Example table: protein structure and design

Create a free account to copy these papers Log in

Papers

Highly accurate protein structure prediction with AlphaFold.
Jumper J, Evans R, Pritzel A, Green T, Figurnov M, Ronneberger O, Tunyasuvunakool K, Bates R · DOI: 10.1038/s41586-021-03819-2 · 2021 · Nature

Publication found
Accurate structure prediction of biomolecular interactions with AlphaFold 3
Josh Abramson, Jonas Adler, Jack Dunger, Richard Evans, Tim Green, Alexander Pritzel, Olaf Ronneberger, Lindsay Willmore · DOI: 10.1038/s41586-024-07487-w · 2024 · Nature
De novo design of protein structure and function with RFdiffusion
Joseph L. Watson, David Juergens, Nathaniel R. Bennett, Brian L. Trippe, Jason Yim, Helen E. Eisenach, Woody Ahern, Andrew J. Borst · DOI: 10.1038/s41586-023-06415-8 · 2023 · Nature
Illuminating protein space with a programmable generative model
John Ingraham, Max Baranov, Zak Costello, Karl W. Barber, Wujie Wang, Ahmed Ismail, Vincent Frappier, Dana M. Lord · DOI: 10.1038/s41586-023-06728-8 · 2023 · Nature
ColabFold: making protein folding accessible to all
Milot Mirdita, Konstantin Schütze, Yoshitaka Moriwaki, Lim Heo, Sergey Ovchinnikov, Martin Steinegger · DOI: 10.1038/s41592-022-01488-1 · 2022 · Nature Methods

Comparison table

19 of 19 filled cells are backed by a passage of the paper.

