Abstract
By using omics, we can now examine all components of biological systems
simultaneously. Deep learning-based drug prediction methods have shown
promise by integrating cancer-related multi-omics data. However, the
complex interaction between genes poses challenges in accurately
projecting multi-omics data. In this research, we present a predictive
model for drug response that incorporates diverse types of omics data,
comprising genetic mutation, copy number variation, methylation, and
gene expression data. This study proposes latent alignment for
information mismatch in integration, which is achieved through an
attention module capturing interactions among diverse types of omics
data. The latent alignment and attention modules significantly improve
predictions, outperforming the baseline model, with MSE = 1.1333,
F1-score = 0.5342, and AUROC = 0.5776. High accuracy was achieved in
predicting drug responses for piplartine and tenovin-6, while the
accuracy was comparatively lower for mitomycin-C and obatoclax. The
latent alignment module exclusively outperforms the baseline model,
enhancing the MSE by 0.2375, the F1-score by 4.84%, and the AUROC by
6.1%. Similarly, the attention module only improves these metrics by
0.1899, 2.88%, and 2.84%, respectively. In the interpretability case
study, panobinostat exhibited the most effective predicted response,
with a value of −4.895. We provide reliable insights for drug selection
in personalized medicine by identifying crucial genetic factors
influencing drug response.
Keywords: multi-omics, drug response, latent alignment, deep learning,
attention module
1. Introduction
Cancer continues to be a significant cause of mortality worldwide, with
a high mortality rate that demands urgent attention. The close
association between cancer and gene alterations has been well
established, arising from both intrinsic and extrinsic factors
[[30]1,[31]2]. Targeted therapy has emerged as a potential solution for
cancer treatment, but the genetic heterogeneity of cancer poses
challenges, leading to diverse responses among patients. The
integration of multi-omics data, including genomics, epigenomics, and
transcriptomics, has shown promise in predicting drug responses for
cancer treatment.
Deep learning models have proven effective in extracting essential
features from multi-omics data to predict drug responses
[[32]3,[33]4,[34]5,[35]6]. For instance, Chiu et al. [[36]7] proposed
using different deep neural networks to integrate gene expression and
gene mutation data and then utilizing the integrated features for drug
response prediction. Building on this approach, Sharifi-Noghabi, H. et
al. [[37]8] introduced the multi-omics late integration (MOLI) method,
which further enriches multi-omics data and employs contrastive loss to
enhance the model performance. However, two aspects of these methods
have not been considered. One overlooked aspect is the absence of
consideration of genomics biology, including the central dogma [[38]9].
In molecular biology, the central dogma elucidates how genetic
information moves through a biological system, illustrating the
coordinated interplay of DNA, RNA, and proteins [[39]10]. The other
aspect is their ineffectiveness in proficiently capturing correlations
within inter-omics data, which play a crucial role in accurately
predicting drug response [[40]11].
Intricate interplay, such as that of the genomics flow from DNA to DNA
or DNA to RNA among different omics data sources, must be considered
during integration to improve prediction performance. A challenge
arises during integration, as different omics datasets undergo
dimensionality transformation without considering information from one
another, resulting in feature misalignment [[41]12] in the latent
space. This misalignment hinders the integration results of the model
and poses a critical problem to address.
In addition, to capture the correlation between different types of
omics data, some studies have proposed using an attention mechanism.
For example, Wang et al. [[42]13] proposed a model that integrates
multi-omics data to increase the richness of the input. An attention
layer was introduced to capture the inter-omics correlations and assign
importance weights to different features. Thus, an attention mechanism
allows for more effective information sharing among different omics
features, resulting in a more comprehensive understanding of the data.
To overcome these obstacles, this study aims to investigate the
correlation between genomics-related multi-omics data and cancer.
Various types of multi-omics data, such as mutations, copy number
variations (CNVs), methylation patterns, and gene expressions specific
to cancer cells, will be utilized to predict drug response
concentrations across different cell line samples. To address the issue
of information misalignment during integration, the study proposes a
method called latent alignment, which aligns multi-omics data into a
common latent space after feature extraction. Additionally, an
attention module will be implemented to enhance the modeling of
interactions between different types of omics data. By successfully
addressing the challenges in multi-omics data integration, this
research aims to contribute to the advancement of precision medicine in
cancer therapeutics.
