Modeling interactions between TgrB1 and TgrC1 proteins

Mitrev, Milosh (author), Curk, Tomaž (mentor)

This thesis addresses the prediction of interaction compatibility between TgrB1 and TgrC1 protein variants in the social amoeba Dictyostelium discoideum, which play a key role in cell recognition and aggregation. A collection of known interacting and non-interacting protein pairs was constructed. Their three-dimensional protein structure models predicted using AlphaFold3 were used together with features derived from protein docking simulations performed with HADDOCK. In addition, protein sequence embeddings obtained with protein language models were included as alternative feature representations of protein sequences. A logistic regression model was trained on the combined feature set to estimate the probability of interaction for protein pairs. During evaluation, one positive and one negative pair were held out, and a prediction was considered correct when the positive pair received a higher interaction probability. The results show that structural features, particularly those derived from AlphaFold and docking, have the strongest impact on predictive performance, while sequence embeddings do not improve accuracy. The best model achieves up to 87% pairwise accuracy.

• Language: English
• Keywords: protein-protein interactions, tgrB1, tgrC1, Dictyostelium discoideum, AlphaFold3, HADDOCK, machine learning
• Work type: Bachelor thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2026
• PID: 20.500.12556/RUL-187590
• Publication date in RUL: 11.09.2026

Secondary language

• Language: Slovenian
• Title: Modeliranje interakcij med proteinoma TgrB1 in TgrC1

Abstract

V diplomski nalogi obravnavamo napovedovanje interakcij med proteinoma TgrB1 in TgrC1 pri socialni amebi Dictyostelium discoideum, ki imata ključno vlogo pri celičnem prepoznavanju in agregaciji. Sestavili smo zbirko znanih interakcij in neinterakcij ter uporabili napovedane tridimenzionalne strukturne modele proteinov, napovedane z orodjem AlphaFold3, skupaj z značilkami, pridobljenimi s simulacijami proteinskega sidranja s programom HADDOCK. Dodatno smo vključili sekvenčne predstavitve proteinov, pridobljene s proteinskimi jezikovnimi modeli. Na združenih značilkah smo učili model logistične regresije za ocenjevanje verjetnosti interakcije med proteinskimi pari. Pri vrednotenju sta bila izločena en pozitiven in en negativen par, napoved pa je bila pravilna, če je pozitiven par prejel višjo verjetnost interakcije. Dosegli smo do 87% točnost pri vrednotenju parov, pri čemer se je izkazalo, da imajo strukturne značilke največji vpliv na uspešnost napovedi, medtem ko sekvenčne predstavitve ne prispevajo k izboljšanju rezultatov.

• Keywords: proteinske interakcije, tgrB1, tgrC1, Dictyostelium discoideum, napovedovanje strukture proteinov, AlphaFold3, HADDOCK, strojno učenje, logistična regresija