University of Crete (UOC)
Classifier-driven generative adversarial networks with applications in protein design
Abstract
dc:descriptionDesigning proteins with specific biological properties is a critical area in bioengineering, drug discovery, and therapeutic development. Traditional experimental methods for protein design are expensive, time-consuming, and lack scalability. Computational approaches, especially Generative Adversarial Networks (GANs), have shown transformative potential in protein design. However, their inherent lack of control over generated outputs poses a significant challenge in designing proteins with specific biological activities such as structural stability, antimicrobial efficacy, and low toxicity. This dissertation introduces a classifier-driven GAN framework that integrates predictive feedback from a classifier into the generative process. By dynamically adjusting the GAN’s loss function based on classifier predictions, the framework guides the generator to produce outputs that align with predefined criteria, thereby enhancing the accuracy and reliability of the generated protein sequences. Furthermore, the classifier is extended to multitask learning, enabling the GAN to generate peptides that satisfy multiple functional requirements, a capability not yet explored in the literature. This innovation addresses key challenges in protein design, offering a scalable and precise solution to balance diverse real-world functional demands. The development of robust classifiers is a cornerstone of this framework. The thesis investigates a broad range of computational techniques, including signal processing, graph-based methods, natural language processing, and transfer learning with large protein language models. Following extensive experimentation, the top-performing classifier was adapted to a multitask learning paradigm, enabling the simultaneous optimization of multiple peptide properties, such as antimicrobial activity and structural stability. The main application of this framework is the design of antimicrobial peptides (AMPs), molecules that play a crucial role in the immune defense of many organisms by exhibiting strong activity against a wide range of pathogens, including drug-resistant bacteria. Recognized as promising alternatives to traditional antibiotics, AMPs address the urgent global challenge of antimicrobial resistance. However, their discovery and optimization are hindered by the complexity of balancing efficacy, stability, and specificity. Our classifier-driven GAN framework streamlines this process by generating peptides with customized antimicrobial activity and structural properties, offering a powerful solution to combat antimicrobial resistance and advance therapeutic innovation. In summary, this thesis demonstrates the capability of classifier-driven GANs to generate bioactive molecules with tailored functionalities, offering a precise and scalable alternative to traditional approaches. The findings highlight the transformative potential of this framework in computational peptide engineering, enabling controlled synthesis and driving innovation in protein design and related fields.
Degree
thesis:*- Grantor dc:publisher
- University of Crete (UOC)
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Zervou, Michaela-Areti
- Ζερβού, Μιχαέλα-Αρετή
Subjects
dc:subject × 18- Παραγωγικά ανταγωνιστικά δίκτυα
- Ταξινόμηση δευτεροταγούς δομής
- Ταξινόμηση αντιμικροβιακών πεπτιδίων
- Σχεδίαση πρωτεϊνών
- Μάθηση πολλαπλών ιδιοτήτων
- Καθοδηγούμενα από ταξινομητή παραγωγικά ανταγωνιστικά δίκτυα
- Generative adversarial networks
- Secondary structure classification
- Antimicrobial peptide classification
- Protein design
- Multi-task learning
- Classifier-driven generative adversarial networks
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Τεχνητή νοημοσύνη
- Computer and Information Sciences
- Natural Sciences
- Artificial Intelligence
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/58573