2. Materials and Methods
2.1. Study Design and Workflow
The Cancer Dependency Map (DepMap) datasets [[43]14] include drug
response and diverse cancer-related omics data from cell lines among
various cancer types. In this study, we proposed a drug response
prediction model that integrates mutation, CNV, methylation, and gene
expression data from the DepMap datasets. The model includes data
preprocessing and a feature extractor for essential omics data features
and integrates a latent alignment and attention module ([44]Figure 1).
Multi-omics integration includes calculating the sample similarity to
correct latent space misalignment. To improve the prediction model, an
attention module was introduced after the latent alignment module
([45]Supplementary Figure S1) to capture relationships among various
omics data. A regression loss function was utilized to quantify the
disparity between the predicted and actual values, enhancing the
precision of drug response concentration prediction.
Figure 1.
[46]Figure 1
[47]Open in a new tab
Study design and framework for drug response prediction. We propose a
drug response prediction model that integrates multi-omics data (gene
expression, mutation, CNV, and methylation data). The model involves
data preprocessing and a feature extractor for omics data features and
incorporates latent alignment and attention modules.
We constructed a deep learning model with a dense embedding layer of
size 1024 for all datasets. The attention module had a dimension of 40,
and the output layer dimension was set to 31. To prevent overfitting, a
dropout rate of 0.2 was applied to each feature extractor. During the
training process, we utilized a learning rate of 10–5 for the feature
extractor and 10–6 for the predictor to optimize the model performance.
The training was conducted on an Nvidia GTX 1080-Ti GPU to accelerate
the computations. To assess the model performance and generalizability,
we employed a threefold validation approach. The dataset was divided
into three subsets, and the model was trained and evaluated three
times, each time with a different subset as the validation set.
We conducted extensive training for 200 epochs to allow the model to
converge and capture complex patterns within the data. By utilizing a
substantial number of epochs, we aimed to enhance the model predictive
capacity.
2.2. Datasets and Data Preprocessing
We utilized the DepMap dataset ([48]https://depmap.org/portal/,
accessed on 10 Augest 2022) for experiments. We proposed a drug
response prediction model, illustrated in [49]Figure 1, that integrates
multi-omics information to forecast drug response and uncover
underlying genomic factors. The drug response dataset includes
IC50(log) values, indicating the drug concentration needed to inhibit
50% of cancer cells. First, we identified all the samples present in
the datasets to confirm the availability of gene information from
multiple omics datasets. [50]Supplementary Table S1 shows dataset sizes
before and after preprocessing, with all datasets containing 543
samples post-preprocessing. The number of genes for multi-omics
datasets were 21,840 for CNV, 19,412 for methylation, 19,144 for
expression, and 223 for mutation. The mutation, CNV, and expression
dataset contains the mutation status (with or without mutations), the
copy number (as an integer), and the normalized read count for each
gene and each cell line, respectively. The methylation dataset includes
the methylation status as a number between 0 and 1 for each CpG site
and each cell line. The drug response dataset includes 31 drugs as
prediction targets.
To prevent biased model predictions, we excluded samples with missing
values exceeding 30% and removed genes with all zeros in each dataset.
The remaining missing values were imputed using the mean of the
original gene data. These steps yielded complete datasets, minimizing
bias and ensuring robust analysis while preserving biological
relevance.
2.3. Feature Extractor
Since not every gene in the multi-omics data carries equal importance
for the final prediction, we employed different feature extractors to
capture the essential features from each multi-omics dataset. Equation
(1) outlines the learning process of each feature extractor, where
z^(i) represents the result of feature extraction for the i-th
multi-omics dataset, denoted as x^(i).
[MATH: z(i)=ReLU(W2(
i) (<
mi mathvariant="italic">ReLU(W1(
i) x(i)+b1(
i)))+b2(
i)), i=1, 2, 3, 4, :MATH]
(1)
where
[MATH: W1(
i) :MATH]
∈
[MATH: Rd1×g(i) :MATH]
,
[MATH: W2(
i) :MATH]
∈
[MATH: Rd2×d1 :MATH]
,
[MATH: b1(
i) :MATH]
∈
[MATH: Rd1×bs :MATH]
, and
[MATH: b2(
i) :MATH]
∈
[MATH: Rd2×bs :MATH]
mean the weight and bias of each omics dataset at different layers.
d[1], d[2] is the dimension of the latent vectors,
[MATH: g(i) :MATH]
is the number of genes in the i-th multi-omics dataset, and bs
represents the batch size during training. By leveraging these feature
extraction techniques, we can identify and capture the relevant
information from each multi-omics dataset.
