Journal: bioRxiv

Article Title: Deep Learning Structural Ensembles as Proxies for Protein Flexibility

doi: 10.64898/2026.05.16.725658

Figure Lengend Snippet: Ranksorted mean squared fluctuations (MSF) from NMR method and three deep learning structure prediction methods (AlphaFold3, AlphaFold2 and RosettaFold). A) Projections of ranksorted MSF onto protein structures for 2lah. Blue-White-Red color palette is used for the projections, where blue indicates low flexibility and red indicates high flexibility. B) 2D comparison of the experimental and the computed MSF for 2lah. Black line (with squares) is for the experimental data, blue line (with circles) is for AlphaFold3, orange line (with inverse triangles) is for AlphaFold2 and green line (with stars) is for RosettaFold. C) Cosine similarity of the experimental and the computed MSF for 70 proteins in the NMR dataset. AlphaFold3 bars are blue, AlphaFold2 bars are orange and RosettaFold bars are green. Averages of the cosine similarities over the entire dataset are also provided as horizontal lines for AlphaFold3 (blue dot-dashed line), AlphaFold2 (orange dashed line) and RosettaFold (green dotted line).

Article Snippet: Twenty AlphaFold3 structural models were generated using the weights provided Google Deepmind .

Techniques: Comparison

Journal: bioRxiv

Article Title: Deep Learning Structural Ensembles as Proxies for Protein Flexibility

doi: 10.64898/2026.05.16.725658

Figure Lengend Snippet: Average similarities with standard deviations as error bars for different approaches to obtain MSF from the NMR experimental data. Blue bars with gray circles are for AlphaFold3, orange bars with gray stripes are for AlphaFold2 and green bars with gray stars are for RosettaFold. A) Average cosine similarities B) Average Pearson similarities.

Article Snippet: Twenty AlphaFold3 structural models were generated using the weights provided Google Deepmind .

Techniques:

Journal: bioRxiv

Article Title: Deep Learning Structural Ensembles as Proxies for Protein Flexibility

doi: 10.64898/2026.05.16.725658

Figure Lengend Snippet: Impact of number of deep learning structural models on the NMR dataset, measured with cosine similarity (left panel) and Pearson similarity (right panel) for A) AlphaFold3 B) AlphaFold2.

Article Snippet: Twenty AlphaFold3 structural models were generated using the weights provided Google Deepmind .

Techniques:

Journal: bioRxiv

Article Title: Deep Learning Structural Ensembles as Proxies for Protein Flexibility

doi: 10.64898/2026.05.16.725658

Figure Lengend Snippet: Ranksorted mean squared fluctuations (MSF) from dual conformations of X-Ray structures and three deep learning structure prediction methods (AlphaFold3, AlphaFold2 and RosettaFold). A) Projections of the ranksorted MSF onto protein structures for 3fweB and 4hzfA. Blue-White-Red color palette is used for the projections, where blue indicates low flexibility and red indicates high flexibility. B) 2D comparison of the experimental and the computed MSF for 3fweB and 4hzfA. Black line (with squares) is for the experimental data, blue line (with circles) is for AlphaFold3, orange line (with inverse triangles) is for AlphaFold2 and green line (with stars) is for RosettaFold. C) Cosine similarity of the experimental and the computed MSF for 43 proteins in the X-Ray dataset. AlphaFold3 bars are blue, AlphaFold2 bars are orange and RosettaFold bars are green. Averages of the cosine similarities over the entire dataset are also provided as horizontal lines for AlphaFold3 (blue dot-dashed line), AlphaFold2 (orange dashed line) and RosettaFold (green dotted line).

Article Snippet: Twenty AlphaFold3 structural models were generated using the weights provided Google Deepmind .

Techniques: Comparison

Journal: bioRxiv

Article Title: Deep Learning Structural Ensembles as Proxies for Protein Flexibility

doi: 10.64898/2026.05.16.725658

Figure Lengend Snippet: Average of similarities for different approaches to obtain MSF from the first conformation set of the X-Ray experimental data. The average cosine similarities are given in the left panel and the average Pearson similarities are provided in the right panel. The average was taken over the similarities of 43 proteins. Blue bars with gray circles are for AlphaFold3, orange bars with gray stripes are for AlphaFold2 and green bars with gray stars are for RosettaFold. A) Two experimental protein conformations were used for the MSF calculations. B) All normal modes of the first conformations was used for the MSF calculations. Only Calpha atoms were used for normal mode analysis. C) Only 10 lowest eigenvalue normal modes of each conformation were used for normal mode analysis. Only Calpha atoms were used for normal mode analysis. D) Bfactors were ranksort normalized and they were used as a proxy for the MSF.

Article Snippet: Twenty AlphaFold3 structural models were generated using the weights provided Google Deepmind .

Techniques:

Journal: bioRxiv

Article Title: Deep Learning Structural Ensembles as Proxies for Protein Flexibility

doi: 10.64898/2026.05.16.725658

Figure Lengend Snippet: Impact of number of deep learning structural models on average similarity of X-ray all normal modes dataset, measured with cosine similarity (left panel, black and gray lines) and Pearson similarity (right panel, blue and light blue lines) for A) AlphaFold3 B) AlphaFold2. Continuous lines are for the first conformations set and dashed lines are for the second conformations set.

Article Snippet: Twenty AlphaFold3 structural models were generated using the weights provided Google Deepmind .

