A deep learning solution for crystallographic structure determination

Ligand Identification using Deep Learning

Motivation

Accurately identifying ligands plays a crucial role in the process of structure-guided drug design. Based on density maps from X-ray diffraction or cryogenic-sample electron microscopy (cryoEM), scientists verify whether small-molecule ligands bind to active sites of interest. However, the interpretation of density maps is challenging, and cognitive bias can sometimes mislead investigators into modeling fictitious compounds. Ligand identification can be aided by automatic methods, but existing approaches are available only for X-ray diffraction and are based on iterative fitting or feature-engineered machine learning rather than end-to-end deep learning.

Results

Here, we propose to identify ligands using a deep learning approach that treats density maps as 3D point clouds. We show that the proposed model is on par with existing machine learning methods for X-ray crystallography while also being applicable to cryoEM density maps. Our study demonstrates that electron density map fragments can be used to train models that can be applied to cryoEM structures, but also highlights challenges associated with the standardization of electron microscopy maps and the quality assessment of cryoEM ligands.

Crystallographic phase identifier of a convolutional self-attention neural network (CPICANN) on powder diffraction patterns

Spectroscopic data, particularly diffraction data, are essential for materials characterization due to their comprehensive crystallographic information. The current crystallographic phase identification, however, is very time consuming. To address this challenge, we have developed a real-time crystallographic phase identifier based on a convolutional self-attention neural network (CPICANN). Trained on 692 190 simulated powder X-ray diffraction (XRD) patterns from 23 073 distinct inorganic crystallographic information files, CPICANN demonstrates superior phase-identification power. Single-phase identification on simulated XRD patterns yields 98.5 and 87.5% accuracies with and without elemental information, respectively, outperforming JADE software (68.2 and 38.7%, respectively). Bi-phase identification on simulated XRD patterns achieves 84.2 and 51.5% accuracies, respectively. In experimental settings, CPICANN achieves an 80% identification accuracy, surpassing JADE software (61%). Integration of CPICANN into XRD refinement software will significantly advance the cutting-edge technology in XRD materials characterization.

Deep-learning map segmentation for protein X-ray crystallographic structure determination

When solving a structure of a protein from single-wavelength anomalous diffraction X-ray data, the initial phases obtained by phasing from an anomalously scattering substructure usually need to be improved by an iterated electron-density modification. In this manuscript, the use of convolutional neural networks (CNNs) for segmentation of the initial experimental phasing electron-density maps is proposed. The results reported demonstrate that a CNN with U-net architecture, trained on several thousands of electron-density maps generated mainly using X-ray data from the Protein Data Bank in a supervised learning, can improve current density-modification methods.

CrysFormer: Protein structure determination via Patterson maps, deep learning, and partial structure attention

Determining the atomic-level structure of a protein has been a decades-long challenge. However, recent advances in transformers and related neural network architectures have enabled researchers to significantly improve solutions to this problem. These methods use large datasets of sequence information and corresponding known protein template structures, if available. Yet, such methods only focus on sequence information. Other available prior knowledge could also be utilized, such as constructs derived from x-ray crystallography experiments and the known structures of the most common conformations of amino acid residues, which we refer to as partial structures. To the best of our knowledge, we propose the first transformer-based model that directly utilizes experimental protein crystallographic data and partial structure information to calculate electron density maps of proteins. In particular, we use Patterson maps, which can be directly obtained from x-ray crystallography experimental data, thus bypassing the well-known crystallographic phase problem. We demonstrate that our method, CrysFormer, achieves precise predictions on two synthetic datasets of peptide fragments in crystalline forms, one with two residues per unit cell and the other with fifteen. These predictions can then be used to generate accurate atomic models using established crystallographic refinement programs.

Unravelling the components of diffuse scattering using deep learning

Many technologically important material properties are underpinned by disorder and short-range structural correlations; therefore, elucidating structure-property relationships in functional materials requires understanding both the average and the local structures. The latter information is contained within diffuse scattering but is challenging to exploit, particularly in single-crystal systems. Separation of the diffuse scattering into its constituent components can greatly simplify analysis and allows for quantitative parameters describing the disorder to be extracted directly. Here, a deep-learning method, DSFU-Net, is presented based on the Pix2Pix generative adversarial network, which takes a plane of diffuse scattering as input and factorizes it into the contributions from the molecular form factor and the chemical short-range order. DSFU-Net was trained on 198 421 samples of simulated diffuse scattering data and performed extremely well on the unseen simulated validation dataset in this work. On a real experimental example, DSFU-Net successfully reproduced the two components with a quality sufficient to distinguish between similar structural models based on the form factor and to refine short-range-order parameters, achieving values comparable to other established methods. This new approach could streamline the analysis of diffuse scattering as it requires minimal prior knowledge of the system, allows access to both components in seconds and is able to compensate for small regions with missing data. DSFU-Net is freely available for use and represents a first step towards an automated workflow for the analysis of single-crystal diffuse scattering.