2.4. Latent Alignment
We obtain the sample similarity S[i,j] by calculating the inner product
between different multi-omics features after feature extraction, as
shown in [51]Supplementary Figure S1, enabling us to capture the
differences between each omics dataset in the same sample in the latent
space.
The formula for calculating the sample similarity is shown as follows:
[MATH: Si,j=z(i)<
/mo>T·z(j)<
/mo>, 1≤i<
/mi>≤4, 1≤j≤4. :MATH]
(2)
where i and j represent the i-th and j-th omics dataset, respectively.
Ideally, we expect the highest similarity between samples of different
omics datasets to occur when they represent the same sample. However,
due to the independent nature of the feature extraction process for
each omics dataset, the actual sample similarity between different data
types in the latent space may deviate from the expected value. To
address this disparity, we introduce a learned target T, which
represents the average sample similarity among the same omics data
features. The formulation of the learned target is defined as
[MATH: T=softmax(14·τ·(S1,1+S2,2+S3,3+S4,4)), :MATH]
(3)
where τ is a hyperparameter used to regulate the value range of the
similarity between different samples.
2.5. Attention Module
[52]Supplementary Figure S2 illustrates the architecture of the
attention module. The module incorporates information exchange between
omics datasets through learnable matrices M[i,j] to learn the affinity
matrices F[i,j] for each pair of omics datasets. This is accomplished
using the following formulas:
[MATH: Fi,j=tanh(z(j)T·Mi,j·z(i)),1 ≤ i≤ 4,1 ≤ j≤ 4, i ≠ j, :MATH]
(4)
where M[i,j] ∈
[MATH: Rd2×d2 :MATH]
.
To integrate this relational information, in Equation (5),
[MATH: Hz(i) :MATH]
represents the result of each omics dataset in acquiring information
from other omics datasets.
[MATH: Hz(i)=
tanh(Wz(i)<
/mrow>·z(i)+Σj=1,j≠i4Wz(j)<
/mrow>z(j)·Fi,j),
1≤i≤4, :MATH]
(5)
where
[MATH: Wz(i)<
/mrow> :MATH]
∈
[MATH: Rk×d2 :MATH]
and k is the dimension of the latent vectors.
Then, the attention weight
[MATH: az(i)<
/mrow> :MATH]
of other omics datasets given to the individual omics dataset is
obtained through
[MATH: Hz(i) :MATH]
, as follows:
[MATH: az(i)<
/mrow>=softmax(WHz(i)<
/mrow>·Hz(i))<
/mrow>, 1≤i≤4
, :MATH]
(6)
where
[MATH: WHz(i)<
/mrow> :MATH]
∈
[MATH: R1×
k :MATH]
.
Finally, the weight
[MATH: az(i)<
/mrow> :MATH]
is multiplied by the original feature vector z^(i) and concatenated to
obtain the features with the correlation information of different omics
datasets,
[MATH: z(i)′=
az(i)·z(i), 1≤i≤4, :MATH]
(7)
[MATH: Z′=concat(z(1)′,
z(2)′,
z(3)′,
z(4)′).<
/mo> :MATH]
(8)
Next, the integrated features are used to predict drug response, and
the prediction can be written as:
[MATH: Pred=W5(ReLU(W4(ReLU(W3·Z′+b3))+b4))+b5,
:MATH]
(9)
where W[3] ∈
[MATH: Rd3×(
4·d2) :MATH]
, W[4] ∈
[MATH: Rd4×d3 :MATH]
, W[5] ∈
[MATH: RD×d4 :MATH]
, b[3] ∈
[MATH: Rd3×bs :MATH]
, b[4] ∈
[MATH: Rd4×bs :MATH]
, b[5] ∈
[MATH: RD×
bs :MATH]
, d[3], d[4] is the dimension of the latent vectors, and D represents
the number of drugs to be predicted.
2.6. Loss Function
The objective function is
[MATH: Lreg=
1N∑i=
1Nyi−y^<
mi>i, :MATH]
(10)
where N is the number of training samples,
[MATH: yi :MATH]
represents the actual drug response concentration, and
[MATH: y^<
mi>i :MATH]
is the predicted drug response concentration.
In the latent alignment module, we updated the feature extractors of
different omics datasets by calculating the similarity between
different features in the latent space and the difference from the
target T. Our goal was to minimize the difference between the same
samples.
[MATH: Lla=Σi=13(Σj=i+14<
mrow>(12·(CE(Si,j,T)+CE(Sj,i,T)))), :MATH]
(11)
where CE( ) is the cross-entropy function.
The objective function of the model, which incorporates these
integration optimization strategies for drug response prediction, can
be expressed as follows:
[MATH: Ltot<
mi>al=γr×Lreg<
/mrow>+γ
l×Lla, :MATH]
(12)
where γ[r] and γ[l] are the ratios for these two losses. Here, we set
the ratio to 1:1.
2.7. Evaluation Metrics
The mean squared error (MSE) is a metric used to quantify the average
squared difference between predicted values and actual values. In
addition, to assess the effectiveness of individual drugs on patients
in the drug concentration dataset, we followed the approach described
by Emdadi et al. [[53]15]. Specifically, the F1-score and AUROC were
employed as metrics to evaluate the model performance based on the
criterion of the median of individual drug response concentrations in
the training dataset.
3. Results
3.1. The Performance of Our Method for Predicting Drug Response
We compared the performance of our model with that of other models
[[54]8,[55]13] that also utilized multi-omics data for predicting drug
response. To facilitate an unbiased comparison, we utilized consistent
datasets and preprocessing methods while implementing the model
architectures as detailed in the referenced studies. [56]Table 1
presents the performance of different model architectures in predicting
drug response concentrations using multi-omics data.
Table 1.
Comparison of our model and reference methods for predicting drug
response.
Method MSE F1-Score AUROC
Sharifi-Noghabi H. et al. [[57]8] 1.3556 ± 0.0062 0.5049 ± 0.0208
0.5151 ± 0.0037
Wang C et al. [[58]10] 1.2455 ± 0.0101 0.5115 ± 0.0262 0.5277 ± 0.0078
Our model 1.1333 ± 0.0057 0.5342 ± 0.0162 0.5776 ± 0.0074
[59]Open in a new tab
MSE: mean squared error; AUROC: area under the receiver operating
characteristic curve.
[60]Table 1 presents the mean and standard deviation values obtained
from 3-fold cross-validation. Our model achieved superior prediction
performance, yielding the lowest mean squared error (MSE) of 1.1333.
Although the MSE difference of 0.1122 between our model and the
second-best performing model proposed by Wang, C. et al. [[61]10]
(1.2455) may seem slight, notable discrepancies emerge in other
metrics. Our model outperforms that of Wang et al. [[62]10], with a
2.27% higher F1-score and a 4.99% improvement in the area under the
receiver operating characteristic curve (AUROC).
[63]Supplementary Table S2 presents the drugs ranked by MSE based on
the predicted drug response concentrations using our model. The table
illustrates MSE variations among different drugs, indicating
disparities in the prediction accuracy. Our model exhibits high
prediction accuracy for drugs such as piplartine and tenovin-6 but
comparatively lower accuracy for mitomycin-C and obatoclax.
3.2. Results of Different Combinations of Multi-Omics Datasets
In this experiment, our aim was to investigate how various combinations
of omics datasets impact the accuracy of predicting drug response
concentrations. We employed a consistent model architecture, varying
the input by integrating subsets of omics data, and the results of drug
response concentration predictions are presented in [64]Table 2. Among
various omics dataset combinations, it is clear that methylation and
expression datasets are pivotal in predicting drug response
concentrations. The model performance is significantly improved by
including these two types of omics data. When employing the CNV and
mutation datasets in the experiments, the MSE, F1-score, and AUROC
values are 1.3209, 51.11%, and 52.21%, respectively. In contrast, by
utilizing methylation and expression datasets, the values improved by
1.2090, 51.04%, and 56.64%, respectively. Despite a marginal 0.07%
decrease in the F1-score, there are notable individual enhancements of
0.1119 in the MSE and 4.43% in the AUROC. This result indicates that
the methylation and expression datasets are highly relevant in
predicting drug response concentrations.
