Sunteți pe pagina 1din 4

TIPS 1620 No.

of Pages 4

Trends in Pharmacological Sciences

Series: Artificial Intelligence


Science & Society
multiple applications, including strategy feature with age, the combination of
Deep Aging Clocks: The games as well as image and text many features can be very predictive. A
Emergence of AI-Based recognition. In healthcare, DL systems human can guess with reasonable
Biomarkers of Aging outperformed human dermatologists, accuracy the age of another human using
ophthalmologists, and radiologists in low- and high-resolution imaging data,
and Longevity various tasks. DL also demonstrated sig- movement patterns, or even scent and
nificant improvement over conventional touch.
Alex Zhavoronkov 1,2,3,* and
ML methods in biomedical data analysis
Polina Mamoshina1,4
[1,2]. Cross-Species Aging Research
Surprisingly, many patterns are similar in
Revolution in Biomarkers of Aging other species and can be used for cross-
First published in 2016, predictors
During this same period of DL progress, species analysis, and transfer learning
of chronological and biological
aging research has also experienced a re- techniques can be very helpful for re-
age developed using deep learning naissance, and new breakthroughs are search. For example, a human who is
(DL) are rapidly gaining popularity rapidly emerging. Multiple data types can shown a macaque for the first time in his
in the aging research community. be used to predict age and associate the or her life is often able to classify it correctly
These deep aging clocks can be prediction with mortality, disease, general into one of three age brackets: young,
used in a broad range of applica- wellbeing, or other biological processes in- middle-aged, or old. In ML this technique
tions in the pharmaceutical indus- cluding methylation, gene expression, is called 'zero-shot learning'. After seeing
try, spanning target identification, microbiome, and imaging data [3]. Since 100 macaques of varying ages, a human
drug discovery, data economics, the publication of the first multitissue is able to achieve much better accuracy.
and synthetic patient data genera- methylation aging clock by Steven This technique is called 'one-shot learn-
Horvath in 2013 [4], multiple methylation ing'. The same techniques may be appli-
tion. We provide here a brief
aging clocks and applications of these cable to learning biological processes by
overview of recent advances in
clocks in humans [5] and mice [6] were de- using age as a feature and then retraining
this important subset, or perhaps
veloped. Even though these clocks were on various diseases with few available
superset, of aging clocks that have developed using traditional ML ap- datasets or cross-species comparisons.
been developed using artificial proaches – notably linear regression with There are many challenges in classifying
intelligence (AI). regularization and the use of a limited aging as a disease in the traditional bio-
number of samples – the results suggest medical paradigm [7], but treating aging
Advances in AI
that gradual changes during aging can be as a process with N100 stages for the de-
tracked using various data types with rea- velopment of deep age predictors helps to
Recent advances in machine learning (ML)
sonable accuracy. capture a broad set of biological pro-
(Box 1), coupled with increases in compu-
tational power and availability of the large cesses in a holistic way. Although the
publicly available datasets, have led to a Age as a Universal Feature of Every study of the classical methylation aging
renaissance in AI. These advances have Organism clocks did not uncover many similarities
generated substantial investment and In 2015, advances in DL and aging bio- between mice and humans [8,9], the ap-
hype, and many data scientists and com- marker development began to converge. plication of AI and multiple data types
panies are exploiting the surge in AI hype AI researchers at Insilico Medicine recog- may help with cross-species research. A
for promotional purposes. This has sown nized that DL requires very large datasets, crowdfunded and crowdsourced project
confusion in the market and triggered crit- and that the most ubiquitous feature called MouseAgei attempts to develop a
icism from scientists working in the phar- among varied and incompatible biological photographic biomarker of aging in mice
maceutical industry, where approval in datasets is age. Age is a universal feature with the aim to apply transfer learning to
clinical trials is the ultimate measure of of every living organism and object on the other animals, and possibly to humans.
success. planet. It is also the most biologically rele- The effectiveness of this approach re-
vant feature because it is most strongly mains to be seen; however, there are
Most of the credible advances in the field correlated with mortality, a broad range clearly many features that are common in
have been in DL and reinforcement learn- of diseases, and remaining quality- rodents and even humans that can be ob-
ing (RL) (Box 1). Since 2013, DL systems adjusted life years (QALYs). Although it served by the naked eye, and DL may help
have surpassed human performance in can be difficult to correlate any individual to uncover these similarities.

Trends in Pharmacological Sciences, Month 2019, Vol. xx, No. xx 1


Trends in Pharmacological Sciences

Box 1. Applied AI Algorithms in digital skin analysis [12]. The deep pho-
Machine learning (ML) refers to data analysis tools that can extract dependencies from the data without being tographic aging clock, using only images
explicitly programmed, thereby providing an attractive alternative to other approaches in areas where few or no
prior data are available about those dependencies or where they are too complex.
of the corners of the eye, can predict the
age of an individual within an accuracy of
Deep machine learning or deep learning (DL) comprises a set of methods that rely on deep architectures with 1.9 years mean absolute error. Although
cascades of multiple layers, and include architectures such as deep neural networks (DNN), generative adver- photographic data are not the most bio-
sarial networks (GAN), deep reinforcement learning, and others.
logically relevant, many genetic and phe-
DNNs are models with multiple hidden layers between the input and output layers. The multilinearity of DNNs notypic disorders can be diagnosed from
combined with non-linear activation functions provides them with exceptional ability to extract complex depen- a picture. For many applications, images
dencies in the data and automatically select features that are most relevant to predictions. In the case of the age
are found to be more valuable than geno-
prediction, networks are trained using biological data as the input to predict age as accurately as possible.
mic data, and are even more valuable in
GANs are a type of a DL model that comprises discriminator and generator networks. A generator produces a combination with other data types. Photo-
candidate vector of synthetic data, and a discriminator networks check the vector validity. Such data gener- graphs are also among the most abundant
ation has been extensively explored for new pharmacological agents, and can also be used to generate syn-
data types, and results can be validated
thetic data for patients.
and interpreted instantaneously by
Reinforcement learning (RL) is a type of goal-oriented algorithm that is trained to attain a complex objective human experts, making images ideal for
over many steps. In case of drug discovery, such an objective could include the drug-likeness of molecules, proof-of-concept experiments.
their ease of synthesis, and other desired properties. RL algorithms could also be deep and have a multilay-
ered architecture.
Deep Transcriptomic Aging Clocks
Transcriptomic data are one of the most
abundant but variable types of data. The
Deep Aging Clocks The deep hematological aging clock study evolution of microarray and RNA sequenc-
Deep Blood Biochemistry and Cell-Count was extended to several million subject re- ing technology since 2000 has resulted in
Aging Clocks cords to evaluate the population specificity the production of millions of gene expres-
The realization that changes during aging and biological relevance of these clocks in sion datasets from multiple tissues, and
can be tracked has led to the search for multiple populations, as well as the associ- varying numbers of genes have been mea-
a biologically relevant data type that has ation of predicted age with mortality [11]. sured using different equipment in diverse
abundant historical datasets as well as a In this study the three DNNs were trained experimental settings. Despite high vari-
small number of highly variable but stan- on anonymized Korean, Canadian, and ability, transcriptomic data are among the
dardized features that can be easily Eastern European blood test samples an- most valuable types of data because they
anonymized. Using one of the broadest notated with age. Testing Korean and enable the identification of the genes
panels of routine blood tests performed Eastern European data with a DNN trained most implicated in specific diseases,
in multiple countries in a standardized on Canadian data revealed that Koreans such as cancer [13]. In 2018, the first
way, the first aging clock study utilizing on average appeared younger than transcriptomic aging clock developed
deep neural networks (DNNs) (Box 1) their chronological age, whereas Eastern using DL and other ML techniques based
was published by the laboratory of Europeans looked significantly older, thus on gene expression data from muscle
Zhavoronkov in 2016 [10]. The scientists demonstrating population differences. In tissue was published [14]. The work pre-
utilized over one million clinical blood addition, through testing on an indepen- sented several ideas on prioritizing specific
tests (blood biochemistry and cell count) dent dataset, researchers found that the genes as possible targets for pharmaceu-
to generate from routine screening tests people predicted to be older had higher tical intervention in sarcopenia and other
a dataset of over 60 000 reasonably mortality rates than those predicted to be muscle-wasting diseases.
healthy subjects annotated with sex and in line with their chronological age,
age. The proof-of-concept study demon- confirming the biological and suggesting Other Data Types
strated the basic application of evaluating clinical relevance of the clock. Wearable and mobile devices provide a
the relevant contributions of each vast amount of biologically relevant data.
simple feature to the accuracy of the pre- Deep Imaging Aging Clocks In 2018, age-associated changes in
dictor. The abundant blood biochemistry Photographic imaging, a highly accessible physical activity were studied for the first
data allowed comparison of the various and prevalent data type used in AI applica- time in context of age prediction using
ML models, and the DNNs clearly tions, has been explored by the research neural networks [15]. A DL-based model
outperformed in every test. team at Haut.AIii, a company specializing trained on activity-monitor data achieved

