Recent advances in Artificial Intelligence have enabled multi-modal systems to model and translate diverse information spaces. Extending beyond text and vision, we introduce OneProt, a multi-modal Deep Learning model for proteins that integrates structural, sequence, text, and binding site data. Using the ImageBind framework, OneProt aligns the latent spaces of protein modality encoders in a lightweight fine-tuning scheme that focuses on pairwise alignment with sequence data, rather than requiring full matches. This novel approach comprises a mix of Graph Neural Networks and transformer architectures. It demonstrates good performance in retrieval tasks and showcases the efficacy of multi-modal systems in Protein Machine Learning through a broad spectrum of downstream baselines, including enzyme function prediction and binding site analysis. Furthermore, OneProt enables the transfer of representational information from specialized encoders to the sequence encoder, enhancing capabilities for distinguishing evolutionarily related and unrelated sequences and exhibiting representational properties where evolutionarily related proteins align in similar directions within the latent space. In addition, we extensively investigate modality ablations to identify the encoders that contribute the most to predictive performance, highlighting the significance of the binding site encoder, which has not been used in similar models previously. This work expands the horizons of multi-modal protein models, paving the way for transformative applications in drug discovery, biocatalytic reaction planning, and protein engineering.

MOTIVATION: TopEnzyme is a database of structural enzyme models created with TopModel and is linked to the SWISS-MODEL repository and AlphaFold Protein Structure Database to provide an overview of structural coverage of the functional enzyme space for over 200 000 enzyme models. It allows the user to quickly obtain representative structural models for 60% of all known enzyme functions. RESULTS: We assessed the models with TopScore and contributed 9039 good-quality and 1297 high-quality structures. Furthermore, we compared these models to AlphaFold2 models with TopScore and found that the TopScore differs only by 0.04 on average in favor of AlphaFold2. We tested TopModel and AlphaFold2 for targets not seen in the respective training databases and found that both methods create qualitatively similar structures. When no experimental structures are available, this database will facilitate quick access to structural models across the currently most extensive structural coverage of the functional enzyme space within Swiss-Prot. AVAILABILITY AND IMPLEMENTATION: We provide a full web interface to the database at https://cpclab.uni-duesseldorf.de/topenzyme/.

Glutamine synthetase (GS) catalyzes ATP-dependent ligation of ammonia and glutamate to glutamine. Two mutations of human GS (R324C and R341C) were connected to congenital glutamine deficiency with severe brain malformations resulting in neonatal death. Another GS mutation (R324S) was identified in a neurologically compromised patient. However, the molecular mechanisms underlying the impairment of GS activity by these mutations have remained elusive. Molecular dynamics simulations, free energy calculations, and rigidity analyses suggest that all three mutations influence the first step of GS catalytic cycle. The R324S and R324C mutations deteriorate GS catalytic activity due to loss of direct interactions with ATP. As to R324S, indirect, water-mediated interactions reduce this effect, which may explain the suggested higher GS residual activity. The R341C mutation weakens ATP binding by destabilizing the interacting residue R340 in the apo state of GS. Additionally, the mutation is predicted to result in a significant destabilization of helix H8, which should negatively affect glutamate binding. This prediction was tested in HEK293 cells overexpressing GS by dot-blot analysis: Structural stability of H8 was impaired through mutation of amino acids interacting with R341, as indicated by a loss of masking of an epitope in the glutamate binding pocket for a monoclonal anti-GS antibody by L-methionine-S-sulfoximine; in contrast, cells transfected with wild type GS showed the masking. Our analyses reveal complex molecular effects underlying impaired GS catalytic activity in three clinically relevant mutants. Our findings could stimulate the development of ATP binding-enhancing molecules by which the R324S mutant can be repaired extrinsically.

Protein thermostability is a crucial factor for biotechnological enzyme applications. Protein engineering studies aimed at improving thermostability have successfully applied both directed evolution and rational design. However, for rational approaches, the major challenge remains the prediction of mutation sites and optimal amino acid substitutions. Recently, we showed that such mutation sites can be identified as structural weak spots by rigidity theory-based thermal unfolding simulations of proteins. Here, we describe and validate a unique, ensemble-based, yet highly efficient strategy to predict optimal amino acid substitutions at structural weak spots for improving a protein's thermostability. For this, we exploit the fact that in the majority of cases an increased structural rigidity of the folded state has been found as the cause for thermostability. When applied prospectively to lipase A from Bacillus subtilis, we achieved both a high success rate (25% over all experimentally tested mutations, which raises to 60% if small-to-large residue mutations and mutations in the active site are excluded) in predicting significantly thermostabilized lipase variants and a remarkably large increase in those variants' thermostability (up to 6.6°C) based on single amino acid mutations. When considering negative controls in addition and evaluating the performance of our approach as a binary classifier, the accuracy is 63% and increases to 83% if small-to-large residue mutations and mutations in the active site are excluded. The gain in precision (predictive value for increased thermostability) over random classification is 1.6-fold (2.4-fold). Furthermore, an increase in thermostability predicted by our approach significantly points to increased experimental thermostability (p < 0.05). These results suggest that our strategy is a valuable complement to existing methods for rational protein design aimed at improving thermostability.

UNLABELLED: Constraint network analysis (CNA) is a graph theory-based rigidity analysis approach for linking a biomolecule's structure, flexibility, (thermo)stability and function. Results from CNA are highly information-rich and require intuitive, synchronized and interactive visualization for a comprehensive analysis. We developed VisualCNA, an easy-to-use PyMOL plug-in that allows setup of CNA runs and analysis of CNA results linking plots with molecular graphics representations. From a practical viewpoint, the most striking feature of VisualCNA is that it facilitates interactive protein engineering aimed at improving thermostability. AVAILABILITY AND IMPLEMENTATION: VisualCNA and its dependencies (CNA and FIRST software) are available free of charge under GPL and academic licenses, respectively. VisualCNA and CNA are available at http://cpclab.uni-duesseldorf.de/software; FIRST is available at http://flexweb.asu.edu.