Table 2.
Performance differences among different combinations of omics datasets.
Combination MSE F1-Score AUROC
Me–Ex–CNV–Mu 1.1333 ± 0.0057 0.5342 ± 0.0162 0.5776 ± 0.0074
Me–Ex–Mu 1.1592 ± 0.0041 0.5260 ± 0.0157 0.5632 ± 0.0077
Me–Ex–CNV 1.1960 ± 0.0203 0.4892 ± 0.0069 0.5539 ± 0.0108
Me–CNV–Mu 1.2170 ± 0.0135 0.4990 ± 0.0147 0.5535 ± 0.0056
Ex–CNV–Mu 1.2040 ± 0.0145 0.5162 ± 0.0280 0.5584 ± 0.0075
Me–Ex 1.2090 ± 0.0063 0.5104 ± 0.0183 0.5664 ± 0.0078
Me–CNV 1.3116 ± 0.0151 0.4733 ± 0.0159 0.5249 ± 0.0019
Me–Mu 1.2242 ± 0.0036 0.4935 ± 0.0153 0.5660 ± 0.0086
Ex–CNV 1.3006 ± 0.0047 0.4788 ± 0.0190 0.5103 ± 0.0107
Ex–Mu 1.2223 ± 0.0122 0.5123 ± 0.0276 0.5683 ± 0.0067
CNV–Mu 1.3209 ± 0.0059 0.5111 ± 0.0335 0.5221 ± 0.0074
Me 1.3660 ± 0.0080 0.4968 ± 0.0134 0.5177 ± 0.0098
EX 1.3397 ± 0.0016 0.4856 ± 0.0321 0.5120 ± 0.0056
CNV 1.3828 ± 0.0076 0.5012 ± 0.0118 0.5067 ± 0.0051
Mu 1.4201 ± 0.0257 0.5091 ± 0.0172 0.5040 ± 0.0031
[65]Open in a new tab
To assess the impact of various omics datasets on 31 drugs, we
established the initial prediction results as the baseline.
Subsequently, we applied a mask, setting the input of one type of omics
dataset to zero, and reperformed drug response concentration prediction
with the modified inputs. Based on the new prediction results, we
calculated the differences between each drug response and the baseline
result to evaluate the impact of excluding a specific type of omics
data on the prediction outcomes. If the new prediction result was worse
than the baseline, it indicated a positive impact of the omitted omics
data on the model predictions. Conversely, if the new prediction result
was better than the baseline, it indicated a negative impact. Using
this approach, we generated [66]Figure 2, which summarizes the obtained
results.
Figure 2.
[67]Figure 2
[68]Open in a new tab
Drug response prediction using different combinations of omics
datasets. The x-axis represents the index of drugs in the dataset. The
y-axis represents different omics datasets, and the darkness of the
color indicates the impact of the respective omics dataset on the drug
response prediction. Darker colors indicate a positive influence, while
lighter colors indicate a negative influence on the model predictions.
The methylation and gene expression datasets make a meaningful
contribution to predicting drug response.
In [69]Figure 2, the x-axis represents the index of 31 drugs in the
dataset, corresponding to the information provided in [70]Supplementary
Table S3. From the chart, it can be observed that the performance of
the methylation and expression datasets in predicting drug response
concentrations is mostly positive, except for a few drugs such as
KIN001-204 and obatoclax, where a negative impact is observed. On the
other hand, for the mutation and CNV datasets, a negative influence is
found. This observation aligns with the notion mentioned in our
previous experiment that the methylation and expression datasets make a
meaningful contribution to predicting drug response.
3.3. Ablation Experiments Focus on the Latent Alignment and Attention Modules
In ablation experiments, our objective was to assess the impact of
integrating the latent alignment and attention modules on the
prediction of drug response concentrations. [71]Table 3 presents a
comparison of the effects of integrating the individual modules on the
prediction of drug response concentrations. The results clearly
demonstrate that both the latent alignment and attention modules
contribute to improved predictions. The latent alignment and attention
modules significantly improve predictions, outperforming the baseline
model, with MSE = 1.1333, F1-score = 0.5342, and AUROC = 0.5776. The
latent alignment module only outperforms the baseline model in terms of
the MSE, F1-score, and AUROC by 0.2375, 4.84%, and 6.1%, respectively.