Paper	Question	Method	Data	Main results	Limitations
1 Highly accurate protein structure prediction with AlphaFold.	Develop a computational method that predicts protein structures with atomic accuracy, including when no similar structure is known. Source · Abstract Here we provide the first computational method that can regularly predict protein structures with atomic accuracy even in cases in which no similar structure is known.	AlphaFold combines geometric and physical inductive biases with learning from PDB structures. Self-distillation adds high-confidence predicted structures to the experimental training data. Source · Training with labelled and unlabelled d… In this procedure, we use a trained network to predict the structure of around 350,000 diverse sequences from Uniclust3036 and make a new dataset of predicted structures filtered to a high-confidence subset. We then train the same architecture again from scratch using a mixture of PDB data and this new dataset of predicted structures as the training data, in which the various training data augmentations such as cropping and MSA subsampling make it challenging for the network to recapitulate the previously predicted structures.	Network inputs include the target amino acid sequence, multiple sequence alignments of related proteins, and homologous structural templates when available. Source · Inputs and data sources Inputs to the network are the primary sequence, sequences from evolutionarily related proteins in the form of a MSA created by standard tools including jackhmmer60 and HHBlits61, and 3D atom coordinates of a small number of homologous structures (templates) where available.	In CASP14, AlphaFold achieved accuracy competitive with experimental structures for a majority of cases and substantially outperformed other methods. No numerical accuracy estimates are supplied in the excerpt. Source · Abstract We validated an entirely redesigned version of our neural network-based model, AlphaFold, in the challenging 14th Critical Assessment of protein Structure Prediction (CASP14) 15 , demonstrating accuracy competitive with experimental structures in a majority of cases and greatly outperforming other methods.	Not stated
2 Accurate structure prediction of biomolecular interactions with AlphaFold 3 Abstract and our summary	Predict the joint structures of complexes containing diverse biomolecular components within a single model. Source · Abstract Here we describe our AlphaFold 3 model with a substantially updated diffusion-based architecture that is capable of predicting the joint structure of complexes including proteins, nucleic acids, small molecules, ions and modified residues.	AlphaFold 3 uses an updated diffusion-based architecture for joint complex structure prediction. Source · Abstract Here we describe our AlphaFold 3 model with a substantially updated diffusion-based architecture that is capable of predicting the joint structure of complexes including proteins, nucleic acids, small molecules, ions and modified residues.	The molecular components covered include proteins, nucleic acids, small molecules, ions and modified residues. Dataset identities and sample sizes are not stated. Source · Abstract Here we describe our AlphaFold 3 model with a substantially updated diffusion-based architecture that is capable of predicting the joint structure of complexes including proteins, nucleic acids, small molecules, ions and modified residues.	AlphaFold 3 outperformed docking tools on protein–ligand interactions, nucleic-acid-specific predictors on protein–nucleic acid interactions, and AlphaFold-Multimer v.2.3 on antibody–antigen prediction. Numerical performance estimates are not stated. Source · Abstract The new AlphaFold model demonstrates substantially improved accuracy over many previous specialized tools: far greater accuracy for protein–ligand interactions compared with state-of-the-art docking tools, much higher accuracy for protein–nucleic acid interactions compared with nucleic-acid-specific predictors and substantially higher antibody–antigen prediction accuracy compared with AlphaFold-Multimer v.2.3 7,8 . Not in this passage: 3	Not stated
3 De novo design of protein structure and function with RFdiffusion Abstract and our summary	Develop a general protein-design framework supporting tasks such as binder design, symmetric assemblies and functional motif scaffolding. Source · Abstract Here we show that by fine-tuning the RoseTTAFold structure prediction network on protein structure denoising tasks, we obtain a generative model of protein backbones that achieves outstanding performance on unconditional and topology-constrained protein monomer design, protein binder design, symmetric oligomer design, enzyme active site scaffolding and symmetric motif scaffolding for therapeutic and metal-binding protein design.	RFdiffusion fine-tunes RoseTTAFold on structure-denoising tasks to generate protein backbones for multiple design objectives. Source · Abstract Here we show that by fine-tuning the RoseTTAFold structure prediction network on protein structure denoising tasks, we obtain a generative model of protein backbones that achieves outstanding performance on unconditional and topology-constrained protein monomer design, protein binder design, symmetric oligomer design, enzyme active site scaffolding and symmetric motif scaffolding for therapeutic and metal-binding protein design.	Experimental characterization covered hundreds of designed symmetric assemblies, metal-binding proteins and protein binders. Source · Abstract We demonstrate the power and generality of the method, called RoseTTAFold diffusion (RFdiffusion), by experimentally characterizing the structures and functions of hundreds of designed symmetric assemblies, metal-binding proteins and protein binders.	A cryo-EM structure of a designed binder bound to influenza haemagglutinin closely matched the design model. No numerical structural deviation is stated. Source · Abstract The accuracy of RFdiffusion is confirmed by the cryogenic electron microscopy structure of a designed binder in complex with influenza haemagglutinin that is nearly identical to the design model.	Not stated
4 Illuminating protein space with a programmable generative model Abstract and our summary	Generate novel protein structures and sequences while allowing users to steer generation toward desired properties and functions. Source · Abstract Here we introduce Chroma, a generative model for proteins and protein complexes that can directly sample novel protein structures and sequences, and that can be conditioned to steer the generative process towards desired properties and functions.	Chroma combines polymer-aware diffusion with a sub-quadratically scaling neural architecture and low-temperature sampling. It treats constrained protein design as Bayesian inference. Source · Abstract To enable this, we introduce a diffusion process that respects the conformational statistics of polymer ensembles, an efficient neural architecture for molecular systems that enables long-range reasoning with sub-quadratic scaling, layers for efficiently synthesizing three-dimensional structures of proteins from predicted inter-residue geometries and a general low-temperature sampling algorithm for diffusion models. Chroma achieves protein design as Bayesian inference under external constraints, which can involve symmetries, substructure, shape, semantics and even natural-language prompts.	Experimental characterization included 310 designed proteins, with crystal structures reported for two designs. Source · Abstract The experimental characterization of 310 proteins shows that sampling from Chroma results in proteins that are highly expressed, fold and have favourable biophysical properties. The crystal structures of two designed proteins exhibit atomistic agreement with Chroma samples (a backbone root-mean-square deviation of around 1.0 Å).	The 310-protein characterization showed expression, folding and favorable biophysical properties. Two crystal structures agreed with generated models at a backbone RMSD of around 1.0 Å; no uncertainty interval is stated. Source · Abstract The experimental characterization of 310 proteins shows that sampling from Chroma results in proteins that are highly expressed, fold and have favourable biophysical properties. The crystal structures of two designed proteins exhibit atomistic agreement with Chroma samples (a backbone root-mean-square deviation of around 1.0 Å).	Not stated
5 ColabFold: making protein folding accessible to all Abstract and our summary	Accelerate protein structure and complex prediction and make it freely accessible through Google Colaboratory. Source · Abstract ColabFold's 40-60-fold faster search and optimized model utilization enables prediction of close to 1,000 structures per day on a server with one graphics processing unit. Coupled with Google Colaboratory, ColabFold becomes a free and accessible platform for protein folding.	ColabFold combines MMseqs2 homology searches with AlphaFold2 or RoseTTAFold for protein structure and complex prediction. Source · Abstract ColabFold offers accelerated prediction of protein structures and complexes by combining the fast homology search of MMseqs2 with AlphaFold2 or RoseTTAFold.	Not stated	Search is reported to be 40–60 times faster, with optimized model use enabling close to 1,000 structure predictions per day on a single-GPU server. Source · Abstract ColabFold's 40-60-fold faster search and optimized model utilization enables prediction of close to 1,000 structures per day on a server with one graphics processing unit.	Not stated

