PhD Scientific Days 2026

Poster Session 3.W - Pharmaceutical Sciences and Health Technologies

ProteGAN: a Generative, Self-Supervised Model with Edge-Level Attention for Network-Based Protein–Protein Interaction Prediction

Előadó neve

Balogh, Olivér, MSc

Neptune code

HYKT8D

Előadó munkahelye

Semmelweis University Department of Pharmacology and Pharmacotherapy

Előadó telefonszáma

06203685079

Előadó e-mail címe

balogh.oliver.marton@semmelweis.hu

Az előadás címe

ProteGAN: a Generative, Self-Supervised Model with Edge-Level Attention for Network-Based Protein–Protein Interaction Prediction

Szerző(k) neve és munkahelye

Olivér M. Balogh¹, Lóránt Sagát¹, András Horváth², Péter Ferdinandy¹, Bence Ágg¹

1: Department of Pharmacology and Pharmacotherapy, Semmelweis University, Budapest
2: Faculty of Information Technology and Bionics, Pázmány Péter Catholic University, Budapest

Bemutatás módja

Poszter

Szekció

Poster Session 3.W - Pharmaceutical Sciences and Health Technologies

Language of the presentation

English

Preferred session

Pharmaceutical Sciences and Health Technologies

Összefoglaló szövege

Introduction
Exploring protein–protein interactions (PPIs) remains a challenging task of drug target discovery due to the time and cost restrains of experimental methods. Numerous in silico prediction tools have been proposed to aid the selection of likely candidate protein pairs. In our prior work (DOI: 10.1186/s12859-022-04598-x), we introduced a conditional generative adversarial network model for the efficient prediction of PPIs, relying solely on the topology of PPI networks.

Aims
Here, our aim was to expand our model architecture with transformer-inspired encoder blocks in order to further improve the performance of our model.

Methods
ProteGAN, the next iteration of our model now includes several encoder blocks before the convolutional u-net part, providing context-aware information via edge-level attention. Version 12 of the STRING (Search Tool for the Retrieval of Interacting Genes/Proteins) database was used as the primary source of networks for training ProteGAN. The area under the precision-recall curve (AUPRC) and the precision in the top 500 predictions (P@500) were used as performance metrics.

Results
Our evaluations on the available PPI networks reveal the importance of simple image preprocessing techniques and state of the art neural architectures. On the H. sapiens network, utilizing 0.7 confidence threshold, our old model achieved AUPRC: 0.092 and P@500: 0.260, while with image preprocessing techniques and encoder blocks it improved to AUPRC: 0.415 and P@500: 0.966.

Conclusion
In conclusion, here we introduce ProteGAN, a generative adversarial network outfitted with transformer-inspired encoder blocks for the prediction of PPIs based on the topology of the interactome. ProteGAN shows considerable performance improvements compared to our previous model and as such, could be a valuable tool for the high throughput prediction of candidate PPIs, contributing to a more efficient drug target identification process.

Funding
Project no. RRF-2.3.1-21-2022-00003 has been implemented with the support provided by the European Union. This study was supported by the 2024-1.2.3.-HU-RIZONT-2024-00026 and the 2025-2.1.1-EKÖP-2025-00014 New National Excellence Program of the Ministry for Culture and Innovation from the source of the National Research, Development and Innovation Fund. O.M.B. was supported by Semmelweis 250+ Excellence Fellowship.

University

Semmelweis University

Supervisor

Dr. Bence Ágg

Publication of my abstract

I do not give consent to the publication of my abstract on the website of the congress.

phd.section.field

in doctoral studies after complex exam (PhD)

Kind

Szabad

Status

elfogadva

Accepted presentation method

poszter

Előadás fájl jóváhagyás

nem rendelkezett róla

Előadó

7403

Start

14:36

End

14:39