AI-driven advances in modeling of protein structure

Overview

Results from the most recent CASP (Critical Assessment of Structure Prediction) experiment show dramatic improvement in computing the three-dimensional structure of proteins from amino acid sequence, with many models rivaling experimental structures in accuracy. These results suggest that deep learning approaches will also be effective for a range of related structural biology applications, including macromolecular assemblies, ligand docking, alternative conformations, disordered states, interpretation of genetic variants, and protein design. The session will bring together researchers from the computational structural biology and machine learning communities to explore this new landscape. We invite contributions addressing relevant questions of methodology, applications, and synergies with experimental structural biology.

Expected progress

Recent developments open the door to further expansion with substantial practical implications:

• Modeling of protein complexes. Although much work is still needed, development of the new generation methods with application to protein complexes is following in the footsteps of single protein methods.

• Protein design. The new generation methods should greatly reduce the need for experimental testing in this area.

• Genetic variant interpretation. Deep learning methods, together with accurate structure models are likely to substantially improve our ability to distinguish pathogenic and benign variants.

• Drug development. As in the case of protein complexes, the successful development of deep-learning methods for docking small molecules to proteins is following suit. Better ligand docking methods should speed up the identification of lead compounds, including drug repurposing opportunities.

• Synergy with experimental structural biology. We expect that computed models will be broadly used to help solve crystal and cryo-EM structures. This is already increasingly seen in CASP, where multiple examples of such applications are already available.

Session topics

Session topics could include the following. Other related topics could be of interest:

• MSA techniques: sensitivity, deep alignments, and using metagenome databases for protein structure prediction

• MSA-free approaches

• Deep template-based modeling

• End-to-end learning: from sequence to structure

• NLP models and architectures for protein sequences

• Protein geometry learning and representation: graphs, point clouds, tessellations, and more

• 3D CNNs and 3D transformers

• Symmetries and equivariance in 3D

• Novel DL architectures: from convolutions to attentions

• Deep learning for model quality estimation and model refinement

• DL-boosted conformational space exploration

• Applications beyond protein structure prediction: protein design, prediction of multiple functional states, predicting protein assemblies, structures of RNAs and IDPs

• Modeling in structural biology: new horizons for X-ray crystallography, NMR, cryo-EM, and SAXS

Submission information

Plesae consult the PSB website for more instructions, https://psb.stanford.edu

• Paper submissions due: August 2, 2021 (firm deadline)

• Notification of paper acceptance: September 13, 2021

• Camera-ready final paper deadline: October 1, 2021

• Abstract deadline for non-reviewed posters: December 6, 2021

All deadlines are at midnight Pacific Standard Time.

PSBs paper format template and instructions are available at http://psb.stanford.edu/psb-online/psb-submit

Articles in the PSB proceedings are archival, rigorously peer-reviewed publications. PSB publications are Open Access and linked directly from MEDLINE/PubMed and Google Scholar for wide accessibility. They are equivalent to journal articles that may be cited on CVs and grant reports.

Session co-chairs