Similarly, adding the attention module yields improvements in these
metrics, with differences of 0.1899, 2.88%, and 2.84%, respectively.
Table 3.
The overlapping genes in chromatin organization.
H2AX KAT5 KAT14 ARID1A
ATF2 RBBP4 PRMT6 ELP6
H2AW RUVBL2 SUZ12 SMARCA2
KMT2D NSD1 NCOA2 GATAD2A
KDM5D SAP130 SMARCC1 SMARCA4
KMT2A EPC1 KDM2B HCFC1
EHMT2 SAP18 PRMT1 TADA2A
EHMT1 SMYD3 EED KANSL1
YEATS4 SUPT20H MCRS1 KAT6B
ART3 JAK2 PRMT3 H2AC20
CHD4 MTA2 SETDB2 TADA3
SGF29 MTA3 ELP3 BRWD3
[72]Open in a new tab
After integrating the latent alignment and attention modules into our
model, we observed significant improvements in its ability to learn and
integrate information from diverse omics data sources. The improved
performance can be attributed to our model’s ability to discern the
genes that truly influence drug response. This discernment is achieved
through the utilization of optimizing integration strategies, enabling
our model to learn the interrelationships among diverse omics datasets
and the genetic information they contain. As a result, this acquired
knowledge is seamlessly incorporated into the process of predicting
drug responses.
3.4. Case Study on the Interpretability of the Model
After integrating latent alignment and attention modules into our
model, we observed significant enhancements in its capacity to
effectively gather and integrate information from various omics data
sources. These advancements prompted us to explore the improved
accuracy of our model in predicting drug response concentrations. The
enhanced performance is primarily attributed to our model’s capability
to identify the pivotal genes influencing drug responses. This ability
stems from optimizing integration strategies, allowing our model to
comprehend the complex relationships among diverse omics datasets and
their genetic insights. Consequently, this acquired knowledge
seamlessly contributes to the prediction of drug responses. We
demonstrate our method in the following two experiments.
3.4.1. Reactome Results Obtained by Integrating the Latent Alignment and
Attention Modules
To determine whether the selected genes are truly associated with drug
responses, first, we randomly selected a sample from the test dataset,
which was a cell line sample of acute myeloid leukemia. Using our
model, we predicted the response concentrations of 31 drugs for this
sample and sorted the results. We found that the drug with the most
effective predicted response was panobinostat, with a predicted value
of −4.895, while the actual value was −5.600. The difference between
these values was only 0.795. Finally, we identified the top 3% of genes
in the model prediction of the patient’s response to panobinostat based
on their contribution among all genes.
The Reactome database is a peer-reviewed biological pathway database
that provides comprehensive cellular processes [[73]16]. [74]Figure 3A
illustrates the enriched signaling pathways obtained through pathway
enrichment analysis using the identified top 3% important genes and the
Reactome database for the specific pathway related to panobinostat,
namely, the chromatin organization signaling pathway. The results of
these 48 overlapping genes, demonstrating a significant −log(p value)
of 3.29, can be observed in [75]Supplementary Table S4. Additionally,
to visualize the location and relationship of drug action within the
signaling pathway associated with panobinostat, relevant information
obtained from the Reactome dataset is displayed in [76]Figure 3B. The
drug enters through the HDAC2 inhibitor and affects the region
indicated by the red line segment.
Figure 3.
[77]Figure 3
[78]Open in a new tab
(A) Enriched signaling pathways and overlapping genes. The x-axis
represents the significance of gene expression in the signaling
pathways, where higher values indicate more significant expression of
the important genes within the pathways. The y-axis represents the
signaling pathways derived from the analysis of these important genes.
The orange color represents the signaling pathways associated with
panobinostat. The size of the circles corresponds to the overlap of
important genes in the signaling pathways, with larger circles
indicating a greater degree of overlap. (B) Chromatin organization
signaling pathway and drug action site. By using Reactome, we found a
pathway–gene network of chromatin organization signaling pathways and
drug action sites. The upper right corner depicts the chromatin
organization signaling pathway, while the left image illustrates the
drug’s response location within this pathway.