Where they agree

Both studies aim to predict molecular structures rather than generate new protein designs; the first focuses on protein structures, while the second extends prediction to complexes with diverse molecular components. [1, 2]
4 passages behind it

1 · Question (Abstract): Here we provide the first computational method that can regularly predict protein structures with atomic accuracy even in cases in which no similar structure is known.
1 · Method (Training with labelled and unlabelled data): In this procedure, we use a trained network to predict the structure of around 350,000 diverse sequences from Uniclust3036 and make a new dataset of predicted structures filtered to a high-confidence subset. We then train the same architecture again from scratch using a mixture of PDB data and this new dataset of predicted structures as the training data, in which the various training data augmentations such as cropping and MSA subsampling make it challenging for the network to recapitulate the previously predicted structures.
2 · Question (Abstract): Here we describe our AlphaFold 3 model with a substantially updated diffusion-based architecture that is capable of predicting the joint structure of complexes including proteins, nucleic acids, small molecules, ions and modified residues.
2 · Method (Abstract): Here we describe our AlphaFold 3 model with a substantially updated diffusion-based architecture that is capable of predicting the joint structure of complexes including proteins, nucleic acids, small molecules, ions and modified residues.
Both protein-design frameworks use diffusion-based generation and support steering designs toward specified objectives or constraints. [3, 4]
4 passages behind it

3 · Question (Abstract): Here we show that by fine-tuning the RoseTTAFold structure prediction network on protein structure denoising tasks, we obtain a generative model of protein backbones that achieves outstanding performance on unconditional and topology-constrained protein monomer design, protein binder design, symmetric oligomer design, enzyme active site scaffolding and symmetric motif scaffolding for therapeutic and metal-binding protein design.
3 · Method (Abstract): Here we show that by fine-tuning the RoseTTAFold structure prediction network on protein structure denoising tasks, we obtain a generative model of protein backbones that achieves outstanding performance on unconditional and topology-constrained protein monomer design, protein binder design, symmetric oligomer design, enzyme active site scaffolding and symmetric motif scaffolding for therapeutic and metal-binding protein design.
4 · Question (Abstract): Here we introduce Chroma, a generative model for proteins and protein complexes that can directly sample novel protein structures and sequences, and that can be conditioned to steer the generative process towards desired properties and functions.
4 · Method (Abstract): To enable this, we introduce a diffusion process that respects the conformational statistics of polymer ensembles, an efficient neural architecture for molecular systems that enables long-range reasoning with sub-quadratic scaling, layers for efficiently synthesizing three-dimensional structures of proteins from predicted inter-residue geometries and a general low-temperature sampling algorithm for diffusion models. Chroma achieves protein design as Bayesian inference under external constraints, which can involve symmetries, substructure, shape, semantics and even natural-language prompts.
Both design studies report experimental structural evidence consistent with their generated models: one reports a closely matching cryo-EM structure of a binder complex, and the other reports two crystal structures with backbone RMSD around 1.0 Å. [3, 4]
4 passages behind it

