In silico identifikacija vezavnih aminokislinskih ostankov v Cu$^{2+}$/Cu$^+$ in Zn$^{2+}$ metaloproteinih z grafovskimi nevronskimi mrežami ter kvantnomehanska analiza njihove koordinacijske sfere

Krašna, Aljaž (author), Hribar Lee, Barbara (mentor)

Kovinski ioni so eni najpogostejših proteinskih kofaktorjev. Delovanje številnih proteinov je odvisno od njihove vezave, saj sodelujejo pri različnih biokemijskih procesih, kot sta stabilizacija strukture in celična signalizacija. Ker so eksperimentalni postopki za identifikacijo in analizo vezavnih mest dragi in zamudni, so se uveljavile različne računske metode. V tej diplomski nalogi smo razvili grafovsko nevronsko mrežo za identifikacijo aminokislinskih ostankov, ki posredno ali neposredno sodelujejo pri vezavi Cu$^{2+}$/Cu$^+$ in Zn$^{2+}$. Proteinske strukture, pridobljene iz podatkovne baze RCSB PDB, smo predstavili v obliki neusmerjenih grafov, kjer vsako vozlišče ustreza enemu aminokislinskemu ostanku. Vozlišča je nevronska mreža (GAT) razvrstila v 3 razrede: aminokislinske ostanke v primarni in sekundarni sferi ter ostale nevezne Napovedane vezavne aminokislinske ostanke smo nato z algoritmom DBSCAN združili v posamezna vezavna mesta. Ker nam naučena nevronska mreža ne poda nobene fizikalne oziroma kemijske informacije o vezavnem mestu, smo izbrana vezavna mesta dodatno analizirali s pomočjo metod kvantne mehanike. Celotno strukturo smo obravnavali s pomočjo teorije gostotnega funkcionala DFT (funkcional r$^2$SCAN-D4 in bazni set def2-TZVPP). Izbranim kompleksom smo optimizirali geometrijo ter izračunali prosto Gibbsovo energijo vezave, analizirali spinsko gostoto (pri Cu$^{2+}$) in izbrane molekulske orbitale. Model je dosegel zadovoljive rezultate, in sicer je za metaloproteine z Zn$^{2+}$ korelacijski koeficient Matthewsa (MCC) znašal 0,51, za metaloproteine s Cu$^{2+}$/Cu$^+$ pa nekoliko manj, 0,38. Razliko v uspešnosti pripisujemo bistveno manjšemu naboru razpoložljivih struktur s kovinskim ionom Cu$^{2+}$/Cu$^+$ (211 proti 2333 struktur). Kvantnomehanska analiza je potrdila pričakovane razlike v vezavni afiniteti in geometriji posameznih kovinskih kompleksov. Razviti pristop ponuja osnovo za nadaljnje delo, predvsem za razširitve na druge kovinske ione in za povečanje nabora struktur, kar bi lahko izboljšalo zlasti napovedne sposobnosti Cu$^{2+}$ kompleksov.

• Language: Slovenian
• Keywords: DFT, strojno učenje, grafovske nevronske mreže, metaloproteini, kvantna mehanika
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FKKT - Faculty of Chemistry and Chemical Technology
• Year: 2026
• PID: 20.500.12556/RUL-184482
• COBISS.SI-ID: 288320515
• Publication date in RUL: 08.07.2026

Secondary language

• Language: English
• Title: In silico identification of metal-binding amino acid residues in Cu$^{2+}$/Cu$^+$ and Zn$^{2+}$ metalloproteins using graph neural networks and quantum-mechanical analysis of the coordination sphere

Abstract

Metal ions are among the most common protein cofactors. The proper functioning of many proteins depends on their binding, as these ions participate in different biochemical processes such as structural stabilisation and cellular signaling. Because experimental procedures for identification of binding sites are expensive and time-consuming, computational methods have become well established. In this thesis we developed a graph neural network for the identification of amino acid residues that participate directly or indirectly in the binding of Cu$^{2+}$/Cu$^+$ and Zn$^{2+}$. Protein structures obtained from RCSB PDB were represented as undirected graphs, in which each node corresponds to a single amino acid residue. Using graph attetion convolutional layers (GAT), these were classified into three classes, namely amino acid residues in the primary and secondary coordination spheres of metal ions, and other non-binding amino acid residues. The predicted binding residues were then merged into individual binding sites using the DBSCAN algorithm. Since the trained netowrk provides no physical or chemical information about the binding site, the selected binding sites were further analyzed using quantum-mechanical methods. The protein structure was treated on the level of density functional theory (DFT functional r$^2$SCAN-D4 and basis set def2-TZVPP). We optimised the geometry of the selected complexes and calculated the Gibbs free binding energy, spin density (for Cu$^{2+}$) and their molecular orbitals. The model achieved solid results, namely for metalloproteins with with Zn$^{2+}$ (MCC = 0.51) and with Cu$^{2+}$/Cu$^+$ (MCC = 0.38). We attribute the difference in results to the substantially smaller set of available structures for Cu$^{2+}$/Cu$^+$ (211 vs. 2333 structues). The quantum-mechanical analysis confirmed the expected differences in binding affinity and geometry of individual metal complexes. The differences between motifs were relatively small, which we attribute to the absence of the secondary coordination sphere in our simplified models. The developed approach provides a basis for further works, in particular extension to other metal ions and enlargment of the structure set, which could improve predictive performance especially for Cu$^{2+}$ complexes.

• Keywords: DFT, machine learning, graph neural networks, metalloprotein, quantum mechanics