3.4.2. Reactome Results by the Base Model
To validate the effectiveness of the latent alignment and attention
modules incorporated into our model in improving the confidence of
predictions, we conducted experiments using the base model on the same
cell line sample afflicted with acute myeloid leukemia. Similarly, we
predicted the drug concentrations of 31 drugs in response to this
sample and ranked them accordingly. Panobinostat remained the most
effective drug in terms of the predicted results, with a value of
−4.110, deviating by 0.785 compared to the value of our model. In
[79]Supplementary Figure S3, it is evident that the ‘chromatin
organization’ signaling pathway associated with panobinostat overlaps
with 12 genes, as shown in [80]Supplementary Table S5, and demonstrates
a significantly low log(p value) of 1.19 × 10^−7.
4. Discussion
In this study, we proposed a novel framework for accurately predicting
drug response concentrations by effectively integrating multi-omics
data. Our datasets comprise genomic-level information, including
mutation and CNV data, as well as transcriptomic-level data
encompassing methylation patterns and gene expression profiles. By
integrating these diverse data types, our framework provides an
interpretable prediction model with enhanced performance. Our results
highlight the following important points. (i) Our approach outperforms
two distinct multi-omics models. (ii) The latent alignment and
attention modules significantly improve predictions, outperforming the
baseline model, with MSE = 1.1333, F1-score = 0.5342, and AUROC =
0.5776. (iii) The case study demonstrates the most effective predicted
response with clinical relevance.
Numerous machine learning models exist for the integration of
multi-omics data [[81]17,[82]18]. Our approach specifically aligns
omics features and captures correlation information among diverse omics
datasets, thereby augmenting the overall performance of the model. To
address feature misalignment, a latent alignment module was introduced
to align different omics features in the same latent space after
feature extraction, improving the integration results. An attention
module was also incorporated to leverage information between omics data
sources, capturing correlations and providing additional insights to
enhance the model’s predictive capability. The performance of our model
surpasses that of the reference study models [[83]8,[84]13], providing
compelling evidence of its capability in drug response concentration
prediction.
Moreover, ablation experiments confirm the significant impact of two
key components: the latent alignment and attention modules. The latent
alignment module facilitates the matching of multi-omics features
during the training process, while the attention module captures the
relationships among different omics datasets. Both modules
significantly contribute to enhancing the performance of drug response
prediction. By integrating both the attention and latent alignment
modules, our final model achieves superior performance in predicting
drug response concentrations, acquiring additional information, and
improving the integration approach.
It is common to encounter missing values in omics data, originating
from various factors [[85]19]. To ensure robust predictions, we adopted
a strategy that encompasses data preprocessing and leverages deep
learning techniques for both feature extraction and prediction. By
utilizing the learned target and considering the different omics
datasets during feature extraction, we improved the similarity of
samples across different data types, thereby enhancing the integration
performance. The results clearly demonstrate that the improvement in
the model performance is not solely attributed to including a larger
amount of available information but also to the advantages of our
integration approach.
To predict drug response concentrations, considering that the impact of
various omics datasets on the prediction results may vary, we conducted
an additional experiment. However, it is important to note that the
negative impact of the mutation and CNV datasets does not necessarily
imply that excluding these datasets would lead to better predictions of
drug response. However, it does indicate that in our trained model,
these two datasets contribute relatively less information and have a
lesser impact on the final prediction results.
This study focused on predicting drug response concentrations
[[86]20,[87]21], providing insights into both the effectiveness of
patient drug responses and the necessary drug concentration for optimal
inhibitory effects. In the case study, importantly, panobinostat is
indeed the most effective drug in the actual response of the patient to
the 31 drugs. Therefore, this prediction result is reasonable.
We observe consistent improvements in both the performance and
interpretability of the model across various cancer types and their
corresponding drug response predictions. The ability of our model to
leverage genes directly associated with drug response further enhances
the reliability of its predictions, even in the presence of diverse
drug effects on patients. Overall, the experimental outcomes
demonstrate the efficacy of our model architecture, particularly the
attention and latent alignment modules, in predicting drug response
concentrations. The successful integration of multi-omics data and
improved interpretability highlight the potential of our model for
enhancing drug response prediction in clinical settings.
