Page 152 - SRPSKO DRUŠTVO ISTRAŽIVAČA RAKA

Serbian Association for Cancer Research SDIRSACR

and stem-cell markers. These molecular tools often outperform classical markers but still leave considerable prognostic
uncertainty. Because outcome prediction in early breast cancer has the potential to guide adjuvant therapy, we sought
to build a deep-learning pipeline that delivers superior accuracy compared with existing approaches. Rationale and
Aim Classification of histopathology images by use of convolutional neural networks has typically been approached
by a single stain and a single image format. We reasoned that complementary stains and multiple colour-depth
representations would expose richer morphology and therefore boost prognostic power. We focused on early-stage
disease, where accurate risk stratification can spare low-risk patients from overtreatment and recommend high-risk
patients toward intensified therapy. Our goal was to predict distant metastasis with maximal accuracy.
Materials and Methods: Whole-slide tumour sections from early breast cancer patients underwent two laboratory
stains: 1. AE1/AE3 pan-cytokeratin (pan-CK), highlighting epithelial malignant patches. 2. Hematoxylin and eosin (H&E),
visualising global histology. Slides were scanned and exported in three image formats: RGB colour, 8-bit grayscale
and binary masks produced at five intensity thresholds, yielding seven biologically distinct datasets per patient.
The final training set thus comprised 2,646 histopathological tumour images. Deep-Learning Pipeline A pretrained
ResNet-50 served as a feature extractor and classifier. We implemented experimental data augmentation rather than
the traditional virtual approach. The data augmentation was thus derived from physically different stains and image-
depth transformations, not from synthetic flips or rotations. Strict separation of development and test sets prevented
data leakage and ensured an unbiased estimate of performance.
Results: Single-dataset performance • Grayscale pan-CK: accuracy 94.4 % AUC 0.982 • Grayscale H&E: accuracy 85.7
% AUC 0.992 Pooled multistain–multidepth model Training on all 14 datasets (two stains x 7 image depths) pushed
performance to 100 % accuracy and an AUC of 1.000 on the test set, confirming that heterogeneous visual inputs
supply complementary prognostic information. Discussion Our findings show that microscopic morphology still holds
untapped prognostic information that is only revealed when a deep learning network is provided with multiple stains
and colour depths. Experimental augmentation with pan-CK and H&E stains provided biologically meaningful diversity.
Grayscale images provided better discriminative performance than full-colour or binary images, yet the merged model
still benefited from every representation. Previous cancer studies combining stains report similar gains, but to our
knowledge, none have combined multiple staining with systematic colour-depth variation and binary thresholding.
Our work supports expanding histopathology pipelines beyond H&E staining and RGB imaging to capture more
comprehensive information about a tumour.
Conclusions: Conclusions and Outlook Deep transfer learning on complementary stains and colour-depth representations
provides better prognostic accuracy for early breast cancer than single-stain models or traditional clinicopathological
markers. This strategy can be integrated with radiomics, genomics, proteomics, and laboratory data to further refine
personalised therapy. Because patients with early disease have the widest treatment choices, they stand to gain the
most from the precise risk estimates offered by this multistain, multidepth deep-learning framework.

Acknowledgments and funding: This research was funded by the MINISTRY OF SCIENCE, TECHNOLOGICAL
DEVELOPMENT AND INNOVATION of the Republic of Serbia, grant number 451-03-136/2025-03/ 200043.

P61

Exploring the Structural Impacts of Missense Variants in USP7 from the perspective of effective inhibitor
selection and MDM2-p53 coordination via classical MD simulations and docking studies

Buse Meriç Açar1, Engin Ulukaya , Aslı Kutlu 3,4
2,3
1Cancer Biology and Pharmacology, Health Science Institute, Istinye University, Istanbul-Türkiye
2Clinical Biochemistry, Faculty of Medicine, Istinye University, Istanbul-Türkiye
3Istinye Molecular Cancer Research Center, Istinye University, Istanbul-Türkiye
4Molecular Biology and Genetics, Faculty of Engineering and Natural Science, Istinye University, Istanbul-Türkiye

Keywords: USP7, p53/MDM2, AutoDock Vina, Molecular Docking, Molecular Dynamic Simulation

Background: Ubiquitin-specific protease 7 (USP7) is a deubiquitinating enzyme involved in the regulation of multiple
signalling pathways. One of the crucial roles of USP7 is the regulation of the tumour suppressor p53 together with
MDM2. Under normal physiological conditions, USP7 indirectly suppresses p53 by stabilizing MDM2, but it stabilizes
p53, thereby inducing apoptosis, under genotoxic stress. The presence of certain USP7 mutations causes excessive
MDM2 stabilization, which triggers p53 degradation and therefore allows cancer cells to escape apoptosis. Due to this

137