Techniques:

Journal: bioRxiv

Article Title: Deep Learning Structural Ensembles as Proxies for Protein Flexibility

doi: 10.64898/2026.05.16.725658

Figure Lengend Snippet: Ranksorted mean squared fluctuations (MSF) from all normal modes of cryo-EM structures and three deep learning structure prediction methods (AlphaFold3, AlphaFold2 and RosettaFold). A) Projections of the ranksorted MSF onto protein structures for 9yin. Blue-White-Red color palette is used for the projections, where blue indicates low flexibility and red indicates high flexibility. B) 2D comparison of the experimental and the computed MSF for 9yin. Black line (with squares) is for the experimental data, blue line (with circles) is for AlphaFold3, orange line (with inverse triangles) is for AlphaFold2 and green line (with stars) is for RosettaFold. C) Cosine similarity of the experimental and the computed MSF for 82 proteins in the cryo-EM dataset. AlphaFold3 bars are blue, AlphaFold2 bars are orange and RosettaFold bars are green. Averages of the cosine similarities over the entire dataset are also provided as horizontal lines for AlphaFold3 (blue dot-dashed line), AlphaFold2 (orange dashed line) and RosettaFold (green dotted line).

Article Snippet: Twenty AlphaFold3 structural models were generated using the weights provided Google Deepmind .

Techniques: Cryo-EM Sample Prep, Comparison

Journal: bioRxiv

Article Title: Deep Learning Structural Ensembles as Proxies for Protein Flexibility

doi: 10.64898/2026.05.16.725658

Figure Lengend Snippet: Average similarities with standard deviations as error bars for different approaches to obtain MSF from all normal modes of the cryo-EM experimental data. The average cosine similarities are given in the left panel and the average Pearson similarities are provided in the right panel. The average was taken over the similarities of 82 proteins and only Calpha atoms were used for normal mode analysis. Blue bars with gray circles are for AlphaFold3, orange bars with gray stripes are for AlphaFold2 and green bars with gray stars are for RosettaFold. A) All normal modes. B) 10 normal modes.

Article Snippet: Twenty AlphaFold3 structural models were generated using the weights provided Google Deepmind .

Techniques: Cryo-EM Sample Prep

Journal: bioRxiv

Article Title: Deep Learning Structural Ensembles as Proxies for Protein Flexibility

doi: 10.64898/2026.05.16.725658

Figure Lengend Snippet: Impact of number of deep learning structural models on the cryo-EM all normal modes dataset, measured with cosine similarity (left panel) and Pearson similarity (right panel) for A) AlphaFold3 B) AlphaFold2.

Article Snippet: Twenty AlphaFold3 structural models were generated using the weights provided Google Deepmind .

Techniques: Cryo-EM Sample Prep

Journal: bioRxiv

Article Title: Deep Learning Structural Ensembles as Proxies for Protein Flexibility

doi: 10.64898/2026.05.16.725658

Figure Lengend Snippet: Ranksorted mean squared fluctuations (MSF) from molecular dynamics (MD) simulations and two deep learning structure prediction methods (AlphaFold3 and AlphaFold2). A) Projections of the ranksorted MSF onto protein structures for 6crk chain G. Blue-White-Red color palette is used for the projections, where blue indicates low flexibility and red indicates high flexibility. B) 2D comparison of the MD and the deep learning MSF for 6crk chain G. Blue (simulation 1), green (simulation 2) and orange (simulation 3) lines (with squares) are for the MD simulation data, black line (with circles) is for AlphaFold3, gray line (with inverse triangles) is for AlphaFold2. C) Cosine similarities of the first set of MD simulations and the deep learning ensemble MSF of 10 proteins in the MD dataset. D) Cosine similarities of the second set of MD simulations and the deep learning ensemble MSF of 10 proteins in the MD dataset. E) Cosine similarities of the third set of MD simulations and the deep learning ensemble MSF 10 proteins in the MD dataset. AlphaFold3 bars are black, AlphaFold2 bars are gray. Averages of the cosine similarities over the entire dataset are also provided as horizontal lines for AlphaFold3 (black dot-dashed line), AlphaFold2 (gray dashed line).

Article Snippet: Twenty AlphaFold3 structural models were generated using the weights provided Google Deepmind .

Techniques: Comparison

Journal: bioRxiv

Article Title: Deep Learning Structural Ensembles as Proxies for Protein Flexibility

doi: 10.64898/2026.05.16.725658

Figure Lengend Snippet: Average similarities with standard deviations as error bars for different approaches to obtain MSF from the molecular dynamics simulations dataset. The average cosine similarities are given in the left panel and the average Pearson similarities are provided in the right panel. The average was taken over the similarities of 10 proteins. Black bars with white circles are for AlphaFold3 and gray bars with white stripes are for AlphaFold2 A) Simulation set 1 (Sim1) B) Simulation set 2 (Sim2) C) Simulation set 3 (Sim3).

Article Snippet: Twenty AlphaFold3 structural models were generated using the weights provided Google Deepmind .

Techniques:

Journal: bioRxiv

Article Title: Deep Learning Structural Ensembles as Proxies for Protein Flexibility

doi: 10.64898/2026.05.16.725658

Figure Lengend Snippet: Impact of number of deep learning structural models on MSF similarities for the proteins in the molecular dynamics dataset, measured with cosine similarity (left panel) and Pearson similarity for A) AlphaFold3 B) AlphaFold2.

Article Snippet: Twenty AlphaFold3 structural models were generated using the weights provided Google Deepmind .

Techniques: