ArticlePDF Available

DeepQA: A Unified Transcriptome‐Based Aging Clock Using Deep Neural Networks

Wiley
Aging Cell
Authors:

Abstract and Figures

Understanding the complex biological process of aging is of great value, especially as it can help develop therapeutics to prolong healthy life. Predicting biological age from gene expression data has shown to be an effective means to quantify aging of a subject, and to identify molecular and cellular biomarkers of aging. A typical approach for estimating biological age, adopted by almost all existing aging clocks, is to train machine learning models only on healthy subjects, but to infer on both healthy and unhealthy subjects. However, the inherent bias in this approach results in inaccurate biological age as shown in this study. Moreover, almost all existing transcriptome‐based aging clocks were built around an inefficient procedure of gene selection followed by conventional machine learning models such as elastic nets, linear discriminant analysis etc. To address these limitations, we proposed DeepQA, a unified aging clock based on mixture of experts. Unlike existing methods, DeepQA is equipped with a specially designed Hinge‐Mean‐Absolute‐Error (Hinge‐MAE) loss so that it can train on both healthy and unhealthy subjects of multiple cohorts to reduce the bias of inferring biological age of unhealthy subjects. Our experiments showed that DeepQA significantly outperformed existing methods for biological age estimation on both healthy and unhealthy subjects. In addition, our method avoids the inefficient exhaustive search of genes, and provides a novel means to identify genes activated in aging prediction, alternative to such as differential gene expression analysis.
This content is subject to copyright. Terms and conditions apply.
1 of 18
Aging Cell, 2025; 0:e14471
https://doi.org/10.1111/acel.14471
Aging Cell
RESEARCH ARTICLE OPEN ACCESS
DeepQA: A Unified Transcriptome- Based Aging Clock
Using Deep Neural Networks
HongqianQi1,2 | HongchenZhao3 | EnyiLi3 | XinyiLu1 | NingboYu3,4 | JinchaoLiu3,4 | JiandaHan3,4
1State Key Laboratory of Medicinal Chemical Biology, Nankai University, Tianjin, China | 2College of Pharmacy, Nankai University, Tianjin,
China | 3College of Artificial Intelligence, Nankai University, Tianjin, China | 4Engineering Research Center of Trusted Behav ior Intelligence, Ministry of
Education, Nankai University, China
Correspondence: Ningbo Yu (nyu@nankai.edu.cn) | Jinchao Liu (liujinchao@nankai.edu.cn) | Jianda Han (hanjianda@nankai.edu.cn)
Received: 11 August 202 4 | Revised: 21 November 2024 | Accepted: 17 December 2024
Funding: This work was supported by National Key Research and Development Program of China (2022YFA1103800, 2022YFB4703204). National Natural
Science Foundation of China (62076140, U1913208). Natural Science Foundation of Tianjin (21JCQNJC00 010).
ABSTRACT
Understanding the complex biological process of aging is of great value, especially as it can help develop therapeutics to prolong
healthy life. Predicting biological age from gene expression data has shown to be an effective means to quantify aging of a subject,
and to identify molecular and cellular biomarkers of aging. A typical approach for estimating biological age, adopted by almost
all existing aging clocks, is to train machine learning models only on healthy subjects, but to infer on both healthy and unhealthy
subjects. However, the inherent bias in this approach results in inaccurate biological age as shown in this study. Moreover, almost
all existing transcriptome- based aging clocks were built around an inefficient procedure of gene selection followed by conven-
tional machine learning models such as elastic nets, linear discriminant analysis etc. To address these limitations, we proposed
DeepQA, a unified aging clock based on mixture of experts. Unlike existing methods, DeepQA is equipped with a specially
designed Hinge- Mean- Absolute- Error (Hinge- MAE) loss so that it can train on both healthy and unhealthy subjects of multiple
cohorts to reduce the bias of inferring biological age of unhealthy subjects. Our experiments showed that DeepQA significantly
outperformed existing methods for biological age estimation on both healthy and unhealthy subjects. In addition, our method
avoids the inefficient exhaustive search of genes, and provides a novel means to identify genes activated in aging prediction, al-
ternative to such as differential gene expression analysis.
1 | Introduction
Biological aging is a very complex process and driven by/related
to many cellular and biological processes. Many efforts have been
made to understand biological aging and to tackle aging- related
diseases. In recent years, thanks to high- throughput sequencing
techniques, large- scale omics data are now available and provide
new opportunities for gaining a deeper understanding of biolog-
ical aging.
Predicting biological age of a subject using machine learn-
ing methods, colloquially termed aging clocks, is an import-
ant means/task to understand the complex biological process
of aging and to facilitate clinical decision- making. A widely
This is a n open access ar ticle under the terms of t he Creative Commons Attr ibution License, which p ermits use, dis tribution and repro duction in any medium, p rovided the orig inal work is
properly cited.
© 2025 T he Author(s). Aging Cell publi shed by Anatomic al Society and Joh n Wiley & Sons Ltd.
Hongqian Qi a nd Hongchen Zhao contr ibuted equally to t his work.
2 of 18 Aging Cell, 2025
accepted hypothesis is that the estimated age of aging clocks
may reflect the aging rate of an individual to some extent, thus
can serve as a measure of biological age (Rutledge, Oh, and
Wyss- Coray2022).
Built upon different types of omics data, aging clocks can be cate-
gorized into different categories, such as DNA methylation aging
clocks (Bocklandt etal.2011; Horvath2013; Hannum etal.2013;
Zhang etal.2017), proteomic ag ing clocks (Pappireddi, Mar tin, and
Wühr2019; Suhre, McCarthy, and Schwenk2021), and transcrip-
tomic aging clocks (López- Otín etal. 2023; Fleischer etal. 2018;
Shokhirev and Johnson2021; Meyer and Schumacher2021). For
instance, Hannum etal. proposed an elastic net- based DNA meth-
ylation aging clock trained on methylomic and clinical parame-
ters, such as gender and body mass index (BMI) and achieved a
validation error of 4.9 years (root- mean- squared- error, RMSE)
(Hannum etal.2013). Zhang etal. developed a second generation
epigenetic clocks where a mortality risk score based on 10 selected
CpGs exhibits strong association with all- cause mortality (Zhang
etal.2017). Though that DNA methylation aging clocks have been
found to be highly reproducible, the resulting models are gener-
ally very difficult to be interpreted to understand the molecular
and cellular causes and consequences of genomic CpG methyla-
tion (Rutledge, Oh, and Wyss- Coray2022). In contrast, transcrip-
tomic aging clocks try to directly link aging to genes (RNA gene
expression levels) and become a promising solution alternative to
other types of aging clocks.
In this study, we focus on the development of transcriptomic
aging clocks. Prior work of this line of research are such as
(López- Otín et al. 2023; Fleischer et al. 2018; Shokhirev and
Johnson 2021; Meyer and Schumacher 2021). Among them,
Fleischer etal. uses an ensemble of linear discriminant analy-
sis as age predictor and the genes for the model training were
selected with explicit filtering process and obtained a predic-
tion error of 7.7 years (The prediction accuracies of aging clocks
in different references may not be directly comparable due to
different datasets and evaluation protocols.) (mean- absolute-
error, MAE) (Fleischer et al. 2018). Shokhirev et al. proposes
to use standard differential expression analyses to select top
1000 differential/variable genes and train random forests on
the top of these genes to predict ages with a prediction accu-
racy of 3.22 years (RMSE) (Shokhirev and Johnson2021). Meyer
etal. employed an expensive gene selection along with elastic
net as age predictor and achieved a prediction error of 6.63 years
(MAE) (Meyer and Schumacher2021).
Though some of these methods showed promising performance
of age prediction, they suffered from several major drawbacks:
Firstly, almost all existing aging clocks follow the protocol of
training machine learning models only on healthy subjects to
predict their chronological age, but to infer the biological age
on both healthy and unhealthy subjects. This protocol seems
to be reasonable since we do not have the accurate biological
age of unhealthy subjects. However, as the model is only trained
on healthy subjects, the deviation of predicted age from the
chronological age of an unhealthy subject in the test set could be
caused by abnormal aging process of the individual and/or the
incompetence of the machine learning model dealing with the
domain gap as the model is not trained on unhealthy subjects.
As a result, the estimated biological age could be inaccurate.
Addressing this problem is challenging and requires novel train-
ing protocol for unhealthy subjects. In this paper, we propose a
simple yet effective training loss, called hinge- mean- absolute-
error loss, which allows unhealthy subjects participate in train-
ing of the aging clock.
Secondly, the prediction accuracy of these methods is still
limited, and many involve iterative gene selection prior to
training a prediction model which demands heavy computa-
tional resources. For large scale datasets, these approaches of
gene selection could be prohibitively expensive to carry out.
Moreover, we observed that some of existing work of aging
clocks failed to avoid the selection bias in gene selection pro-
cess (Ambroise and McLachlan 2002). In other words, they
did not follow the standard testing protocol which is widely
accepted to secure a valid evaluation of generalization ability
of methods. Consequently, the prediction performance can be
seriously overestimated.
One potential solution to address these limitations is deep learn-
ing which has shown to excel in learning complex representa-
tions from a broad spectrum of tasks involving different kinds of
high- dimensional data w ith proper design and tra ining strategies
(Noothout etal.2022; Krizhevsky, Sutskever, and Hinton2012;
Salehinejad et al. 2018; Zhang et al. 2020; Liu et al. 2021; Wu
et al. 2024). Especially, deep learning has found many suc-
cessful applications in biology and healthcare etc. (Korsunsky
etal. 2019; Du etal.2019; Eraslan etal.2019; Zhao etal.2021;
Ceglia etal.2023). Recently Mohamadi etal. presents an attempt
to apply deep neural networks to human age estimation from
gene expression data without gene selection as a prior step. In
their approach, the input features (genes) were reshaped to 2D
image- alike signals so that they can be processed by convolu-
tional layers. However, unlike images where neighboring pixels
are indeed spatially close to one another, genes, and their neigh-
bors in these image- alike 2D signals are not necessarily related.
This may explain why the performance of this method, as shown
in the experimental section, was not entirely satisfactory. This
motivates us to explore the feasibility of using deep learning for
aging estimation on transcriptomic data.
Moreover, deep learning has shown great potential as prediction
models in aging estimation involving various data types (Bao
etal.2023), especially images of X- rays or MRI etc. For instance,
lens photographs and deep learning were used to predict bio-
logical age of humans and access the risks of age- related eye
and systemic diseases (Ma et al.2021; Li etal.2023). Another
example is brain aging estimation which could involve three
categories of biomarkers, namely functional, imaging, and body
fluid ones. For imaging itself, it includes more than t wo different
kinds of data: functional magnetic resonance imaging (fMRI),
optical coherence tomography (OCT) etc. (Mishra, Beheshti,
and Khanna2023; Aging Biomarker Consortium etal.2023).
Owing to the power of deep models, even more complicated and
unconventional types of data can be used as biomarkers to pre-
dict the degree of aging. In particular, Savcisens etal. proposed a
transformer- based deep model life2vec which takes information
about life- events related to health, education, occupation, in-
come, address and working hours, and predict diverse outcomes
3 of 18
ranging from early mortality to personality nuances (Savcisens
etal.2023). This method has shown to significantly outperform
state- of- the- art methods.
In this paper, we proposed a transcriptome- based aging clock
for biological age estimation, named as “DeepQA” which
stands for Deep learning for Quantifying Aging. DeepQA is
equipped with a novel loss function, named hinge- mean-
absolute- error (Hinge- MAE) loss, so that it can be trained
on unhealthy subjects and estimate biological age more ac-
curately. To the best of our knowledge, this is the first aging
clock which trains on both healthy and unhealthy subjects
from multiple cohorts. Inside DeepQA, a mixture of expert
model (MoE) is employed to model the mapping from the gene
expression data to biological ages. DeepQA takes expression
data of a full set of genes as inputs and does not need any prior
gene selection. Instead, the prominent genes which may be
related to aging process are identified through analyzing the
saliency map of the trained deep model. DeepQA enjoys the
merits of achieving superior performance while avoiding the
expensive or even biased gene selection during the process
of model fitting. Figure1 presents a graphical illustration of
DeepQA with a comparison to existing methods.
2 | Materials and Methods
From the machine learning point of view, the difficulty of using
deep learning for aging prediction with gene expression data is
mainly due to two reasons. Firstly, gene expression data can be
regarded as so- called tabular data and deep neural networks
have been experienced difficulties in dealing with tabular data
in contrast to signals, for example, images, words and so on
(Shwartz- Ziv and Armon 2022). Secondly, for age estimation,
there are often limited number of samples for training com-
pared to the high- dimensional inputs of genes (> 14 K). This is
usually referred to as small data or few shot learning in the ma-
chine learning community which has been attracting much at-
tentions. This makes aging prediction using deep learning even
harder. Technically, this is also the reason why existing methods
for age prediction usually require gene selection prior to learn-
ing a predictor, as their ability of handling high- dimensional
data is limited compared to deep learning methods (Rutledge,
Oh, and Wyss- Coray2022).
One way to tackle these difficulties is to use multiple layer per-
ceptrons (MLPs), also called dense layers, which is suitable for
processing unordered gene expression data (Cheng etal.2024;
Agarwal et al. 2021), and to combine multiple cohorts of sam-
ples and create a larger database for training. Furthermore, by
stacking MLPs to form a Mixture of Experts (MoEs), we are able
to train a unified aging clock which predict biological age across
diverse types of samples from different tissues, sexes, or health
conditions.
2.1 | DeepQA: Quantifying Aging With Mixture
of Experts on Both Healthy and Unhealthy Subjects
Mixt ure of experts has been w idely used in various applications
in healthcare, recognition etc. It was originally introduced by
(Jacobs etal.1991) which is an ensemble of several experts/
sub- models controlled by a learnable gate. The gate learns to
dispatch each of the inputs only to one of the experts. This ar-
chitecture makes MoE an excellent learner especially suitable
for tasks with heterogeneous data where multiple modes exist
(Yuksel, Wilson, and Gader2012; Chen etal.2022).
A graphical illustration of the proposed DeepQA is shown in
Figure1. Inside DeepQA, a mixture of experts model is trained
to predict biological age given gene expression data as inputs.
The loss term for training on healthy subjects is straightforward
as their biological age is equal to the chronological age, and we
adopt mean- absolute- error (MAE) which has been widely used
for training aging clocks. However, it is nontrivial to train on
unhealthy subjects as their biological age could deviate from
their chronological age. We therefore propose a novel loss term,
named hinge- mean- absolute- error (Hinge- MAE), for unhealthy
subjects.
Formally, assume that the gene expression data of a sample is
denoted as x, a 1D vector of length d where d is the number of
genes involved. In each minibatch, DeepQA receives a tensor
X of shape
d×m
which is formed by m samples in the current
minibatch. Each of the experts
i(
)
takes X as input and makes
prediction
Yi
of 1 × m,
The total loss is then calculated as
where
pi
is the output of the gating network to weight the pre-
diction of the
ith
expert
i(
)
.
Yc
is the training targets, that is,
chronological age.
(
,
)
is the prediction error. For healthy
subjects, it is simply mean- absolute- error. For unhealthy sub-
ject, it is our proposed Hinge- MAE which is defined below.
LAHK
is the alignment to human knowledge loss.
and
γ
are positive
weights to balance the loss terms and have been set to 30 and
100, respectively.
where
Δ
is a positive number and serves as a margin to allow the
model to output predictions that deviate from the labels, that is,
chronological age of unhealthy subjects. In practice,
Δ
should be
regarded as a hyperparameter and it was set to five in our exper-
iments. The rationale behind this loss is that, the prior knowl-
edge that the biological age of unhealthy subjects could deviate
from their chronological age can be modeled as “learning with
inaccurate labels,” which is then addressed by designing an ex-
plicit margin for the prediction error.
The number of experts is a hyperparameter to tune and has been
set to two in our experiments. Each expert contains five dense
layers with 640, 256, 128, 64, and 1 neurons, respectively, along
with batch normalization and the nonlinear activation function
(1)
Yi=i(X)
(2)
L
total =
i
piL
(
Y,Yc
)
=
i
pi
(
λ
(
Yi,Yc
)
+𝛾LAHK
)
(3)
Hinge
MAE =
Ypred Yc
1,
Ypred Yc
1
Δ
0,
Ypred Yc
1<Δ
4 of 18 Aging Cell, 2025
FIGUR E  | A graphical illustration of the proposed DeepQA compared to existing methods. (a) Existing methods train an AI model for each co-
hort with homogeneous dat a as aging clocks to eva luate biological ages. The models , typically random forest, LDA with inefficient gene selection, are
usually trained only on healthy/control subjects and tested on both healthy and unhealthy subjects. (b) We propose a unified transcriptome- based
aging clock based on deep neural networks, named DeepQA. Unlike existing methods, DeepQA is trained on data of both healthy and unhealthy
subjects on multiple cohorts and is able to infer biological age much more accurate than existing methods. (c) Architecture of the proposed DeepQA
as well as specially designed loss functions for the training on both healthy and unhealthy subjects. Our DeepQA has 26.3 M trainable parameters
in total.
5 of 18
ReLU. Dropout was also applied to counter overfitting (Srivastava
etal.2014). The gating network has the same architecture with
the experts except for the head for the output. Overall, the pro-
posed DeepQA has 26.3 M trainable parameters. Detailed archi-
tectures of the experts and the gating network can be found in
Figure1c.
2.2 | Identifying Important Genes via
Saliency Map
In prior work, there are three typical ways of selecting genes,
among which the most popular one is selecting genes of can-
didates via, for example, Differential/Variable Expression
Analysis (DEA), filtering or through unsupervised learning
(Mohamadi and Adjeroh2021). It is usually computationally
efficient than iterative search since no model fitting is involved.
However, this strategy is not immune to selection bias if the la-
bels of age are used within the selection process. For instance,
when performing DEA, one at least needs to set up two groups,
for example, young and old subjects where the information of
age goes into the gene selection process. Additional care needs
to be taken to avoid selection bias. Secondly, researchers have
also proposed to find a good subset of all genes by iterative/
exhaustive search. Theoretically, it can produce good results but
is extremely inefficient as it needs to train a large number of
models. For small models, such as LDA, random forest, etc., the
computation time is almost unbearable if without a substantial
amount of computing resources. Obviously, it is infeasible for
large deep learning models. The third way is using the genes
that are known to be related to aging in literature. The disad-
vantage is also obvious, as the overall performance is limited by
using only existing human knowledge, especially when some
of them may be outdated or questionable.
Saliency map has been a very popular and effective tool to ex-
plain/visualize the behaviors of neural networks (Selvaraju
etal.2017). It can be calculated efficiently within the framework
of deep learning such as Pytorch. For our application, explaining
how DeepQA works is of course valuable, but more importantly,
saliency map allows us to infer genes that are important for age
estimation. This is much more efficient computationally than
traditional approaches such as SHAP (Lundberg and Lee2017;
Hartman etal.2023).
The saliency map for each expert
i
is basically the partial de-
rivative of its output w.r.t the input and can be calculated as
Overall, the saliency map can be calculated
In practice, this can be easily realized using automatic differen-
tiation in PyTorch.
2.3 | AHK Loss: Alignment to Human Knowledge
As a powerful data- driven method, deep neural networks are
typically able to learn knowledge purely from the training data.
So given enough data, neural networks could learn biological
knowledge spontaneously to some extent. However, it is often
beneficial to incorporate human knowledge into the archi-
tecture of the networks. For applications related to biology or
healthcare, these models are often referred to as biologically
informed neural networks (Hartman et al.2023; Fortelny and
Bock2020; Elmarakeby etal.2021).
For transcriptome- based aging estimation, datasets are rela-
tively small compared to applications in such as computer vision
or natural language processing, it becomes even more critical
for deep neural networks to embrace the biological human
knowledge accumulated over years. In this study, we propose to
align the learning process of DeepQA to the human knowledge
by adding an additional loss term, named as AHK loss, which
measures the similarity between important genes in the current
epoch and the genes known to be related to aging in the litera-
ture (defined as aging gene pool in our work). During training,
the AHK loss is combined with other losses, for example, MAE
to guide the learning process of the prediction model.
Another motivation of us proposing the AHK loss is sparsity or
gene selection. Since gene selection does not exist in our method
as a preprocessing step, a mechanism of forcing the prediction
model to use a subset of important genes such as sparsity con-
straint would be crucial. In DeepQA, we do not use sparsity
constraint such as
L1
regularization as its shrinking direction
is controlled purely by optimizing the age prediction error and
could be biased to a wrong set of genes. Instead, we designed the
AHK loss which can serve as a sparsity constraint or a soft ver-
sion of “gene selection.” Imaging an extreme case of using only
the AHK loss, the prediction would be carried out only with a
small set of known genes which are indubitably sparse.
We decide a set of genes, referred to as Aging Gene Pool, which
are known to be related to aging in the literature. We denote
this pool as
G
=
(
g
1
,,g
n)
. For every epoch during train-
ing, we compute the saliency map (gene importance) and
produce a list of genes sorted by their importance, denoted as
G
(t)=
(
g(t)
1,,g(t)
k
)
. The AHK loss is given by
Obviously,
LAHK [1, 0]
. We minimize this loss to encourage
the model to use genes in the aging gene pool
G
to make predic-
tion. When only genes in
G
are used for prediction,
LAHK =−1
.
When none in
G
is used,
LAHK =0
.
In our implementation, we compiled
N=646
genes which
are known to be related to aging as the aging gene pool (Mao
etal.2023). In practice, it can be chosen as any genes that serve
the users' applications. Particularly in our experiments, we
(4)
i=
||||
𝛿
i
(x)
𝛿x||||
(5)
total =
i
pii=
i
pi
||||
𝛿
i
(x)
𝛿x
||||
(6)
L
AHK =−
g(t)G
g
(t)
g
(t)
g(t)
6 of 18 Aging Cell, 2025
chose
n=10N
of total genes randomly to compute the AHK
loss.
n
should be treated as a hyperparameter.
2.4 | Simulating Inherent Noises in Gene Expression
Data Using Principal Component Analysis
To examine the robustness of aging clocks against inher-
ent noises in the gene expression data, we employed a classic
technique principal component analysis (PCA) to find factors
which may act as sources of noises (to be exact, small variations)
(Bishop2006; Manjón, Coupé, and Buades2015). We simply will
refer to this kind of noises as PCA noises. By manipulating these
factors, we can simulate inherent noises which are used to test
the robustness of an aging clock.
Formally, any (normalized) sample of gene expression data can
be decomposed into a linear weighted combination of principal
components (PCs, which are learned from the data),
where the first
L
terms correspond to PCs contributed to 95% of
the variances of the data. The rest of the terms (
NL
) in total
contributed to 5% of the variances which are often regarded as
noises. This allows us to simulate inherent noises by manipulat-
ing the weights of these “noisy” PCs, that is,
where
𝜌>0
and
x
is the noisy version of the original sample
x
.
More details of generating PCA noises can be found in the sup-
port document.
3 | Results
3.1 | Database and Evaluation Protocol
The database used in our study is one of large publicly avail-
able database which is a collection of multiple human bulk
RNA- Seq datasets compiled by the reference (Shokhirev and
Johnson 2021). It is publicly available in the Sequence Read
Archive (https:// www. ncbi. nlm. nih. gov/ sra). The raw gene
count table and metadata can be downloaded from Mendeley
(https:// doi. org/ 10. 17632/ 92rgn swtn8. 1).
This database consists of 31 datasets produced by different
labs from the references, and contains 3060 samples in total
which cover young, adult, and old human subjects with var-
ious health conditions such as Alzheimer's disease (AD),
Schizophrenia, Age- related Macular Degeneration (AMD),
Dilated Cardiomyopathy (DCM), Dysplasia, and others. Samples
were collected across different organs. Details can be found in
Table 1. For data preprocessing and filtering, we followed the
protocol in the reference (Shokhirev and Johnson 2021). For
convenience, we will refer to this database as “MCATS” to high-
light that it includes samples of multiple cohorts across tissues,
healthy conditions and sex.
To evaluate the performance of an aging clock, two cases need
to be considered:
For healthy subjects, it is straightforward and we can simply use
the mean- absolute- error between the predicted and chronologi-
cal age, since the biological age of a healthy subject is (assumed
to be) equal to its chronological age.
For unhealthy subjects, their biological age often does not match
the chronological age, we adopt a qualitative analysis used in
(Jonsson et al. 2019), termed as PAD significance, where PAD
stands for “Predicted Age Difference” or more accurately the
difference between the predicted and chronological age. PAD
has been widely used for estimating the deviation of a subject
from normal aging. Note that PAD on subjects of interest itself is
insufficient to determine normal or abnormal aging, unless the
aging clock works perfectly for healthy subjects with zero pre-
diction error which rarely happens in reality. So instead of look-
ing into PADs of subjects of interest themselves, one prepares
a control group of healthy subjects unseen by the aging clock
during training, and calculate their PADs. By testing whether
these two groups of PADs are statistically different from one
another, one can determine whether abnormal aging exists
(Supporting Information: SectionS1).
In our study, besides healthy subjects, we chose three unhealthy
conditions that are known to be related to accelerated aging,
(7)
x=
N1
k=0
𝜇kPCk=
L1
k=0
𝜇kPCk+
N1
k=L
𝜇kPC
k
(8)
x=
L1
k=0
𝜇kPCk+
N1
k=L
𝜌𝜇kPC
k
TABLE  | Details of the samples used in this study including the organs where the samples were taken from, number of healthy/unhealthy
samples from different organs, age distributions etc.
Organ #References Healthy Unhealthy Male Female Age (years)
Brain 391 Alzheimer's disease (AD) 133 144 80 32 ~ 103
Schizophrenia 57 125 23 24 ~ 99
Retina 6176 Age- related Macular Degeneration (AMD) 406 285 297 47 ~ 107
Heart 2162 Dilated Cardiomyopathy (DCM) 166 174 154 15 ~ 83
Lung 125 Dysplasia 50 51 24 47 ~ 77
Others 23 901 893 736 953 0.2 5 ~ 86
Total 35 1355 1705 1440 1515 0.25 ~ 107
Note: This database is a collection of 31 datasets from the references. Note that 105 samples of the category of “Other s” do not have the gender specified.
7 of 18
namely Alzheimer's disease, Schizophrenia, age- related macu-
lar degeneration (AMD), and two unhealthy conditions that are
not known to be related to unhealthy aging, namely dilated car-
diomyopathy (DCM) and dysplasia (Jonsson et al.2019; Chang
etal.2019; Bashyam etal.2020). Details of samples of these con-
ditions including gender, age distribution are shown in Table1.
3.2 | Competing Methods and Implementation
Details
Almost all non- deep- learning methods require explicit gene
selection, we therefore adopted four popular methods, namely
Gene expression filtering (Fleischer et al. 2018), variable and
differential genes (Shokhirev and Johnson2021), and AgingMap
(Mao etal. 2023). For iterative searching of genes (Meyer and
Schumacher 2021), we had to exclude it from comparison be-
cause it demands too much computational resources. With our
hardware and experimental settings, it takes months to complete
a single run. If we adopt the leave- one- out scheme as in (Meyer
and Schumacher2021), it would take many years to finish the
experiments.
Methods of LDA, random forest and elastic net were imple-
mented based on sklearn (Pedregosa etal.2011). Our investiga-
tion showed that only random forest among these conventional
methods benefited from data augmentation with Gaussian
noise, and the corresponding results were obtained with data
augmentation. CNN- 2D were proposed with data augmentation
and our implementation followed the design and training proto-
col in (Mohamadi etal.2021).
DeepQA was implemented in Python (version 3.8.10) and based
on the popular deep learning framework PyTorch (Paszke
etal.2019) (version 1.12.1). The number of experts in MoE was
2. The optimizer was Adam (Kingma and Ba2015). The learning
rate was set initially as 5e- 3 and decayed in step wise by 0.8. The
batch size was 512. For each training sample, 30 artificial sam-
ples augmented with Gaussian noises were generated for train-
ing. The experiments were conducted on servers with multiple
NVIDIA GeForce RTX 3090 and 3080 GPUs.
3.3 | Accuracy in Biological Age Prediction
As discussed previously, the evaluation protocols of “training
and test” in many existing works are flawed when gene selection
is involved and lead to selection bias. So in our study, following
the suggestion in (Ambroise and McLachlan 2002), we carried
out 10- fold cross- validation. For the implementation of compet-
ing methods, we have ensured no “test set leakage” would occur
by performing gene selection within each fold. We reported the
averaged prediction accuracy.
Table2 presents the comparison of the proposed DeepQA with
existing state- of- the- art methods on both healthy and unhealthy
subjects. The first two row shows the prediction performance of
models trained on healthy subjects of each cohort, respectively.
For each cohort, we trained a model to predict their biological age.
This is a common practice for biological age prediction in existing
work. Row 3 ~ 10 show results of models trained on healthy sub-
jects of multiple cohorts, and the last row shows the performance
of our DeepQA trained on both healthy and unhealthy subjects
from multiple cohorts. A graphical comparison of DeepQA and
the competing methods can also be found in Figure3a.
For healthy subjects, DeepQA outperformed all the compared
methods signif icantly in terms of both MA E and
R2
and achie ved
TABLE  | Prediction accuracy of the compared aging prediction methods on healthy and unhealthy subjects.
