Right now, doctors often use the same treatment for everyone with the same type of cancer.
But this doesn't always work well because every person's cancer is different, and people react
differently to the same medicine.
My motivation for this research comes from three key issues:
• First, minimizing harm by avoiding ineffective or toxic treatments.
• Second, improving prediction of drug effectiveness, which currently varies widely
between patients.
• Third, overcoming the limitations of existing models, which often rely on single-omics
data or simple concatenation, resulting in weaker predictive power.
Problem statement
Accurate prediction of cancer drug response remains a major challenge.
Existing deep learning models often demand high computational resources, long training times,
and substantial memory, which makes them impractical for large datasets.
Although vast amounts of patient data are now available, current models still struggle to
effectively leverage complex multi-omics data.
My research directly addresses this gap by proposing a time-efficient and memory-optimized
model capable of handling large-scale data.”
Objective
My objectives are:
• To develop a time-efficient and memory-optimized deep learning model for accurate
cancer drug response prediction.
• To integrate multi-omics data and drug features into a sparse, computationally efficient
drug-cell-pair network while solving the over-smoothing problem common in graph-
based learning.
My key contributions include:
• Designing an optimized architecture that significantly reduces computation time and
memory usage while maintaining or improving predictive performance.
• Demonstrating superior scalability and practical applicability by achieving faster training
and inference across large drug–cell line datasets compared to existing deep learning
methods.”
Methodology:
Data Collection:
Drug information was sourced from PubChem, while multi-omics data—including RNA-seq,
DNA methylation, microRNA expression, and copy number variation—came from the Cancer
Cell Line Encyclopedia (CCLE). Drug response data were obtained from the GDSC portal.
Once the raw data were collected, we performed initial preprocessing to prepare it for
modeling.
First, we mapped and integrated the datasets because each source uses different annotation
schemes.
For example, RNA-seq data contained about 400 cell lines, while CNV data included around 450.
To ensure consistency, we selected only those cell lines that were common across all datasets.
We then handled missing values using imputation, removed outliers to improve quality, and
applied normalization to bring all features to a comparable scale.
Cell Similarity – Slide
We started with cell similarity because identifying how similar cell lines behave is crucial for
predicting their drug response.
To comput Cell Similarity, we integrate two data types—miRNA and CpG methylation. After
applying PCA, we reduce them to 200 and 1000 components, respectively. To balance their
contribution, we assign weights: 0.4 for miRNA and 0.6 for CpG.
We used dynamic neighbout selection between 10 to 20 per cell and calculate the cell similarity
using cosine similariy.
After then
Physicochemical features were reduced to 150 dimensions and cosine similarity with adaptive
thresholding built a balanced drug similarity network.
------------------------
After computing cell and drug similarities, we integrate them into a Drug–Cell Pair network,
which is Bipartite graph between drug and cell lines, then it transforming to homogeneous
graph of 561,824 nodes. covering all possible combinations.
For each DCP node, we combine drug neighbors and cell neighbors, and compute a weighted
similarity. We then select the top 10 most similar DCPs per node. And construct the full DCP
network.
So, we already constructed the DCP node. Now we have to find the feature of each DCP node .
RNAseq Expression data and Copy number variation data are highly responsible for drug
response prediction. But these data is very high dimensional.
So we used to autoencoder to reduce the dimension .
Because high-dimensional omics data pose computational challenges, I applied a symmetric
encoder–decoder architecture with GELU activation, skip connections, and multi-head self-
attention to compress RNA-seq and CNV data into 400 informative features.
The reduced features were then fused to form a unified representation.
Finally, the fused data were input into my NeuroGraphDRP model,
First we concatenate the features from cell features and Drug features through the attention
mechanism , then we combine the features with DCP node. Then we find the DCP features , it is
the actual input of my model.
This DCP features then fed in to the iterative layer GCN layers that learn Complex interaction
between drug and cells. After then it fed into Quantum Inspired Processing block including
entanglement , Rotation, measurement . it helps capture more complicated pattern in the data
by treating them quantum states and encoding them. Now it marge all learned features into a
single representation , then it fed in to fc layer which predict the IC50.
Then the result section
which applies quantum-inspired graph layers to efficiently learn complex patterns.”