Deep Template-based Protein Structure Prediction

Wu, Fandi; Xu, Jinbo

doi:10.1101/2020.12.26.424433

Cited by 12 publications

(18 citation statements)

References 38 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…NDThreader [77] and ProALIGN [78] are specifically designed to optimally align the query with the template in template-based modeling. Both methods exploit predicted or observed inter-residue distances to improve the sequence alignments, a strategy that proved powerful already in CASP13 [72,79,80].…”

Section: Leveraging (Meta-)genomicsmentioning

confidence: 99%

Protein sequence‐to‐structure learning: Is this the end(‐to‐end revolution)?

et al. 2021

View full text Add to dashboard Cite

The potential of deep learning has been recognized in the protein structure prediction community for some time, and became indisputable after CASP13. In CASP14, deep learning has boosted the field to unanticipated levels reaching nearexperimental accuracy. This success comes from advances transferred from other machine learning areas, as well as methods specifically designed to deal with protein sequences and structures, and their abstractions. Novel emerging approaches include (i) geometric learning, that is, learning on representations such as graphs, threedimensional (3D) Voronoi tessellations, and point clouds; (ii) pretrained protein language models leveraging attention; (iii) equivariant architectures preserving the symmetry of 3D space; (iv) use of large meta-genome databases; (v) combinations of protein representations; and (vi) finally truly end-to-end architectures, that is, differentiable models starting from a sequence and returning a 3D structure. Here, we provide an overview and our opinion of the novel deep learning approaches developed in the last 2 years and widely used in CASP14.

show abstract

Section: Leveraging (Meta-)genomicsmentioning

confidence: 99%

Protein sequence‐to‐structure learning: Is this the end(‐to‐end revolution)?

et al. 2021

View full text Add to dashboard Cite

show abstract

“…Sequence information is limited to the target protein itself in contrast to RPDD based methods, where multiple sequence alignment information is included as a critical part of input. In AlphaFold, 20 AlphaFold2 58 and many other RPDD based studies, 21,22,[24][25][26][27][28][29]59,60 the core information obtained is explicit protein ( family ) specific RPDD, which are in fact marginalization of the GJD after integrating away all other variables except the distance between the concerning residues. While marginalization in general is an extremely difficulty task in high dimensional space, it is trivial for an approximate GJD represented by a trajectory of configurations with heavy statistical weights confined within the corresponding manifold.…”

Section: Connection To Conventional Ai Driven Protein Structure Studiesmentioning

confidence: 99%

“…The connection among coarse graining, enhanced sampling and LFEL as various forms of applying "dividing and conquering" and "caching" principle in molecular modeling was summarized previously. 19 Like all present protein structure prediction, design and refinement studies, [20][21][22][23][24][25][26][27][28][29] there is an implicit and extremely crude assumption that all high resolution experimental structures were solved under similar environmental (thermodynamic) conditions. Alternatively, differences in thermodynamic and environmental conditions are deemed not important for all high resolution structural data utilized to train models.…”

Section: Introductionmentioning

confidence: 99%

The repetitive local sampling and the local distribution theory

Tian

2021

Preprint

View full text Add to dashboard Cite

Molecular simulation is a mature and versatile tool set widely utilized in many subjects with more than 30,000 publications each year. However, its methodology development has been struggling with a tradeoff between accuracy/resolution and speed, significant improvement of both beyond present state of the art is necessary to reliably substitute many expensive and laborious experiments in molecular biology, materials science and nanotechnology. Previously, the ubiquitous issue regarding severe wasting of computational resources in all forms of molecular simulations due to repetitive local sampling was raised, and the local free energy landscape approach was proposed to address it. This approach is derived from a simple idea of first learning local distributions, and followed by dynamic assembly of which to infer global joint distribution of a target molecular system. When compared with conventional explicit solvent molecular dynamics simulations, a simple and approximate implementation of this theory in protein structural refinement harvested acceleration of about six orders of magnitude without loss of accuracy. While this initial test revealed tremendous benefits for addressing repetitive local sampling, there are some implicit assumptions need to be articulated. Here, I present a more thorough discussion of repetitive local sampling; potential options for learning local distributions; a more general formulation with potential extension to simulation of near equilibrium molecular systems; the prospect of developing computation driven molecular science; the connection to mainstream residue pair distance distribution based protein structure prediction/refinement; and the fundamental difference of utilizing averaging from conventional molecular simulation framework based on potential of mean force. This more general development is termed the local distribution theory to release the limitation of strict thermodynamic equilibrium in its potential wide application in general soft condensed molecular systems.

show abstract

“…This computational graph was successfully utilized to achieve the only end-to-end and the most efficient protein structural refinement pipeline 17 up to date. Like all present protein structure prediction, design and refinement studies, [20][21][22][23][24][25][26][27][28][29] there is an implicit and extremely crude assumption that all high resolution experimental structures were derived under similar environmental (thermodynamic) conditions. Alternatively, differences in thermodynamic and environmental conditions are deemed not important for all high resolution structural data utilized to train models.…”

Section: Introductionmentioning

confidence: 99%

The repetitive local sampling and the local distribution theory

Tian

2021

Preprint

View full text Add to dashboard Cite

Previously, the ubiquitous issue regarding severe wasting of computational resources in all forms of molecular simulations due to repetitive local sampling was raised, and the local free energy landscape approach was proposed to address it. This approach is derived from a simple idea of first learning local distributions, and followed by dynamic assembly of which to infer global joint distribution of a target molecular system. When compared with conventional explicit solvent molecular dynamics simulations, a simple and approximate implementation of this theory in protein structural refinement harvested acceleration of about six orders of magnitude without loss of accuracy. While this initial test revealed tremendous benefits for addressing repetitive local sampling, there are some implicit assumptions need to be articulated.Here, I present a more thorough discussion of repetitive local sampling; potential options for learning local distributions; a more general formulation with potential extension to simulation of near equilibrium molecular systems; generalization of repetitive local sampling to repetitive local computation and potential application in accelerating numerical solving of complex equations; the prospect of developing computation driven molecular science; and the connection to mainstream residue pair distance distribution based protein structure prediction/refinement. 1This more general development is termed the local distribution theory to release the limitation of strict thermodynamic equilibrium in its potential wide application in both soft condensed molecular systems and complex equations.

show abstract

Deep Template-based Protein Structure Prediction

Cited by 12 publications

References 38 publications

Protein sequence‐to‐structure learning: Is this the end(‐to‐end revolution)?

Protein sequence‐to‐structure learning: Is this the end(‐to‐end revolution)?

The repetitive local sampling and the local distribution theory

The repetitive local sampling and the local distribution theory

Contact Info

Product

Resources

About