5. Conclusions
In this study, we introduced the “Optimizing Integration Strategies”
framework, designed to predict drug response concentrations accurately
by integrating multi-omics data effectively. This framework
incorporates genomic data such as mutations and CNV, alongside
transcriptomic data including methylation patterns and gene expression
levels. By combining these diverse data types, our framework offers an
interpretable prediction model with enhanced performance.
Experimental results demonstrated the superiority of our “Optimizing
Integration Strategies” framework compared to previous approaches for
the same task. Further, ablation experiments confirmed the significant
impact of two critical components: the latent alignment module, which
aligns multi-omics features during training, and the attention module,
which captures relationships among different omics data types. Both
modules contributed to improved drug response prediction performance.
These findings underscore the substantial advancements achieved by our
proposed framework in predictive accuracy and its ability to leverage
diverse omics datasets effectively.
Through comprehensive experimental analysis, we observed that
methylation and gene expression datasets positively influenced drug
response prediction, highlighting their significant contributions to
our model. In contrast, mutation and CNV datasets had relatively
smaller effects on the final prediction results, allowing for a more
effective prioritization of relevant omics data for drug response
prediction tasks.
Additionally, our investigation into the correlation between selected
genes and actual drug response validated the reliability of our model’s
predictions. Pathway enrichment analysis indicated that these
identified important genes are directly involved in pathways relevant
to the corresponding drugs.
By successfully integrating multiple omics data types, our model not
only improves prediction accuracy but also enhances interpretability,
facilitating the identification of key genetic factors influencing drug
response. These results provide reliable insights for drug selection in
medical decision-making and establish a robust framework for
integrating diverse types of data.
Supplementary Materials
The following supporting information can be downloaded at:
[88]https://www.mdpi.com/article/10.3390/jpm14070694/s1, Figure S1: The
latent alignment module; Figure S2: Architecture of the attention
module; Figure S3: Pathway analysis_based models; Table S1: The dataset
sizes before and after preprocessing; Table S2: Individual drug
response prediction results; Table S3: Drug-Index correspondence; Table
S4: The overlapping genes in Chromatin Organization; Table S5: The
overlapping genes in Chromatin Organization (Base model).
[89]jpm-14-00694-s001.zip^ (1MB, zip)
Author Contributions
Research design: J.-H.C. and P.-C.L.; development of methodology:
Y.-C.C. and H.-O.C.; acquisition of data: Y.-C.C., H.-O.C. and P.-C.L.;
statistical and computational analysis: Y.-C.C., H.-O.C. and P.-C.L.;
writing, review, and revision of the manuscript: Y.-C.C., H.-O.C.,
P.-C.L. and J.-H.C.; Study supervision: J.-H.C. All authors have read
and agreed to the published version of the manuscript.
Institutional Review Board Statement
Not applicable.
Informed Consent Statement
Not applicable.
Data Availability Statement
The datasets used and analyzed during the current study are available
from the corresponding author on reasonable request. The source code is
available at the GitHub repository
([90]https://github.com/Chei-YuanChi/Matster_Thesis). The raw data from
the DepMap dataset (accessed on 10 August 2022) can be downloaded at
[91]https://drive.google.com/open?id=10O4lwyLxg5nLx7rB4i_--moxM6p2Gy1K&
authuser=dpc0628%40gmail.com&usp=drive_fs (accessed on 10 August 2022).
Conflicts of Interest
The authors declare no conflicts of interest.
Funding Statement
This work was funded in part by the National Science and Technology
Council (NSTC), Taiwan, under research grant numbers NSTC
111-2634-F-006-011, NSTC 112-2314-B-006-085, and NSTC
112-2622-8-006-019-IE, and the National Cheng Kung University (grant
numbers NCKUH-11202042, NCKUH-11210015 and NCKUH-11209008). This
research was funded in part by the Higher Education Sprout Project,
Ministry of Education to the Headquarters of University Advancement at
National Cheng Kung University (NCKU). All authors have read and
approved the manuscript.
Footnotes
Disclaimer/Publisher’s Note: The statements, opinions and data
contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI
and/or the editor(s) disclaim responsibility for any injury to people
or property resulting from any ideas, methods, instructions or products
referred to in the content.
References