Page 96 - AJWEP-v22i3

Sonsare, et al.

compounds. With the increasing availability of curated as vertices (nodes) and bonds as edges. This graph
datasets and advances in deep learning, particularly structure allows GNNs to propagate information through
GNNs, machine learning models now offer a promising bonded atoms, capturing both topological and chemical
alternative to traditional quantitative structure- features essential for property prediction. GNNs offer
activity relationship (QSAR) models. This study a revolutionary method for predicting molecular
aims to investigate the applicability of modern GNN properties by allowing for the direct handling of graph-
architectures to these tasks and assess their reliability structured information. The GNN model utilizes a
across different molecular datasets. recursive framework to disseminate information across
The organization of this paper is outlined as graph structures. Kipf and Welling presented GCNs
16
4
follows: Section 2 explores pertinent studies that that generalized conventional convolutions to graphs
illustrate advancements in GNN-based molecular through the aggregation of information from a node’s
property forecasting. The methodology detailing neighboring nodes, utilizing a normalization approach.
model architectures, training methods, and dataset GCNs emerged as a favored option for semi-supervised
preparation is thoroughly explained in Section 3. The learning tasks, including the classification of molecules.
findings from the experiment are presented in Section Nonetheless, GCNs frequently faced issues with over-
4, and the discussion of those results is in Section 5. smoothing, resulting in node representations becoming
The conclusion and potential future study topics are indistinguishable throughout distant neighborhoods.
presented in Section 6. Xu et al. introduced GINs, which enhanced the
5
expressiveness of GNNs by utilizing a sum-based
2. Literature review aggregation function. This architecture has proven to be
as effective as the test for graph isomorphism. It enables
Predicting molecular properties has been a persistent the effective differentiation of non-isomorphic graphs.
17
challenge in computational chemistry, driven by the The attention mechanism is a concept in machine
necessity to uncover essential molecular characteristics learning, especially in neural networks, that allows a
that affect properties such as carcinogenicity, solubility, model to dynamically focus on the most relevant parts
and toxicity. The prediction of molecular properties of the input data when generating output. Originally
has greatly enhanced due to the progress of machine introduced in natural language processing, it has since
learning methods, especially GNNs. This segment been successfully applied in various domains, including
provides a comprehensive review of the current GNNs, vision, and cheminformatics. GINs have shown
literature regarding molecular property prediction better performance on molecular datasets such as
utilizing GNNs, emphasizing their usage, effectiveness, MUTAG, surpassing GCNs in tasks that need accurate
and architectural advancements. structural representation. In the aggregation stage,
Before the introduction of GNNs, conventional Veličković et al. proposed GATs that assign trainable
machine learning models such as random forests, support weights to local nodes using an attention mechanism. 18
vector machines, and feed-forward neural networks GATs have proven to be especially useful for datasets
were widely utilized for predicting molecular properties. where the interactions among particular nodes hold
These models demanded manually created molecular greater significance than the overall structure. Several
descriptors that encompassed physicochemical studies improved GNNs by explicitly including edge
properties and structural fingerprints (e.g., extended- characteristics that indicate bond types and lengths
connectivity fingerprints). These methods frequently within molecular graphs. Schütt et al. created the SchNet
did not succeed in representing the intricate relational model, which employed continuous-filter convolutional
framework of molecular graphs. For instance, Duvenaud layers for datasets in quantum chemistry. 19
et al. presented a model that directly learned molecular The MUTAG dataset has been extensively utilized
fingerprints from data through deep convolutional to evaluate GNNs for predicting molecular properties.
architectures, highlighting an early transition from MUTAG consists of 188 molecular diagrams depicting
manually crafted features to learnable representations. nitroaromatic substances. Each is categorized as
15
Nevertheless, these approaches did not possess the either carcinogenic or non-carcinogenic. The dataset
capability to fully utilize graph structure, since they were presents a binary classification challenge that assesses
not specifically created for graph-structured information. a model’s capability to accurately capture molecular
In graph-based molecular modeling, molecules graph characteristics. GNNs have likewise been utilized
are abstracted as graphs where atoms are represented on various datasets, including a quantum chemistry

Volume 22 Issue 3 (2025) 90 doi: 10.36922/AJWEP025070041