3 · Data (Abstract): We demonstrate the power and generality of the method, called RoseTTAFold diffusion (RFdiffusion), by experimentally characterizing the structures and functions of hundreds of designed symmetric assemblies, metal-binding proteins and protein binders.
3 · Main results (Abstract): The accuracy of RFdiffusion is confirmed by the cryogenic electron microscopy structure of a designed binder in complex with influenza haemagglutinin that is nearly identical to the design model.
4 · Data (Abstract): The experimental characterization of 310 proteins shows that sampling from Chroma results in proteins that are highly expressed, fold and have favourable biophysical properties. The crystal structures of two designed proteins exhibit atomistic agreement with Chroma samples (a backbone root-mean-square deviation of around 1.0 Å).
4 · Main results (Abstract): The experimental characterization of 310 proteins shows that sampling from Chroma results in proteins that are highly expressed, fold and have favourable biophysical properties. The crystal structures of two designed proteins exhibit atomistic agreement with Chroma samples (a backbone root-mean-square deviation of around 1.0 Å).
Both approaches use sequence homology information in protein structure prediction: one takes multiple sequence alignments and available structural templates as inputs, while the other combines accelerated homology searches with existing predictors. [1, 5]
3 passages behind it

1 · Method (Training with labelled and unlabelled data): In this procedure, we use a trained network to predict the structure of around 350,000 diverse sequences from Uniclust3036 and make a new dataset of predicted structures filtered to a high-confidence subset. We then train the same architecture again from scratch using a mixture of PDB data and this new dataset of predicted structures as the training data, in which the various training data augmentations such as cropping and MSA subsampling make it challenging for the network to recapitulate the previously predicted structures.
1 · Data (Inputs and data sources): Inputs to the network are the primary sequence, sequences from evolutionarily related proteins in the form of a MSA created by standard tools including jackhmmer60 and HHBlits61, and 3D atom coordinates of a small number of homologous structures (templates) where available.
5 · Method (Abstract): ColabFold offers accelerated prediction of protein structures and complexes by combining the fast homology search of MMseqs2 with AlphaFold2 or RoseTTAFold.

Where they disagree

The table establishes no direct contradictory findings or common numerical head-to-head comparison among these five papers. The prediction studies report different benchmarks: experimental-level accuracy for a majority of CASP14 cases versus improvements over docking tools, nucleic-acid-specific predictors and an earlier multimer predictor. Neither entry supplies numerical accuracy estimates. [1, 2]
2 passages behind it

1 · Main results (Abstract): We validated an entirely redesigned version of our neural network-based model, AlphaFold, in the challenging 14th Critical Assessment of protein Structure Prediction (CASP14) 15 , demonstrating accuracy competitive with experimental structures in a majority of cases and greatly outperforming other methods.
2 · Main results (Abstract): The new AlphaFold model demonstrates substantially improved accuracy over many previous specialized tools: far greater accuracy for protein–ligand interactions compared with state-of-the-art docking tools, much higher accuracy for protein–nucleic acid interactions compared with nucleic-acid-specific predictors and substantially higher antibody–antigen prediction accuracy compared with AlphaFold-Multimer v.2.3 7,8 .
The design studies differ in the specificity of their reported validation. One describes characterization of hundreds of designs and a cryo-EM match without a numerical structural deviation; the other reports 310 characterized proteins and two crystal structures with backbone RMSD around 1.0 Å. These results do not establish which framework is more accurate. [3, 4]
4 passages behind it

