Page 102 - AJWEP-v22i3

Sonsare, et al.

Perform forward propagation to obtain or max) combines node-level representations into a
predictions . graph-level embedding.
Compute test accuracy. This embedding is subsequently processed using fully
Calculate additional performance metrics: connected layers to generate the expected values for K ,
ow
Confusion Matrix K , and K . The model is trained using the MSE loss
aw
d
AUC (Area Under the ROC Curve). function, optimized with the Adam optimizer (learning
Precision, Recall rate: 0.001, weight decay: 5e-4), and regularized with
------------------------------------------------------------------ dropout (p=0.3) to prevent overfitting. In addition, data
augmentation methods such as random edge masking
3.4. Scalability and robustness validation are used to improve model resilience.
Further experiments were carried out to confirm The dataset is divided into 70% training, 15%
the scalability and resilience of the suggested GNN validation, and 15% test sets with stratified sampling
architectures. The models underwent training and to balance different partition coefficient ranges. Model
evaluation using the QM9 and ZINC datasets, which performance is tested using mean absolute error (MAE),
feature considerably more samples and intricate RMSE, and R² scores, with expected results indicating
molecular graphs than MUTAG. Metrics such as good predictive power (R scores of 0.88 for K , 0.85
2
training duration, memory consumption, and precision for K , and 0.91 for K ). ow
were observed to evaluate the scalability of every aw d
architecture. Noise was added to the node features and 4. Experimental results
edge connections in the MUTAG dataset to mimic real-
world flaws. The models were additionally assessed The models were evaluated on the MUTAG dataset
on graphs with randomly omitted nodes or edges to with a stratified train–test split (80 – 20%) over five
check their robustness against incomplete input data. runs. The metrics used for evaluation were accuracy,
The models that were trained on MUTAG underwent precision, recall, and ROC-AUC. GIN achieved the
testing on a QM9 subset to evaluate their generalization highest accuracy and ROC-AUC score, showcasing its
capabilities. superior ability to capture graph structure with its sum-
based aggregation. GAT demonstrated competitive
3.5. Partition coefficients identification performance, particularly in recall, due to its attention
A structured technique was used to estimate the mechanism that highlights relevant neighbors. The
Kₒw, Kₐw, and K_d utilizing the MoleculeNet ROC-AU score is a performance metric used to evaluate
dataset and GINs. The first phase comprises data the classification ability of machine learning models.
gathering and preparation, where key datasets such It quantifies the area under the ROC curve, which
as FreeSolv (for solubility, connected to ), ESOL plots the true positive rate against the false positive
d
(for aqueous solubility, beneficial for K ), and rate at various threshold levels. A higher ROC-AU
ow
Lipop (for lipophilicity, directly related to K ) are score (closer to 1) indicates better model performance,
ow
picked. Because MoleculeNet provides molecular whereas a score of 0.5 suggests random guessing.
representations in SMILES format, they are converted GCN showed slightly lower accuracy and ROC-AUC,
into graph structures with RDKit, which extracts potentially due to its limitations in handling complex
node features (atomic types, hybridization, and graph structures. Figure 1 shows the analysis of loss
electronegativity), edge features (bond orders and and accuracy.
aromaticity), and global molecular features (molecular All models showed a consistent reduction in training
weight and polar surface area). Partition coefficient and validation loss. GIN converged faster than GCN
values are standardized with log transformations to and GAT. GIN and GAT exhibited a stable increase in
guarantee model stability. accuracy, with GCN slightly lagging.
GINs are used to predict molecular properties The relationship between each model’s true positive
because of their high capacity to differentiate graph rate and false positive rate is depicted by the ROC
structures, as suggested by the Weisfeiler–Lehman curves. Figure 2 depicts the confusion matrix and
test. The model uses a five-layer GIN with sum shows the ROC curve comparison. GIN demonstrates
aggregation to adequately represent structural and its superior ability to capture molecular structure,
physicochemical features. After passing through achieving the best performance across all metrics.
the GIN layers, a global pooling layer (mean, sum, GAT effectively leverages attention mechanisms to

Volume 22 Issue 3 (2025) 96 doi: 10.36922/AJWEP025070041