (85.0, 0.53 Å), rsvf_ii_158 (82.9, 0.51 Å), rsvf_ii_171 (88.4, 0.69 Å); rsvfv_hal_1 (82, 0.7 Å); rsvfv_hal_2 (88, 0.64 Å); rsvfv_hal_3 (86, 0.65 Å). (B) Design of COVID-19 receptor trap based on ACE2 interface helix (6VW1 chain A residues 24–42). Design metrics: ace2_76 (89.1, 0.55 Å); ace2_1157 (80.4, 0.47 Å); ace2_1007 (83.3, 0.57 Å). Colors: native protein scaffold (light yellow); native functional motif (orange); hallucinated scaffold (gray); hallucinated motif (purple); binding partner (blue). See Table S2 for additional metrics on each design. (C) Normalized maximum SPR signal (response units) of purified RSV-F epitope scaffolds and point mutants at various concentrations of hRSV90 antibody, with sigmoid fits. RSV-F refers to purified trimeric native F protein. KD values for each design are shown in legend. (D) Mean residue ellipticity (MRE) versus wavelength, from CD spectroscopy, for the 3 RSV-F site V hallucinations with binding activity.
Figure 2. Design of epitope scaffolds and receptor traps. (A) Design of proteins scaffolding immunogenic epitopes on RSV protein F (site II: PDB 3IXT chain P residues 254–277; site V: 5TPN chain A residues 163–181). Comparisons of the RF hallucinated models to AF2 structure predictions from the design sequence are in Fig. S9; here because of space constraints we show only the AF2 model; the two are very close in all cases. Here and in the following figures, we assess the extent of success in designing sequences which fold to structures harboring the desired motif through two metrics computed on the AF2 predictions: prediction confidence (AF pLDDT), and the accuracy of recapitulation of the original scaffolded motif (motif RMSD AF versus native). For RSV-F designs, these metrics are rsvf_ii_141 (85.0, 0.53 Å), rsvf_ii_158 (82.9, 0.51 Å), rsvf_ii_171 (88.4, 0.69 Å); rsvfv_hal_1 (82, 0.7 Å); rsvfv_hal_2 (88, 0.64 Å); rsvfv_hal_3 (86, 0.65 Å). (B) Design of COVID-19 receptor trap based on ACE2 interface helix (6VW1 chain A residues 24–42). Design metrics: ace2_76 (89.1, 0.55 Å); ace2_1157 (80.4, 0.47 Å); ace2_1007 (83.3, 0.57 Å). Colors: native protein scaffold (light yellow); native functional motif (orange); hallucinated scaffold (gray); hallucinated motif (purple); binding partner (blue). See Table S2 for additional metrics on each design. (C) Normalized maximum SPR signal (response units) of purified RSV-F epitope scaffolds and point mutants at various concentrations of hRSV90 antibody, with sigmoid fits. RSV-F refers to purified trimeric native F protein. KD values for each design are shown in legend. (D) Mean residue ellipticity (MRE) versus wavelength, from CD spectroscopy, for the 3 RSV-F site V hallucinations with binding activity.CC BY · PubMed Central
Figure 3. Design of metal binding (A) Di-iron binding site from E. coli cytochrome b1 (1BCF chain A residues 18–25, 27–54, 94–97, 123–130). Colors: native protein scaffold (light yellow); native functional motif (orange); hallucinated scaffold (gray); hallucinated motif (purple); bound metal (blue). Active site residues shown in boxes for di-iron and EF-hand respectively. (B) Absorbance spectra showing of dife_inp_1 (or mutant) in the presence (or not) of an 8-fold molar excess of Co2+. Note the peaks at 520 nm, 555 nm and 600 nm, consistent with Co2+ binding to the desired scaffolded motif (33). The mutant design was the same sequence but with the 6 coordinating residues (sidechains shown in (A)) mutated to alanine [E16A, E55A, H58A, E89A, H92A, E115A]). Protein concentration was 200 μM. (C) Titration analysis of Co2+ against