Method Trained on
Healthy Unhealthy
MAER2
AD AMD Schi. DCM Dysp. Score
(A.) (A.) (A.) (N.) (N.) (5)
RF- Differential Healthy, per cohort 8.307 0.574 N. A. A. N. N. 4
RF- Variable Healthy, per cohort 7.9 67 0.599 N. A. A. N. N. 4
ELDA- GEF Healthy 6.366 0.689 N. N. A. A. A. 1
CNN- 2D Healthy 8.778 0.488 A. A. A. A. N. 4
RF- Differential Healthy 7.284 0.686 N. N. A. A. A. 1
RF- Variable Healthy 6.821 0.711 N. N. A. A. A. 1
EN- Differential Healthy 8.512 0.580 A. N. A. A. A. 2
EN- Variable Healthy 7.995 0.634 N. N. A. A. A. 1
EN- AgingMap Healthy 8.292 0.582 N. N. A. A. A. 1
EN- All Healthy 6.931 0.693 N. N. A. A. A. 1
Deep QA(ours) Healthy 5.295 0.744 A. N. A. A. A. 2
Healt hy + unhe althy 4.820 0.789 A.A.A.N.N.5
Note: For unhealthy subjects of five conditions, A. stands for ac celerated aging and N. stands for no accelerated aging. Schi. stands for Schizophrenia. Dysp. stands for
Dysplasia.
8 of 18 Aging Cell, 2025
the lowest prediction error of 4.820 in MAE which is 1.5 year
more accurate than the second best method ELDA- GEF. For
unhealthy subjects, only our DeepQA correctly identified the
aging status of all the conditions. Please refer to TablesS1–S11
for detailed results on unhealthy subjects. It is worth noting that
even for DeepQA itself, training on both healthy and unhealthy
subjects improved the prediction performance significantly,
demonstrating the necessity of exploiting data of unhealthy sub-
jects (Table2).
The good performance of ELDA- GEF on healthy subjects was
mainly due to an ensemble of a substantial number of trained
models. However, this design also makes this method very inef-
ficient. Using ensemble technique here indeed improved the pre-
diction performance, but also increased the risk of overfitting.
Another interesting observation is that elastic net performed the
best when fed with all the genes which indicated that its sparsity
constraint played a key role in selecting genes in the process of
model fitting. On the other hand, as we will discuss in the next
subsection, the robustness against noise of these methods was
unsatisfactory which indicated a certain degree of overfitting.
Interestingly, the accuracy of the CNN- 2D method on healthy
subjects is rather low, while it did perform well on unhealthy
subjects. This may indicate that deep neural networks could
be especially suitable for the task of aging estimation, yet the
specific architecture of 2D convolutional networks may not be
a good choice though. This is due to the inherent flaw of con-
verting gene expression data into 2D images, as neighborhood
relations of genes can be arbitrary. In other words, unlike real
images, these pseudo- images of gene expression data contain
false neighboring information which leads to poor performance,
especially when combined with the age- bin strategy.
Since our database is heterogeneous and contains multiple data-
sets or cohorts, we therefore also calculated the prediction per-
formance of DeepQA along with existing competing methods
on healthy subjects from different groups of samples, namely
male, female, young, adult, old, brain, retina, heart, lung, and
other diseases in Figures2 and 3b. It is evident that the proposed
DeepQA outperformed all the other methods. This suggested
that DeepQA has the potential to be a universal aging clock
which learns to predict biological age of samples from transcrip-
tome data across different organs, sex and age with sufficient
data for training.
3.4 | Robustness to Noises in Gene Expression
The raw data of RNA sequencing are often contaminated by
multiple types of noises, for example, technical noise, biological
variation, ex perimental settings, and ma ny other factors (Conesa
etal.2016). Though that denoising and imputation methods for
data cleaning are available, the performance of which is not al-
ways satisfactory (Eraslan etal.2019; Moutsopoulos etal.2021;
Li, Brouwer, and Luo2022). As a result, the downstream anal-
ysis such as age prediction here becomes even more challeng-
ing due to the presence of noises. It is therefore important to
evaluate and improve the robustness of the prediction methods
against noises.
Following the general practice in both biology and machine
learning (Bishop 2006; Li, Brouwer, and Luo2022), and more
importantly to simulate the effect of many different noise factors
combining together (according to the Central Limit Theorem),
we injected Gaussian noise into the gene expression data, and
tested the robustness of the proposed method accordingly at dif-
ferent levels of noises (Figure3c). When the magnitude of noises
was smaller than 5, our method significantly outperformed
other competing methods. While extremely large noises were
imposed and too much of information and structure in data
were destroyed, none of these methods were able to predict with
an error of less than 10. Some of them performed extremely bad
though.
To have a clear view of the comparison, we have also plotted
the prediction accuracies of the competing methods at a typical
level of noises (1.5 for Gaussian, 3% for Dropout) (Figure3e,f).
DeepQA is much more robust than any other competing meth-
ods with a large margin (> 3 years in MAE). In contrast, the
second best method ELDA- GEF performed extremely unsatis-
factorily against the interference of the noises, which indicated
that overfitting might have occurred in its good result in Table2
(Bishop2006; Xu, Caramanis, and Mannor2009; Goodfellow,
Bengio, and Courville2016). Similarly, EN with all genes per-
formed much better than EN with variable or differential genes,
but showed comparable robustness which also indicated a cer-
tain degree of overfitting. It is expectable that CNN- 2D has
shown to be quite robust, but its prediction accuracy was too low
to be an appealing prediction model. Besides Gaussian noise, we
have also tested robustness of these methods against dropout
noises (Figure3e,f).
Besides injecting noises directly into the gene expression data,
we are also interested in how DeepQA and other methods would
respond to the inherent noises in the gene expression data.
Hence, following the popular approach of using PCA to isolate
small variations which may be regarded as “noises.” Results are
reported in Figure 3d. Overall, our DeepQA showed superior
robustness among the compared methods. Interestingly, the
compared methods exhibited three distinct dynamics: DeepQA,
CNN- 2D, RF- Variable and RF- Differential formed a leading
group which significantly outperformed the other two groups.
When the magnitude of the noises reached to 20, the MAEs
were still about impressively 10. The second group of methods
were elastic- net- based ones where the prediction errors almost
linearly grew when the noises become larger, indicating a very
poor robustness. Lastly, the response cur ve of ELDA- GEF is rad-
ically different from the other two groups where it grew very fast
when the noise was small, but saturated beyond 7. These behav-
iors may indicate that the second and three groups of methods
were biased to rely more heavily on factors with small variations
(including inherent noises) or large. In contrast, the first group
of methods including DeepQA may be able to learn features in
a balanced way.
3.5 | Robustness to Noises in Age Labels
The MoE model in DeepQA are trained with pairs of data<gene
expression data, age label>and we have shown that our method
9 of 18
FIGUR E  | Legend on next page.
10 of 18 Aging Cell, 2025
was remarkably robust to the noises in the input gene expression
data. Moreover, the age labels may also be corrupted by noises.
Particularly they in fact already contain a certain degree of in-
herent uncertainty due to the inaccuracy in quantifying age in
years. A simple and straightforward example is that the “true”
chronological age of two persons with the same age label (e.g.,
60 years) could differ in 1 year if they were born in the beginning
and the end of the same year, respectively.
To examine how our method responds to this kind of “label
noise,” we designed and conducted a set of experiments where
Gaussian noises were added on age labels of all training sam-
ples. Results were reported in Figure3g. The x- axis represents
the std. of the Gaussian noises, and the y- axis shows the predic-
tion accuracy on healthy subjects. Overall, DeepQA was able to
handle label noises of reasonably large magnitude very well (std
< 3 years). When label noises were too strong, it is no surprise
that the performance of DeepQA decreased.
3.6 | Computational Efficiency
We examined the computational efficiency and scalability of
DeepQA and reported the training hours and prediction accu-
racy on a single N VIDIA GeForce RT X 3090 GPU when different
number of training samples were used (Figure3h). The training
time grew basically linearly when the size of the training data
increased. Note that we haven't used sophisticated techniques to
speed up the training process, other than standard optimization.
When a large dataset is available, techniques such as distributed
and parallel training can significantly boost the training speed
(Li etal.2020).
3.7 | Impact of the Hyperparameters
The values of hyperparameters in DeepQA were determined
manually following the standard protocol (Goodfellow,
Bengio, and Courville2016). To investigate how DeepQA is
affected by its key hyperparameters, we conducted a series
of experiments where DeepQA was trained with different
values of hyperparameters and the corresponding prediction
accuracy on both healthy and unhealthy subjects were re-
ported (Figure 4). Namely, these hyperparameters were the
margin for the Hinge- MAE loss
Δ
, the weight to balance the
loss terms
𝛾
, the percentage of genes used in the AHK loss
n
N
,
and the number of experts. For clarity, we have also plotted
the MA E alone when different hyperparameters were chosen
(Figure4e).
The parameters
n
N
and
𝛾
control how the AHK loss contributes
to the learning process, and interestingly the prediction error in
MAE on healthy subjects varied slightly when both of these two
parameters became larger (Figure 4e), while the significances
on unhealthy subjects showed radical changes (Figure 4b,c).
This demonstrated that the AHK loss was crucial for detecting
aging acceleration (or not).
In the Hinge- MAE loss, when the parameter
Δ
became larger
than 7, the prediction error in MAE on healthy subjects in-
creased significantly, and the significances on unhealthy sub-
jects of AD, AMD and Dysplasia flipped wrongly (Figure4a,e).
The number of experts controls the model capacity of DeepQA,
and its impact on age prediction accuracies were also quite typi-
cal. Small or large values of this parameter decreased the perfor-
mance as the model was experiencing underfitting or overfitting
(Figure 4d,e). In practice, if a large amount of data is available
for training, we can increase the number of experts to enlarge
the model to learn from big data.
Interestingly, our method is relatively insensitive to its hyperpa-
rameters. This suggests that the training of DeepQA converged to
flat minima of the loss function (Equation2), which usually indi-
cated good generalization (Keskar etal.2017; Cha etal.2021).
3.8 | Explainability and Clinical Applicability via
Identifying Important Genes
In general, understanding how deep neural networks work is very
challenging. For aging clocks, we can start from investigating im-
portant genes identified by DeepQA and checking their biological
plausibility. To this end, we performed the gene enrichment anal-
ysis on the top- 1000 important genes identified by DeepQA on all
the samples (Figure 5a). We showed results of healthy subjects,
Alzheimer's disease, Schizophrenia, age- related macular degener-
ation, dilated cardiomyopathy and dysplasia.
The healthy samples (Figure 5a) can be categorized into three
groups, namely young (< 30 years), adult (30 ~ 70 years), and old
(> 70 years), so we also used DeepQA to identify genes prominent
in each of these three groups, and performed KEGG, respectively.
The results ind icate that in the young g roup, there is an enrichment
of pathways in cancer, metabolic pathways, and nicotinate and
nicotinamide metabolism (Figure5b). The enrichment of various
cancer pathways indicated stem cells properties (Reya etal.2011).
In the adult group, the enrichment is observed in metabolic path-
ways, pathways in cancer, and the cell cycle pathway related to
p53 (Figure5c) (Hafner etal.2019). Metabolic processes, ferropto-
sis, legionellosis (common in the elderly people [Flanagan2020]),
and protein degradation pathways such as protein processing in
endoplasmic reticulum are enriched in the old group (Figure5d)
(Krshnan, van de Weijer, and Carvalho2022).
We also examined the senescence- associated secretor y pheno-
type (SASP)- related genes among the top 1000 genes identified
by DeepQA on healthy and unhealthy samples (Figure 5e,f;
Figure S2; TableS12) (Suryadevara etal.2024; Qi etal.2024),
followed by GO analysis detecting gene functions of the over-
lapped genes. While subjects of AD and AMD share similar age
FIGUR E  | Prediction performa nce of DeepQA compared with ex isting methods on healthy subject s of different groups, namely all, brain , retina,
heart, lung, and others from top to bottom. Columns correspond to DeepQA, ELDA- GEF, RF- Variable, and EN- All, respectively. In each plot, the x-
axis is the true age and the y- axis is the predicted age. DeepQA outperformed existing methods on all groups.
11 of 18
FIGUR E  | Legend on next page.
12 of 18 Aging Cell, 2025
range, AD was enriched of negative regulation of cell popula-
tion proliferation, cell cycle, and regulation of cell population
proliferation (Figure 5g). While AMD was enriched of cellu-
lar response to hormone stimulus, negative regulation of cell
population proliferation, response to estradiol and cell cycle
(Figure 5h). These showed that both aging- related diseases
including AD and AMD are enriched in cell cycle and cell
proliferation.
DeepQA was able to learn to identify and made use of genes for
age prediction in a biological plausible manner. It learned to
predict biological ages using not only genes that are associated
to aging in general, but also ones that are linked to the charac-
teristics/conditions of a particular group that a sample belongs
to. This serves as a biological validation of the explainability of
DeepQA.
More importantly, the ability of DeepQA identifying important
genes can also be an effective tool for discovery of novel gene
biomarkers, alternative to (not necessarily replacing) such as
differentially expressed genes (DEGs) analysis. This is another
clinical usage of DeepQA besides age prediction as a diagnosis
or evaluation tool for certain diseases. Here we demonstrated
how DeepQA can be jointly used with DEG analysis to identify
potential important genes.
We chose a batch of brain samples containing both healthy con-
trols and unhealthy samples with AD (Friedman et al. 2018).
We used both DESeq2 and DeepQA to analyze the samples
of AD or healthy. The top 1000 genes of DeepQA were over-
lapped with the DEGs, showing that 57 genes were overlapped
(Figure6a). Among overlapped genes, most of them were upreg-
ulated DEGs in AD (Figure6b). GO analysis revealed biological
processes related to negative regulation of cell population pro-
liferation, positive regulation of interleukin8 production and
extracellular matrix organization (Figure 6d). KEGG analysis
of overlapped target genes identified upregulation of various
aging related pathways, including MAPK signaling pathway,
longevity regulating pathway and PI3K Akt signaling pathway
(Figure6e) (López- Otín etal.2023; Nikoletopoulou, Kyriakakis,
and Tavernarakis2014). In addition, PPI network for overlapped
genes was generated using an online tool String (Szklarczyk
etal.2023) (Figure6c). Nodes with different colors represented
different proteins and the edges between nodes represented
known interactions or predicted interactions between these
proteins. In summary, these results showed that DeepQA can
produce meaningful genes overlapped with DEGs which are as-
sociated with aging in AD. In other words, it can help to further
filter out genes generated by DEG analysis, and be an effective
and efficient tool in complementary to DEG analysis. Please
note that under this scenario, train- test split is no longer a con-
cern, and any sample can be used to produce potential genes of
importance.
4 | Discussion
We proposed a deep- learning- based aging clock, named as
DeepQA, which predicts biological age of an individual based
on transcriptome data, and discovers genes of importance
via analyzing saliency maps efficiently. With a simple yet ef-
fective Hinge- MAE loss and the architecture of mixture of
experts which is especially suitable for modeling heteroge-
neous data, DeepQA can train on both healthy and unhealthy
subjects of multiple cohorts to improve the performance of
estimating biological age, especially for unhealthy subjects.
Compared to existing methods, our method enjoys the follow-
ing advantages.
Firstly, it can predict biological age of healthy and unhealthy
subjects from gene expression data much more accurately than
existing methods. In our experiments, DeepQA outperformed
existing methods with a large margin on a database which con-
sists of several public datasets. On healthy subjects, our DeepQA
achieved superior prediction performance which was 1.5 years
better than the second best method. For unhealthy subjects,
DeepQA was the only one with correct prediction on subjects of
all five unhealthy conditions.
Secondly, it does not require gene selection as a preprocessing
step and can identify important genes via analyzing saliency
maps after model fitting. DeepQA avoids the extremely ineffi-
cient exhaustive search of genes, and is therefore also immune
to the selection bias in gene selection process which numer-
ous existing methods have suffered from. It provides a novel
means to identify genes activated in aging prediction, alterna-
tive to such as differential gene expression analysis. We val-
idated this by showing important genes selected by DeepQA
were indeed meaningful, and consistent with known human
knowledge.
It should be emphasized that all existing methods (aging clocks)
of estimating biological age were proposed to train on (data
of) healthy subjects only. To the best of our knowledge, our
DeepQA is the first method which can train on both healthy and
unhealthy subjects and shows superior performance of predict-
ing biological age compared to existing methods. The key is a
new loss we proposed, called Hinge- MAE loss, combined with
FIGUR E  | (a) Age prediction accuracy of DeepQA and existing methods on both healthy and unhealthy subjects. T his plot is a graphical illustra-
tion of Table2. For healthy subjects, MAE was used to measure the good ness of prediction. For unhealthy subjects, statistical signif icance, log10(p-
value), was plotted to show the plausibility of prediction. (b) Comparison of DeepQA and existing methods on different groups of healthy samples.
DeepQA outperformed existing methods in all the groups. (c) Robustness of DeepQA and other methods against simulated Gaussian noises in gene
expression data at different levels. (d) Robustness of DeepQA and other methods against simulated PCA noises in gene expression data at different
levels. (e) Robustness in MAE of DeepQA and other methods against Gaussian or Dropout noises at a ty pical magnitude of 1.5. (f) Robustness in R2
of DeepQA and other methods against Gaussian or Dropout noises at a typical magnitude of 1.5. (g) Impact of noises in ages (labels) on prediction
accuracy of DeepQA on healthy subjects. (h) Training time of DeepQA w.r.t the size of the training data. The diameters of the blue solid dots corre-
spond to the prediction error in MA E.
13 of 18
FIGUR E  | Age prediction accuracy of DeepQA with different hyperparameter. (a)
Δ
, the margin for the Hinge- MAE loss. (b)
𝛾
, weight to bal-
ance the loss terms. (c)
n
N
, the percentage of genes used in the A HK loss. (d) The number of experts. The red circle corresponds to - log10(0.05) which
is the threshold for the signif icance test. For AD, AMD, and Schizophrenia, the data points should be located outside the red circle which indicates
aging acceleration has been detected. For the other two conditions, data points should be inside the red circle. (e) MAE on the healthy subjects with
different hyperparameters, among which the number of experts was the most sensitive one af fecting the prediction accuracy on the healthy subjects.
The optimal values of these hyperparameters were shown in the main text.
14 of 18 Aging Cell, 2025
the network architecture of mixture of experts, to train the AI
model on unhealthy subjects. This loss for unhealthy subjects is
general and can be used for any AI model.
One of key findings in our study is that existing methods per-
formed rather poorly on unhealthy subjects when trained on
heterogeneous samples of multiple cohorts, compared to on
each homogeneous cohort, as shown in Table2. A disadvantage
of the latter is multiple models are required to be trained and
deployed. This is certainly an inefficient solution which also
cannot benefit from large (heterogeneous) data. In contrast,
our DeepQA demonstrated clearly its ability and advantages of
learning from heterogeneous data, thanks to the architecture
of mixture- of- experts and the Hinge- MAE loss. In this sense,
DeepQA provides a “unified” solution for estimating biological
age from gene expression data.
To demonstrate the clinical usage of DeepQA, we examined the
SASP genes predicted in six groups and found that they were en-
riched in metabolism, cell cycle, and cellular proli feration, show-
ing similarities in physiological processes. Moreover, among
healthy individuals of different ages, the pathways in the old
group significantly differed from those in the other two groups,
while the data range for AD and AMD spanned from adult to
old age, sharing similar enriched pathways with the healthy old
group, indicating the plausibility of the prediction results. For
FIGUR E  | (a) KEGG pathways enriched on the top- 1000 genes identified to be important for age prediction on the group of healthy subjects
and five unhealthy conditions by DeepQA. (b–d) KEGG analysis of genes identified by DeepQA from young (b), adult (c), and old healthy groups (d),
respectively. (e–f) DeepQA genes overlapped with SA SP genes in AD (e) or AMD (f). (g–h) GO analysis of overlapped genes from AD (g) or AMD (h).
In all graphs, top 1000 genes with DeepQA were selected. In all the bubble charts, color gradient indicates the level of statistical significance (log10
(p- value)) and the dot size indicates the number of genes. The gene enrichment analysis was performed using the online DAVID tools (Sherman
etal.2022).
15 of 18
non- aging- related diseases, such as DCM, enriched pathways
were mostly related to the characteristics of the disease itself,
instead of age, including fatty acids, purine, and nucleotide
(Gigli etal. 2024). In subsequent analyses using a batch of AD
data, DeepQA was jointly used with DESeq2. Compared with
healthy individuals, there were 57 age- related genes predicted
by DeepQA that showed differential expression in AD. Among
these, MAPK and PI3K- Akt are related to the pathogenesis of
FIGUR E  | DeepQA and DESeq2 joint analysis of RNA- seq dates from Alzheimer's disease on samples from (Friedman etal.2018). (a) Overlap
between DEGs of AD and top 1000 genes of DeepQA from AD. (b) Expression heatmap of overlapped genes. Average expression level of each age bin
was determined and presented as Z- score across different bins. (c) The protein–protein interaction (PPI) network of the overlapped genes. GO anal-
ysis (d) or KEGG analysis of pathways (e) related to upregulated genes of DEGs from overlapped genes. The analysis was done with DAVID. Color
shade indicated significance and dot size indicated gene count of the corresponding pathway.
16 of 18 Aging Cell, 2025
AD or aging (Kitagishi et al. 2014; Dai et al. 2016), while le-
gionellosis and longevity regulating pathways exhibit charac-
teristics of the elderly. Notably, the impact of legionellosis and
measles on the aging process in AD has not been reported and
merits further exploration.
Our current aging clock is gene- centric, focusing on genes
within the transcriptome. For future work, we will investigate
the possibility of including more data, for example, transcripts
from noncoding sequences like transposable elements which
contribute to the aging process and can serve as aging mark-
ers (Lu 2023; Bourque et al. 2018; Liu et al. 2023; De Cecco
et al. 2019), and for example, DNA methylation data, to form
multimodal aging clocks to further improve the prediction ac-
curacy through multimodal learning. Moreover, we shall also
further investigate the DeepQA's explainability and how it can
be used to discover novel biomarkers, besides predicting biolog-
ical age.
Author Contributions
Hongqian Q i: conceptualiz ation, methodology, data curation a nd anal-
ysis, writing – original draft, writing – review and editing. Hongchen
Zhao: methodology, data curation and analysis, writing – rev iew and
editing. Enyi Li: methodology, data curation and analysis, w riting –
review and editing. Xinyi Lu: data analysis, funding acquisition, writ-
ing – review and editing. Ningbo Yu: conceptualization, data analysis,
funding acquisition, writing – review and editing. Jinchao Liu: con-
ceptualization, data analysis, funding acquisition, supervision, writing
– origina l draft, and wr iting – review and editing. Jianda Han: concep-
tualization, data analysis, funding acquisition, supervision and writing
– review and editing.
Acknowledgments
This work was supported by National Key Research and Development
Progra m of China, Grant Number: 2 022YFA110380 0, 2022YF B4703204 ,
National Natural Science Foundation of China, Grant Number:
62076140, U1913208, Natural Science Foundation of Tianjin, Grant
Number: 21JCQNJC00010.
Conflicts of Interest
The authors declare no conflicts of interest.
Data Availability Statement
The data that we have used in this study are publicly available. The de-
tails can be found in the main text. The code for DeepQA will be re-
leased publicly at https:// github. com/ Chaos cende nce/ DeepQA.
References
Agar wal, R., L. Melnick, N. Frosst , etal. 2021. “Neural Additive Models:
Interpretable Machine Learning With Neural Nets .Advances in Neur al
Information Processing Systems 34: 4699– 4711.
Aging B iomarker Consor tium, Y. Jia, J. Wang, etal. 2023. “A Framework
of Biomarkers for Brain Aging: A Consensus Statement by the Aging
Biomarker Consortium.” Lifest yle Medicine 2, no. 3: lnad017.
Ambroise, C., and G. J. McLachlan. 2002. “Selection Bias in Gene
Extraction on the Basis of Microarray Gene- Expression Data.”
Proceedings of the National Academy of Sciences of the United States of
America 99, no. 10: 6562–6566.
Bao, H., J. Cao, M. Chen, et al. 2023. “Biomarkers of Aging.” Science
China Life Sciences 66, no. 5: 893–1066.
Bashyam, V., G. Erus, J. Doshi, etal. 2020. “MR I Signatures of Brain
Age and Disease Over the Lifespan Based on a Deep Brain Network and
14 468 Individuals Worldwide.” Brain: A Journal of Neurology 143, no.
7: 2312–2324.
Bishop, C. 2006. Pattern Recognition and Machine Learning. Berlin,
Germany: Springer.
Bocklandt, S., W. Lin, M. E. Sehl, et al. 2011. “Epigenetic Predictor of
Age.” PLoS One 6, no. 6: e14821.
Bourque, G., K. H. Burns, M. Gehring, etal. 2018. “Ten Things You
Should Know About Transposable Elements.” Genome Biology 19:
1–12.
Ceglia, N., Z. Sethna, S. S. Freeman, et al. 2023. “Identification of
Transcriptional Programs Using Dense Vector Representations Defined
by Mutual Information With GeneVector.” Nature Communications 14,
no. 1: 440 0.
Cha, J., S. Chun, K. Lee, etal. 2021. “SWAD: Domain Generalization
by Seeking Flat Minima.” Advances in Neural Information Processing
Systems 34: 22405–22418.
Chang, A. Y., V. F. Skirbekk, S. Tyrovolas, N. J. Kassebaum, and J. L.
Dieleman. 2019. “Measuring Population Ageing: An Analysis of the
Global Burden of Disease Study 2017.Lancet Public Health 4, no. 3:
e159–e167.
Chen, Z., Y. Deng, Y. Wu, Q. Gu, and Y. Li. 20 22. “Towa rds Understanding
the Mixture- Of- Experts Layer in Deep Learning.” Advances in Neural
Information Processing Systems 35: 23049–23062.
Cheng, Y., R. Hu, H. Ying, X. Shi, J. Wu, and W. Lin. 202 4. “Arithmetic
Feature Inter action Is Necessar y for Deep Tabular Le arning.” Proceedings
of the AA AI Conference on Artificial Intelligence 38: 11516–11524.
Conesa, A., P. Madrigal, S. Tarazona, et al. 2016. “A Survey of Best
Practices for RNA- Seq Data Analysis.” Genome Biology 17, no. 1: 13.
Dai, H., W. Hu, L. Jiang, L. Li, X. Gaung, and Z. Xiao. 2016. “p38 MAPK
Inhibition Improves Synaptic Plasticity and Memory in Angiotensin II-
Dependent Hypertensive Mice.” Scientific Reports 6, no. 1: 27600.
De Cecco, M., T. Ito, A. P. Petrashen, et al. 2019. “L1 Drives IFN in
Senescent Cells and Promotes Age- Associated Inf lammation.” Nature
566, no. 7742: 73–78.
Du, J., P. Jia, Y. Dai, C. Tao, Z. Zhao, and D. Zhi. 2019. “Gene2vec:
Distributed Representation of Genes Based on Co- Expression.” BMC
Genomics 20, no. 1: 7–15.
Elmarakeby, H. A., J. Hwang, R. Arafeh, et al. 2021. “Biologically
Informed Deep Neural Network for Prostate Cancer Discovery.” Nature
598, no. 7880: 348–352.
Eraslan, G., L. M. Simon, M. Mircea, N. S. Mueller, and F. J. Theis. 2019.
“Single- Cell RNA- Seq Denoising Using a Deep Count Autoencoder.”
Nature Communications 10, no. 1: 390.
Flanagan, N. M. 2020. “Legionellosis in Long- Term Care: An Update.”
Caring Ages 21, no. 1: 3.
Fleischer, J. G., R. Schulte, H. H. Tsai, etal. 2018. “Predicting Age From
the Transcriptome of Human Dermal Fibroblasts.” Genome Biology 19:
1–8.
Fortelny, N., and C. Bock. 2020. “Knowledge- Primed Neural Networks
Enable Biologically Interpretable Deep Learning on Single- Cell
Sequencing Data.” Genome Biology 21, no. 1: 190.
Friedman, B. A., K. Srinivasan, G. Ayalon, etal. 2018. “Diverse Brain
Myeloid Express ion Profiles Reveal D istinct Microglia l Activation States
and Aspects of Alzheimer's Disease Not Evident in Mouse Models.” Cell
Reports 22, no. 3: 832–847.
17 of 18
Gigli, M., D. Stolfo, M. Merlo, G. Sinagra, M. R. Taylor, and L. Mestroni.
2024. “Pathophysiology of Dilated Cardiomyopathy: From Mechanisms
to Precision Medicine.” Nature Reviews Cardiology: 1–16. https:// doi.
org/ 10. 1038/ s4156 9- 024 - 01074 - 2.
Goodfellow, I., Y. Bengio, and A. Cour ville. 2016. Deep Learning.
Cambridge, M A: MIT Press.
Hafner, A., M. L. Bulyk, A. Jambhekar, and G. Lahav. 2019. “The
Multiple Mechanisms T hat Regulate p53 Activit y and Cell Fate.” Nature
Reviews Molecular Cell Biology 20, no. 4: 199–210.
Hannum, G., J. Guinney, L. Zhao, et al. 2013. “Genome- Wide
Methylation Profiles Reveal Quantitative Views of Human Aging
Rates.” Molecular Cell 49, no. 2: 359–367.
Hartman, E., A. M. Scott, C. Karlsson, et al. 2023. “Interpreting
Biologically Informed Neural Networks for Enhanced Proteomic
Biomarker Discovery and Pathway Analysis.” Nature Communications
14, no. 1: 5359.
Horvath, S. 2013. “DNA Methylation Age of Human Tissues and Cell
Type s.” Genome Biology 14: 1–20.
Jacobs, R . A., M. I. Jordan, S. J. Nowla n, and G. E. Hinton. 1991. “Adaptive
Mixtures of Local Experts.” Neural Computation 3, no. 1: 79–87.
Jonsson, B. A., G. Bjornsdottir, T. E. Thorgeirsson, et al. 2019. “Brain
Age Prediction Using Deep Learning Uncovers Associated Sequence
Var ia nt s.” Nature Communications 10, no. 1: 5409.
Keskar, N. S., D. Mudigere, J. Nocedal, M. Smelyanskiy, and P. T. Tang.
2017. “On Large- Batch Training for Deep Learning: Generalization
Gap and Sharp Minima” International Conference on Learning
Representations.
Kingma, D. P., and J. Ba. 2015. “Adam: A Method for Stochastic
Optimization.” International Conference on Learning
Representations.
Kitagishi, Y., A. Nakanishi, Y. Ogura, and S. Matsuda. 2014. “Dietary
Regulation of PI3K/AKT/GSK- 3β Pathway in A lzheimer's Disease.”
Alzheimer's Research & Therapy 6: 1–7.
Korsunsky, I., N. Millard, J. Fan, et al. 2019. “Fast, Sensitive and
Accurate Integration of Single- Cell Data With Harmony.” Nature
Methods 16, no. 12: 1289–1296.
Krizhevsky, A., I. Sutskever, and G. E. Hinton. 2012. “ImageNet
Classification With Deep Convolutional Neural Networks.” Advances in
Neural Information Processing Systems 60: 84–90.
Krshnan, L., M. L. van de Weijer, and P. Carvalho. 2022 . “Endoplasmic
Reticulum–Associated Protein Degradation.” Cold Spring Harbor
Perspectives in Biology 14, no. 12: a041247.
Li, H., C. R. Brouwer, and W. Luo. 2022. “A Universal Deep Neural
Network for In- Depth Cleaning of Single- Cell RNA- Seq Data.” Nature
Communications 13, no. 1: 1901.
Li, R., W. Chen, M. Li, etal. 2023. “LensAge Index as a Deep Learning-
Based Biological Age for Self- Monitoring the Risks of Age- Related
Diseases and Mortality.Nature Communications 14, no. 1: 7126.
Li, S., Y. Zhao, R. Varma, etal. 2020. “P yTorch Distributed: Experiences
on Accelerating Data Parallel Training.” Proceedings of the V LDB
Endowment 13, no. 12: 30 05–3018.
Liu, J., D. Zhang, D. Yu, M. Ren, and J. Xu. 2021. “Machine Learning
Powered Ellipsometry.” Light: Science & Applications 10, no. 1: 55.
Liu, X., Z. Liu, Z. Wu, et al. 2023. “Resurrection of Endogenous
Retroviruses During Aging Reinforces Senescence.Cell 186, no. 2:
287–304.
López- Otín, C., M. A. Blasco, L . Partridge, M. Serrano, and G. Kroemer.
2023. “Hallmarks of Aging: An Expanding Universe.” Cell 186, no. 2:
243–278.
Lu, X. 2023. “Regulation of Endogenous Retroviruses in Murine
Embryonic Stem Cells and Early Embryos.” Journal of Molecular Cell
Biology 15, no. 8: mjad052.
Lundberg, S. M ., and S. I. Lee. 2017. “A Unified A pproach to Interpreting
Model Predictions.” Advances in Neural Information Processing Systems
30: 4765 –4774.
Ma, J., X. Xu, M. Li, et al. 2021. “Predictive Models of Aging of the
Human Eye Based on Ocular Anterior Segment Morphology.” Journal
of Biomedical Informatics 120: 103855.
Manjón, J. V., P. Coupé, and A. Buades. 2015. “MRI Noise Estimation
and Denoising Using Non- Local PCA.” Medical Image Analysis 22, no.
1: 35 47.
Mao, S., J. Su, L. Wang, X. C. Bo, C. Li, and H. Chen. 2023. “A
Transcriptome- Based Single- Cell Biological Age Model and Resource for
Tissue- Specific Aging Measures.” Genome Research 33, no. 8: 1381–1394.
Meyer, D. H., and B. Schumacher. 2021. “BiT Age: A Transcriptome-
Based Aging Clock Near the Theoretical Limit of Accuracy.” Aging Cell
20, no. 3: e13320.
Mishra , S., I. Beheshti, a nd P. Khan na. 2023. “A Review of Neu roimaging-
Driven Brain Age Estimation for Identification of Brain Disorders
and Health Conditions.” IEEE Reviews in Biomedical Engineering 16:
371–385.
Mohamadi, S., and D. A. Adjeroh. 2021. “An Information- Theoretic
Framework for Identif ying Age- Related Genes Using Human Dermal
Fibroblast Transcriptome Data.” In IEEE International Conference on
Bioinformatics and Biomedicine, 2294–2300. Cambridge, MA: IEEE .
Mohamadi, S., N. M. Nasrabadi, G. Doretto, and D. A. Adjeroh. 2021.
“Human Age Estimation From Gene Expression Data Using Artificial
Neural Networks.” In IEEE International Conference on Bioinformatics
and Biomedicine, 3492–3497. Cambridge, MA: IEEE.
Moutsopoulos, I., L . Maischak, E. Lauzikaite, et al. 2021. “noisyR:
Enhancing Biological Signal in Sequencing Datasets by Characterizing
Random Technical Noise.” Nucleic Acids Research 49, no. 14: e83.
Nikoletopoulou, V., E. Kyriakakis, and N. Tavernarakis. 2014. “Cellular
and Molecular Longevity Pathways: The Old and the New.” Trends in
Endocrinology and Metabolism 25, no. 4: 212–223.
Noothout, J. M., B. D. De Vos, J. M. Wolterink, et al. 2022. “Deep
Learning- Based Regression and Cla ssification for Automatic Landmark
Localization in Medical Images.” IEEE Transactions on Medical
Imaging 39, no. 12: 4011–4022.
Pappireddi, N. , L. Martin , and M. Wühr. 2019. “A Review on Quantitat ive
Multiplexed Proteomics.” Chembiochem 20, no. 10: 1210–1224.
Paszke, A., S. Gross, F. Massa, et al. 2019. “PyTorch: An Imperative
Style, High- Performance Deep Learning Library.” Advances in Neural
Information Processing Systems 32: 8024–8035.
Pedregosa, F., G. Varoquaux, A. Gramfort, et al. 2011. “Scikit- Learn:
Machine Learning in Python.” Journal of Machine Learning Research
12: 2825–2830.
Qi, H., Y. Wu, W. Zhang, N. Yu, X. Lu, and J. Liu. 202 4. “The Syntaxin-
Binding Protein STXBP5 Regulates Progerin Expression.” Scientific
Reports 14, no. 1: 23376.
Reya, T., S. J. Morrison, M. F. Clarke, and I. L. Weissman. 2011. “Stem
Cells, Cancer, and Cancer Stem Cells.” Nature 414, no. 6859: 105–111.
Rutledge, J., H. Oh, and T. Wyss- Coray. 2022. “Measuring Biological
Age Using Omics Data.” Nature Reviews Genetics 23, no. 12: 715–727.
Salehinejad, H., E. Colak, T. Dowdell, J. Barfett, and S. Valaee. 2018.
“Synthesizing Chest X- Ray Pathology for Training Deep Convolutional
Neural Networks.” IEEE Transactions on Medical Imaging 38, no. 5:
1197–1206.
18 of 18 Aging Cell, 2025
Savcisens , G., T. Eliassi- Rad, L. K. H ansen, etal. 2023. “Using S equences
of Life- Events to Predict Human Lives.” Nature Computational Science
4, no. 1: 43–56.
Selvaraju, R . R., M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D.
Batra. 2017. “Grad- Cam: Visual Explanations From Deep Networks via
Gradient- Based Localization.” In Proceedings of the IEEE International
Conference on Computer Vision, 618–626. Cambridge, MA: IEEE.
Sherman, B. T., M. Hao, J. Qiu, etal. 2022. “DAVID: A Web Server for
Functional Enrichment Analysis and Functional Annotation of Gene
Lists (2021 Update).” Nucleic Acids Research 50, no. W1: W216–W221.
Shokhirev, M. N., and A. A. Johnson. 2021. “Modeling the Human
Aging Transcriptome Across Tissues, Health Status, and Sex.” Aging
Cell 20, no. 1: e13280.
Shwartz- Ziv, R., and A. Armon. 2022. “Tabular Data: Deep Learning Is
Not all You Need.” Information Fusion 81: 84 –90.
Srivastava, N., G. Hinton, A. Krizhevsky, I. Sutskever, and R.
Salakhutdinov. 2014. “Dropout: A Simple Way to Prevent Neural
Networks From Overfitting.” Journal of Machine Learning Research 15,
no. 1: 1929–1958.
Suhre, K., M. I. McCarthy, and J. M. Schwenk. 2021. “Genetics Meets
Proteomics: Perspectives for Large Population- Based Studies.” Nature
Reviews Genetics 22, no. 1: 19–37.
Suryadevara, V., A. D. Hudgins, A. Rajesh, et al. 2024. “SenNet
Recommendations for Detecting Senescent Cells in Different Tissues.”
Nature Reviews Molecular Cell Biolog y 25: 1–23.
Szklarczyk, D., R. Kirsch, M. Koutrouli, et al. 2023. “The STR ING
Database i n 2023: Protein –Protein Ass ociation Networks and F unctional
Enrichment Analyses for any Sequenced Genome of Interest.” Nucleic
Acids Research 51, no. D1: D638–D646.
Wu, Y., J. Liu, Y. Wang, S. Gibson, M. Osadchy, and Y. Fang. 2024.
“Reconstructing Randomly Masked Spectra Helps DNNs Identify
Discriminant Wavenumbers.” IEEE Transactions on Pattern Analysis
and Machine Intelligence 46, no. 5: 3845–3861.
Xu, H., C. Caramanis, and S. Mannor. 2009. “Robustness and
Regularization of Support Vector Machines.” Journal of Machine
Learning Research 10, no. 7: 1485–1510.
Yuksel, S. E., J. N. Wilson, and P. D. Gader. 2012. “Twenty Years of
Mixture of Experts.” IEEE Transactions on Neural Networks and
Learning Systems 23, no. 8: 1177–1193.
Zhang, L ., X. Wang, D. Yang, etal. 2020. “Generalizing Deep Learning
for Medical Image Segmentation to Unseen Domains via Deep Stacked
Tran sfor mation.” IEEE Transactions on Medical Imaging 39, no. 7:
2531–2540.
Zhang, Y., R. Wilson, J. Heiss, etal. 2017. “DNA Methylation Signatures
in Peripheral Blood Strongly Predict All- Cause Mortality.” Nature
Communications 8, no. 1: 14617.
Zhao, Y., H. Cai, Z. Zhang, J. Tang, and Y. Li. 2021. “Learning
Interpretable Cellular and Gene Signature Embeddings From Single-
Cell Transcriptomic Data.” Nature Communications 12, no. 1: 5261.
Supporting Information
Additional supporting information can be found online in the
Supporting Information section.
ResearchGate has not been able to resolve any citations for this publication.
Article
Full-text available
Hutchinson–Gilfor progeria syndrome (HGPS) is caused by a mutation in Lamin A resulting in the production of a protein called progerin. The accumulation of progerin induces inflammation, cellular senescence and activation of the P53 pathway. In this study, through public dataset analysis, we identified Syntaxin Binding Protein 5 (STXBP5) as an influencing factor of progerin expression. STXBP5 overexpression accelerated the onset of senescence, while STXBP5 deletion suppressed progerin expression, delayed senility, and decreased the expression of senescence-related factors. STXBP5 and progerin have synergistic effects and a protein-protein interaction. Through bioinformatics analysis, we found that STXBP5 affects ageing-related signalling pathways such as the mitogen-activated protein kinase (MAPK) pathway, the hippo pathway and the interleukin 17 (IL17) signalling pathway in progerin-expressing cells. In addition, STXBP5 overexpression induced changes in transposable elements (TEs), such as the human endogenous retrovirus H internal coding sequence (HERVH-int) changes. Our protein coimmunoprecipitation (Co-IP) results indicated that STXBP5 bound directly to progerin. Therefore, decreasing STXBP5 expression is a potential new therapeutic strategy for treating ageing-related phenotypes in patients with HGPS.
Article
Full-text available
Once considered a tissue culture-specific phenomenon, cellular senescence has now been linked to various biological processes with both beneficial and detrimental roles in humans, rodents and other species. Much of our understanding of senescent cell biology still originates from tissue culture studies, where each cell in the culture is driven to an irreversible cell cycle arrest. By contrast, in tissues, these cells are relatively rare and difficult to characterize, and it is now established that fully differentiated, postmitotic cells can also acquire a senescence phenotype. The SenNet Biomarkers Working Group was formed to provide recommendations for the use of cellular senescence markers to identify and characterize senescent cells in tissues. Here, we provide recommendations for detecting senescent cells in different tissues based on a comprehensive analysis of existing literature reporting senescence markers in 14 tissues in mice and humans. We discuss some of the recent advances in detecting and characterizing cellular senescence, including molecular senescence signatures and morphological features, and the use of circulating markers. We aim for this work to be a valuable resource for both seasoned investigators in senescence-related studies and newcomers to the field.
Article
Full-text available
Here we represent human lives in a way that shares structural similarity to language, and we exploit this similarity to adapt natural language processing techniques to examine the evolution and predictability of human lives based on detailed event sequences. We do this by drawing on a comprehensive registry dataset, which is available for Denmark across several years, and that includes information about life-events related to health, education, occupation, income, address and working hours, recorded with day-to-day resolution. We create embeddings of life-events in a single vector space, showing that this embedding space is robust and highly structured. Our models allow us to predict diverse outcomes ranging from early mortality to personality nuances, outperforming state-of-the-art models by a wide margin. Using methods for interpreting deep learning models, we probe the algorithm to understand the factors that enable our predictions. Our framework allows researchers to discover potential mechanisms that impact life outcomes as well as the associated possibilities for personalized interventions.
Article
Full-text available
Age is closely related to human health and disease risks. However, chronologically defined age often disagrees with biological age, primarily due to genetic and environmental variables. Identifying effective indicators for biological age in clinical practice and self-monitoring is important but currently lacking. The human lens accumulates age-related changes that are amenable to rapid and objective assessment. Here, using lens photographs from 20 to 96-year-olds, we develop LensAge to reflect lens aging via deep learning. LensAge is closely correlated with chronological age of relatively healthy individuals (R2 > 0.80, mean absolute errors of 4.25 to 4.82 years). Among the general population, we calculate the LensAge index by contrasting LensAge and chronological age to reflect the aging rate relative to peers. The LensAge index effectively reveals the risks of age-related eye and systemic disease occurrence, as well as all-cause mortality. It outperforms chronological age in reflecting age-related disease risks (p < 0.001). More importantly, our models can conveniently work based on smartphone photographs, suggesting suitability for routine self-examination of aging status. Overall, our study demonstrates that the LensAge index may serve as an ideal quantitative indicator for clinically assessing and self-monitoring biological age in humans.
Article
Full-text available
The incorporation of machine learning methods into proteomics workflows improves the identification of disease-relevant biomarkers and biological pathways. However, machine learning models, such as deep neural networks, typically suffer from lack of interpretability. Here, we present a deep learning approach to combine biological pathway analysis and biomarker identification to increase the interpretability of proteomics experiments. Our approach integrates a priori knowledge of the relationships between proteins and biological pathways and biological processes into sparse neural networks to create biologically informed neural networks. We employ these networks to differentiate between clinical subphenotypes of septic acute kidney injury and COVID-19, as well as acute respiratory distress syndrome of different aetiologies. To gain biological insight into the complex syndromes, we utilize feature attribution-methods to introspect the networks for the identification of proteins and pathways important for distinguishing between subtypes. The algorithms are implemented in a freely available open source Python-package (https://github.com/InfectionMedicineProteomics/BINN).
Article
Full-text available
Endogenous retroviruses (ERVs) are an important component of the transposable elements, which constitute ∼50% of the mammalian genome. They exhibit dynamic expression during early embryonic development and are engaged in numerous biological processes. Therefore, their expression must be closely monitored by cells. Studies on the expression regulation of ERVs have been focused on mouse embryonic stem cells (ESCs) and early embryonic development. This review encapsulates recent advances in the modulation of ERVs in mouse ESCs and preimplantation embryos. The review first touches on the classification, expression, and functions of ERVs in mouse ESCs and early embryos. Next, the review mainly discusses the ERV modulation strategies from the perspectives of transcription, epigenetic modification, nucleosome/chromatin assembly, and post-transcriptional control.
Article
Full-text available
Accurately measuring biological age is crucial for improving healthcare for the elderly population. However, the complexity of aging biology poses challenges in how to robustly estimate aging and in how to interpret the biological significance of the traits used for estimation. Here we present SCALE, a statistical pipeline that quantifies biological aging in different tissues using explainable features learned from literature and single-cell transcriptomic data. Applying SCALE to the "Mouse Aging Cell Atlas" ( Tabula Muris Senis ) data, we identified tissue-level transcriptomic aging programs for over 20 murine tissues and created a multi-tissue resource of mouse quantitative aging-associated genes. We observe that SCALE correlates well with other age indicators, such as the accumulation of somatic mutations, and can distinguish subtle differences in aging even in cells of the same chronological age. We further compared SCALE with other transcriptomic and methylation ‘clocks’ in data of aging muscle stem cells, Alzheimer's disease, and heterochronic parabiosis. Our results confirm that SCALE is more generalizable and reliable in assessing biological aging in aging-related diseases and rejuvenating interventions. Overall, SCALE represents a valuable advancement in our ability to measure aging accurately, robustly, and interpretably in single cells.
Article
Dilated cardiomyopathy (DCM) is a complex disease with multiple causes and various pathogenic mechanisms. Despite improvements in the prognosis of patients with DCM in the past decade, this condition remains a leading cause of heart failure and premature death. Conventional treatment for DCM is based on the foundational therapies for heart failure with reduced ejection fraction. However, increasingly, attention is being directed towards individualized treatments and precision medicine. The ability to confirm genetic causality is gradually being complemented by an increased understanding of genotype-phenotype correlations. Non-genetic factors also influence the onset of DCM, and growing evidence links genetic background with concomitant non-genetic triggers or precipitating factors, increasing the extreme complexity of the pathophysiology of DCM. This Review covers the spectrum of pathophysiological mechanisms in DCM, from monogenic causes to the coexistence of genetic abnormalities and triggering environmental factors (the 'two-hit' hypothesis). The roles of common genetic variants in the general population and of gene modifiers in disease onset and progression are also discussed. Finally, areas for future research are highlighted, particularly novel therapies, such as small molecules, RNA and gene therapy, and measures for the prevention of arrhythmic death.
Article
Until recently, the question of the effective inductive bias of deep models on tabular data has remained unanswered. This paper investigates the hypothesis that arithmetic feature interaction is necessary for deep tabular learning. To test this point, we create a synthetic tabular dataset with a mild feature interaction assumption and examine a modified transformer architecture enabling arithmetical feature interactions, referred to as AMFormer. Results show that AMFormer outperforms strong counterparts in fine-grained tabular data modeling, data efficiency in training, and generalization. This is attributed to its parallel additive and multiplicative attention operators and prompt-based optimization, which facilitate the separation of tabular samples in an extended space with arithmetically-engineered features. Our extensive experiments on real-world data also validate the consistent effectiveness, efficiency, and rationale of AMFormer, suggesting it has established a strong inductive bias for deep learning on tabular data. Code is available at https://github.com/aigc-apps/AMFormer.
Article
Nondestructive detection methods, based on vibrational spectroscopy, are vitally important in a wide range of applications including industrial chemistry, pharmacy and national defense. Recently, deep learning has been introduced into vibrational spectroscopy showing great potential. Different from images, text, etc. that offer large labeled data sets, vibrational spectroscopic data is very limited, which requires novel concepts beyond transfer and meta learning. To tackle this, we propose a task-enhanced augmentation network (TeaNet). The key component of TeaNet is a reconstruction module that inputs randomly masked spectra and outputs reconstructed samples that are similar to the original ones, but include additional variations learned from the domain. These augmented samples are used to train the classification model. The reconstruction and prediction parts are trained simultaneously, end-to-end with back-propagation. Results on both synthetic and real-world datasets verified the superiority of the proposed method. In the most difficult synthetic scenarios TeaNet outperformed CNN by 17%. We visualized and analysed the neuron responses of TeaNet and CNN, and found that TeaNet's ability to identify discriminant wavenumbers was excellent compared to CNN. Our approach is general and can be easily adapted to other domains, offering a solution to more accurate and interpretable few-shot learning.