3 · Data (Abstract): We demonstrate the power and generality of the method, called RoseTTAFold diffusion (RFdiffusion), by experimentally characterizing the structures and functions of hundreds of designed symmetric assemblies, metal-binding proteins and protein binders.
3 · Main results (Abstract): The accuracy of RFdiffusion is confirmed by the cryogenic electron microscopy structure of a designed binder in complex with influenza haemagglutinin that is nearly identical to the design model.
4 · Data (Abstract): The experimental characterization of 310 proteins shows that sampling from Chroma results in proteins that are highly expressed, fold and have favourable biophysical properties. The crystal structures of two designed proteins exhibit atomistic agreement with Chroma samples (a backbone root-mean-square deviation of around 1.0 Å).
4 · Main results (Abstract): The experimental characterization of 310 proteins shows that sampling from Chroma results in proteins that are highly expressed, fold and have favourable biophysical properties. The crystal structures of two designed proteins exhibit atomistic agreement with Chroma samples (a backbone root-mean-square deviation of around 1.0 Å).
The accessibility study reports computational efficiency rather than a comparable structural-accuracy result: homology search is 40–60 times faster, and optimized prediction approaches 1,000 structures per day on a single-GPU server. The two structure-prediction studies instead emphasize predictive accuracy without numerical accuracy estimates in the table. [1, 2, 5]
6 passages behind it

5 · Question (Abstract): ColabFold's 40-60-fold faster search and optimized model utilization enables prediction of close to 1,000 structures per day on a server with one graphics processing unit. Coupled with Google Colaboratory, ColabFold becomes a free and accessible platform for protein folding.
5 · Main results (Abstract): ColabFold's 40-60-fold faster search and optimized model utilization enables prediction of close to 1,000 structures per day on a server with one graphics processing unit.
1 · Question (Abstract): Here we provide the first computational method that can regularly predict protein structures with atomic accuracy even in cases in which no similar structure is known.
1 · Main results (Abstract): We validated an entirely redesigned version of our neural network-based model, AlphaFold, in the challenging 14th Critical Assessment of protein Structure Prediction (CASP14) 15 , demonstrating accuracy competitive with experimental structures in a majority of cases and greatly outperforming other methods.
2 · Question (Abstract): Here we describe our AlphaFold 3 model with a substantially updated diffusion-based architecture that is capable of predicting the joint structure of complexes including proteins, nucleic acids, small molecules, ions and modified residues.
2 · Main results (Abstract): The new AlphaFold model demonstrates substantially improved accuracy over many previous specialized tools: far greater accuracy for protein–ligand interactions compared with state-of-the-art docking tools, much higher accuracy for protein–nucleic acid interactions compared with nucleic-acid-specific predictors and substantially higher antibody–antigen prediction accuracy compared with AlphaFold-Multimer v.2.3 7,8 .

What none of them addresses

How do the methods compare on shared, task-appropriate benchmarks with matched datasets, evaluation metrics and computational budgets? The table provides no such controlled comparison.
What are the methods' failure modes, applicability boundaries and explicitly acknowledged limitations? Every limitations entry is marked as not stated.
How uncertain and reproducible are the reported performance estimates? The table gives no confidence intervals, repeated-run variability or uncertainty estimates for the reported numerical results.
How well do the methods capture molecular dynamics, alternative conformations or changes in structure under different environmental conditions? None of the listed questions or results explicitly addresses these issues.
How reliably do designed proteins achieve quantitatively specified biological functions? The table reports structural agreement and favorable biophysical properties but supplies no quantitative functional-performance measurements.
How do training-data overlap and similarity to known structures affect generalization? Although prediction without a similar known structure is an explicit goal, the table provides no stratified results or data-leakage analysis addressing this question.