the design (protein concentration = 200 μM). Quantification of the absorbance at 550 nm, using a predicted extinction coefficient of 155 for Co2+ binding the motif (33), is consistent with both binding sites being recapitulated in the dife_inp_1 design. (D) CD spectra of design in the presence and absence of Co2+. Both spectra are consistent with the predicted helical structure. (E) CD melt curve in the presence and absence of Co2+. Note that the coordination of Co2+ in the protein core significantly stabilizes dife_inp_1 (protein concentration in CD experiments = 6.7 μM, Co2+ concentration = 53.3 μM). (F) AF2 prediction of inpainted design EFhand_inp_1 scaffolding the double EF-hand motif with input motif residues in purple, input non-motif residues in green, and overlaid with the native motif from 1PRW (orange). (G) Tryptophan-enhanced terbium fluorescence spectra of EFhand_inp_1 matches known spectra (57) and suggests the design can bind terbium. (H) CD spectra of EFhand_inp_1 incubated with (4X protein concentration) and without CaCl2 suggest stabilization of the protein upon binding calcium. Design metrics (AF pLDDT, motif RMSD AF versus native): dife_inp_1 (92 /0.65 Å), EFhand_inp1 (84, 0.7 Å).
Figure 3. Design of metal binding (A) Di-iron binding site from E. coli cytochrome b1 (1BCF chain A residues 18–25, 27–54, 94–97, 123–130). Colors: native protein scaffold (light yellow); native functional motif (orange); hallucinated scaffold (gray); hallucinated motif (purple); bound metal (blue). Active site residues shown in boxes for di-iron and EF-hand respectively. (B) Absorbance spectra showing of dife_inp_1 (or mutant) in the presence (or not) of an 8-fold molar excess of Co2+. Note the peaks at 520 nm, 555 nm and 600 nm, consistent with Co2+ binding to the desired scaffolded motif (33). The mutant design was the same sequence but with the 6 coordinating residues (sidechains shown in (A)) mutated to alanine [E16A, E55A, H58A, E89A, H92A, E115A]). Protein concentration was 200 μM. (C) Titration analysis of Co2+ against the design (protein concentration = 200 μM). Quantification of the absorbance at 550 nm, using a predicted extinction coefficient of 155 for Co2+ binding the motif (33), is consistent with both binding sites being recapitulated in the dife_inp_1 design. (D) CD spectra of design in the presence and absence of Co2+. Both spectra are consistent with the predicted helical structure. (E) CD melt curve in the presence and absence of Co2+. Note that the coordination of Co2+ in the protein core significantly stabilizes dife_inp_1 (protein concentration in CD experiments = 6.7 μM, Co2+ concentration = 53.3 μM). (F) AF2 prediction of inpainted design EFhand_inp_1 scaffolding the double EF-hand motif with input motif residues in purple, input non-motif residues in green, and overlaid with the native motif from 1PRW (orange). (G) Tryptophan-enhanced terbium fluorescence spectra of EFhand_inp_1 matches known spectra (57) and suggests the design can bind terbium. (H) CD spectra of EFhand_inp_1 incubated with (4X protein concentration) and without CaCl2 suggest stabilization of the protein upon binding calcium. Design metrics (AF pLDDT, motif RMSD AF versus native): dife_inp_1 (92 /0.65 Å), EFhand_inp1 (84, 0.7 Å).CC BY · PubMed Central
