About DeepUMQA server

Overview of DeepUMQA server

The DeepUMQA server is a web-based platform for model quality assessment of protein monomer and complex structure. Figure 1(A) shows the pipeline of DeepUMQA. For the input complex structure, overall complex features, intra-monomer features, and inter-monomer features are used to represent it. An enhanced residual neural network is proposed to assess the quality of protein models. The backbone of the network consists of triangle update, axial attention, and feed-forward layers to update residual pair information. Two branches of the network based on residual network are used to predict the residue-residue distance deviation and the contact map with a threshold of 15Å, respectively, which are used to calculate per-residue lDDT and interface residues accuracy. For the input monomer structure, model-dependent features, MSA features and template features are used to describe it. The enhanced residual neural network same as the complex structure is used to predict the lDDT of all residues. Figure 1(B) shows detailed features of the complex structure. A complex structure is represented from three levels. The first is the overall complex level, which treat the complex as one large monomer structure. The residue order-independent features, such as monomer USR, residue voxelization, residue-pair distance and orientations, and amino acid properties, are extracted. The second is the intra-monomer level, the residue order dependent features, such as sequence embeding, secondary structure, and rosetta energy terms, are extracted for each monomer structures. The third is inter-monomer level, the paired sequence attention and complex USR are used to characterize the inter-monomer relationships from sequence and structure levels, respectively. Figure 1(C) shows the schematic of Monomer USR. For each residue O_i of the structure, the farthest residue f_i from it, and the farthest residue ff_i from f_i are selected. These three positions represent the center of residue Oi and its extremes in the structure. The distance from each residue in the structure to these three residues constitute three distance distributions, and the first three moments of all distribution are used to characterize the relationship between residues and the overall structure. Figure 1(D) shows the schematic of complex USR. For each residue i of monomer A, the nearest residue r from it, the farthest residue m from it, and the farthest residue n from residue m in monomer B are selected as representative residues. The first three moments of the distance distributions of all residues in monomer B to these four residues, and the spatial position relationship between residue i and the triangle form by these three representative residues are extracted, to characterize the relationship between the residue of one monomer and the topology of other monomers.

logo Figure 1. Overview of DeepUMQA3 for model quality assessment of protein monomer and complex structure.

Performance of DeepUMQA

DeepUMQA3 (Group name: GuijunLab-RocketX) ranked first in CASP15’s complex interface residue accuracy estimation, and significantly outperforms other competing methods (as shown in Figure 2). DeepUMQA3 achieved the best performance on 27 out of 39 targeets (3 of which were not submitted in time due to computational resource issues), as shown in Figure 3.

Figure 2. Z-score ranking of interface residue accuracy estimaton.

Figure 3. Per-target performance of top5 groups in interface residue accuracy estimaton.

DeepUMQA ranked first in the 1-year (2021-12-03 to 2022-11-26) blind test of protein monomer model quality assessment of CAMEO (as shown in Figure 4), and DeepUMQA2 shows state-of-the-art performance in the continuous blind test of CAMEO ( Link ).

Figure 4. CAMEO 1-year blind test ranking.

Input of DeepUMQA server

Monormer structure assessment:

(Mandatory) At least 1 protein monomer structure model in PDB format
Users can click the button "Add domain" to add text boxes for input more structure models of same protein monomer. The server page is set to receive up to 10 structure models. If you need to evaluate more than 10 structure models at the same time, please email all structural models to guijunlab01@163.com. Please set the email subject as "Request DeepUMQA Batch Evaluation".
(Optional) Choose not to use templates
(Optional) Choose not to use any homologous sequences
(Optional) Provide your email address
(Optional) Assign your target name

Figure 5. The "Submit" section of monomer structure assessment in DeepUMQA home page.

Complex structure assessment:

(Mandatory) At least 1 protein complex structure model in PDB format
Users can click the button "Add domain" to add text boxes for input more structure models of same protein complex The server page is set to receive up to 10 structure models. If you need to evaluate more than 10 structure models at the same time, please email all structural models to guijunlab01@163.com. Please set the email subject as "Request DeepUMQA Batch Evaluation".
(Optional) Provide your email address
(Optional) Assign your target name

Figure 6. The "Submit" section of complex structure assessment in DeepUMQA home page.

Output of DeepUMQA server

Monormer structure assessment:

The evaluated monomer structure model
Inter-residue distance deviation of the monomer structure model
Per-residue lDDT and global lDDT of the monomer structure model
Ranking of all input monomer structure model (According to the global lDDT)
The refined monomer structure model (If you choose to refine the input monomer structure model)

Figure 7. The output of monomer structure assessment in DeepUMQA result page.

Complex structure assessment:

The evaluated complex structure model
Per-residue lDDT and global lDDT of the complex structure model
Interface residues lDDT of the complex structure model
Interface residues are defined as having a contact with at least one residue from another chain (CB-CB distance ≤ 8Å, CA in case of Glycine).
The refined monomer structure model (If you choose to refine the input monomer structure model)

Figure 8. The output of complex structure assessment in DeepUMQA result page.

How to cite DeepUMQA?

Please cite the following articles when you use the trRosetta server:

Jun Liu, Kailong Zhao and Guijun Zhang*. Improved model quality assessment using sequence and structural information by enhanced deep neural networks. Briefings in Bioinformatics, 2022, https://doi.org/10.1093/bib/bbac507.
Sai-Sai Guo^✝, Jun Liu^✝, Xiao-Gen Zhou, Gui-Jun Zhang*. DeepUMQA: Ultrafast Shape Recognition-based Protein Model Quality Assessment using Deep Learning. Bioinformatics, 2022, 38(7): 1895-1903.

Need more help?

If you have more questions or comments about the server, please eamil guijunlab01@163.com.