Figure 4. In silico design of enzyme active sites. (A-B) Hallucinations using backbone description of site using RF. (C-D) Hallucination using sidechain description of site using AF2 augmented with trRosetta (Materials and Methods). (A) Carbonic anhydrase II active site (5YUI chain A residues 62–65, 93–97, 118–120). (B) Δ5-3-ketosteroid Isomerase active site (1QJG chain A residues 14, 38, 99). Colors: native protein scaffold (light yellow); native functional motif (orange); hallucinated scaffold (gray); hallucinated motif (purple); bound metal (blue). Active site residues shown for boxed designs in panel B and for carbonic anhydrase II, and Δ5-3-Ketosteroid Isomerase respectively. Design metrics (AF pLDDT, motif RMSD AF versus native): hcA_1 (73, 1.04 Å), hcA_2 (71, 0.62 Å), KSI_1 (84, 0.30 Å Cb), KSI_2 (72, 0.53 Å Cb)
Figure 4. In silico design of enzyme active sites. (A-B) Hallucinations using backbone description of site using RF. (C-D) Hallucination using sidechain description of site using AF2 augmented with trRosetta (Materials and Methods). (A) Carbonic anhydrase II active site (5YUI chain A residues 62–65, 93–97, 118–120). (B) Δ5-3-ketosteroid Isomerase active site (1QJG chain A residues 14, 38, 99). Colors: native protein scaffold (light yellow); native functional motif (orange); hallucinated scaffold (gray); hallucinated motif (purple); bound metal (blue). Active site residues shown for boxed designs in panel B and for carbonic anhydrase II, and Δ5-3-Ketosteroid Isomerase respectively. Design metrics (AF pLDDT, motif RMSD AF versus native): hcA_1 (73, 1.04 Å), hcA_2 (71, 0.62 Å), KSI_1 (84, 0.30 Å Cb), KSI_2 (72, 0.53 Å Cb)CC BY · PubMed Central
Figure 5. Design of protein-binding proteins. Designs containing target-binding interfaces built around native-complex-derived binding motifs. Targets are in blue, native scaffolds in yellow or pink, native motifs in orange, designed scaffolds in gray and designed motifs in purple. (A) Crystal structure of high-affinity consensus (HAC) PD-1 in complex with PD-L1. (B) Inpainted PD-L1 binder superimposed on PD-1 interface motif. (C) Max BLI binding signal versus PD-L1 concentration. (D) Crystal structure of previously designed TrkA minibinder in complex with TrkA, superimposed on TrkA receptor dimer. (E) Hallucinated bivalent TrkA binder. Protein topologies of (D-E) are shown to the right. (F) Max BLI binding signal versus TrkA concentration, showing that both binding sites bind TrkA. (G) Hallucinated Mdm2 binder designs superimposed on native p53 helix in complex with Mdm2 (see also Fig. S17D–E). New binding interactions (hallucinated residues within 5 Å of the target) are in green. Inset: Overlay of mdm2_hal_1 and native p53 helix showing key sidechains for binding.
Figure 5. Design of protein-binding proteins. Designs containing target-binding interfaces built around native-complex-derived binding motifs. Targets are in blue, native scaffolds in yellow or pink, native motifs in orange, designed scaffolds in gray and designed motifs in purple. (A) Crystal structure of high-affinity consensus (HAC) PD-1 in complex with PD-L1. (B) Inpainted PD-L1 binder superimposed on PD-1 interface motif. (C) Max BLI binding signal versus PD-L1 concentration. (D) Crystal structure of previously designed TrkA minibinder in complex with TrkA, superimposed on TrkA receptor dimer. (E) Hallucinated bivalent TrkA binder. Protein topologies of (D-E) are shown to the right. (F) Max BLI binding signal versus TrkA concentration, showing that both binding sites bind TrkA. (G) Hallucinated Mdm2 binder designs superimposed on native p53 helix in complex with Mdm2 (see also Fig. S17D–E). New binding interactions (hallucinated residues within 5 Å of the target) are in green. Inset: Overlay of mdm2_hal_1 and native p53 helix showing key sidechains for binding.CC BY · PubMed Central

Verification

Claims the paper makes

9 Valid · 0 Not valid · 12 Not checked · 71 Own results and interpretations

59 more claims · 3 with their cited sources checked

Create a free account to see all 92 claims and to verify the 16 references still open.

See all claims — free account
  • Probably not a substantial claim. 5
  • It may state two separate claims as one. 23

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Citations

Cites

Show all 94 references

Research works printed but not linked: 1

  1. Z. Lin T. Sercu Y. LeCun A. Rives “Deep generative models create new and diverse protein structures ” 35th Conference on Neural Information Processing Systems (NeurIPS 2021) 6 to 14 December 2021.

Cited by

See all 489

Metadata

Authors
Wang, J., Lisanza, S., Juergens, D., Tischer, D., Watson, J. L., Castro, K. M., Ragotte, R., Saragovi, A., Milles, L. F., Baek, M., Anishchenko, I., Yang, W., Hicks, D. R., Expòsit, M., Schlichthaerle, T., Chun, J.-H., Dauparas, J., Bennett, N., Wicky, B. I. M., Muenks, A., DiMaio, F., Correia, B., Ovchinnikov, S., & Baker, D.
Affiliations
  • University of Washington (US)
  • École Polytechnique Fédérale de Lausanne (CH)
  • Harvard University (US)
  • Howard Hughes Medical Institute (US)
Corresponding author
  • Wang, Jue
  • Lisanza, Sidney
  • Juergens, David
  • Tischer, Doug
  • Watson, Joseph L.
  • Ovchinnikov, Sergey
  • Baker, David
Venue
Science
Published
2022-07-22
Volume
377
Issue
6604
Pages
387-394
ISSN
  • 0036-8075 (ISSN-L)
  • 1095-9203
OpenAlex
W4286488966
Cited by
487
References
94
Citations by year
2026: 84 · 2025: 107 · 2024: 124 · 2023: 121 · 2022: 49
FWCI
32.74
Type
journal article
Language
en
Open access
Licence
cc-by-nc-nd
Locations
  • Science (publishedVersion)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
  • PubMed Central (acceptedVersion, cc-by)
  • Infoscience (Ecole Polytechnique Fédérale de Lausanne) (submittedVersion, cc-by-nc-nd)
  • MPG.PuRe (Max Planck Society) (submittedVersion)
  • Seoul National University Open Repository (Seoul National University) (submittedVersion)
Funders
  • Howard Hughes Medical Institute
  • Washington Research Foundation
  • Amgen
  • Open Philanthropy Project
  • Fundación Bancaria Caixa d'Estalvis i Pensions de Barcelona
  • National Institutes of Health (DP5OD026389, HHSN272201700059C)
  • National Institute on Aging (U19AG065156)
  • National Cancer Institute (R01CA240339)
  • National Institute of Allergy and Infectious Diseases (HHSN272201700059C)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Protein Structure and Dynamics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Machine Learning in Bioinformatics
  • Life Sciences › Pharmacology, Toxicology and Pharmaceutics › Drug Discovery › Computational Drug Discovery Methods
Keywords
protein scaffolding, protein design, deep learning, inpainting, RoseTTAFold, immunogen, metalloproteins, enzyme design, protein structure prediction
MeSH terms
Deep Learning, Binding Sites, Catalysis, Protein Binding, Proteins, Protein Engineering, Protein Structure, Secondary, Protein Folding

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-05, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-05 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Scaffolding protein functional sites using deep learning. https://app.cochat.ai/papers/10.1126/science.abn2100
The paper
Wang, J., Lisanza, S., Juergens, D., Tischer, D., Watson, J. L., Castro, K. M., Ragotte, R., Saragovi, A., Milles, L. F., Baek, M., Anishchenko, I., Yang, W., Hicks, D. R., Expòsit, M., Schlichthaerle, T., Chun, J.-H., Dauparas, J., Bennett, N., Wicky, B. I. M., . . . Baker, D. (2022). Scaffolding protein functional sites using deep learning. Science. https://doi.org/10.1126/science.abn2100

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.