2 Trends in Pharmacological Sciences, Month 2019, Vol. xx, No. xx


Trends in Pharmacological Sciences

a relatively high accuracy in predicting anonymization of data while maintaining experiments. Deep aging clocks provide
age, but showed lower association with the most biologically relevant features. It a simple way for a pharmaceutical
mortality compared to a less accurate also enables the identification of potential company to evaluate which type of data
age-prediction model. To address this targets that drive aging and disease- is affected by a drug or intervention, lead-
lack of mortality association, the authors related processes [19]. ing to a clearer understanding of which
proposed a DL mortality predictor as a data are most important in a clinical trial.
tool for the identification of various Use of Deep Aging Clocks Age predictors can also help to evaluate
health risks. The intersection of recent advances in AI the quality of the data as well as their im-
and aging research yields many new pact both on the accuracy of the predictor
Generation of Synthetic Data as a tools and applications for the pharmaceu- and on the importance of specific features.
Tool for Target Identification in tical industry to exploit – at every step of Multimodal age predictors enable the inte-
Aging the R&D process as well as in personaliza- gration of previously incompatible data
In addition to expanding the scope of tion, marketing, and real-world evidence. types, such as dynamic wearable data
aging clocks, neural networks can be Over a dozen of these possible applica- combined with photographic and tissue-
used to generate synthetic data in large tions are summarized in Figure 1. We high- specific gene expression data.
volumes. Generative adversarial networks light a few of these below.
(GANs) (Box 1), a new ML technique first Moreover, aging research is a broad multi-
introduced by Ian Goodfellow in 2014 For pharmaceutical companies, multi- disciplinary field that converges with many
[16] and now commonly used in drug dis- modal age predictors (predictors that inte- other scientific disciplines directed at age-
covery [17,18], enable the generation of grate multiple data types) at the very related diseases [20]. Many interventions
biologically relevant synthetic data with minimum can provide deeper insights in immuno-oncology rely on the state of
specified conditions. Synthesizing new into biological data management. Not all the patient's immune system and general
patient data using GANs trained on data types are of equal biological impor- health. Aging clocks may be used to
millions of samples, using only age as a tance and relevance. Age predictors are track immunosenescence levels and iden-
generation condition, allows massive excellent tools for a broad range of tify new interventions designed to boost

Trends in Pharmacological Sciences

Figure 1. Prospective Applications of Deep Aging Clocks in the Pharmaceutical and Biotechnology Industries. Abbreviation: CS, stem cell.

Trends in Pharmacological Sciences, Month 2019, Vol. xx, No. xx 3


Trends in Pharmacological Sciences

the immune system in the elderly. For corporations, investment funds, and References
1. Mamoshina, P. et al. (2016) Applications of deep learning
companies specializing in vaccines and startups. After the newly announced in biomedicine. Mol. Pharm. 13, 1445–1454
looking for immediate revenue gains from call by famed Silicon Valley incubation 2. Aliper, A. et al. (2016) Deep learning applications for
predicting pharmacological properties of drugs and drug
AI, aging clocks can provide a way to program Y Combinator to provide repurposing using transcriptomic Data. Mol. Pharm. 13,
track response rates. If a meta-analysis of seed funding for extending longevityiii, 2524–2530
3. Xia, X. et al. (2017) Molecular and phenotypic biomarkers
clinical trials demonstrates that patients several startups in data-driven longevity of aging. F1000Res. 6, 860
predicted to be older than their chronolog- are sure to emerge in 2019 and beyond. 4. Hannum, G. et al. (2013) Genome-wide methylation pro-
files reveal quantitative views of human aging rates. Mol.
ical age respond better to an alternative Further, we see that large holding Cell 49, 359–367
dosage or vaccination protocol, then nec- companies in longevity, including 5. Field, A.E. et al. (2018) DNA methylation clocks in aging:
categories, causes, and consequences. Mol. Cell 71,
essary additional doses of the vaccine Juvenescence.AIiv, Longevity Vision Fundv, 882–895
may be sold. and Life Biosciencesvi, are also growing 6. Meer, M.V. et al. (2018) A whole lifespan mouse multi-
tissue DNA methylation clock. Elife 7, e40675
AI-powered longevity companies in an 7. Zhavoronkov, A. and Bhullar, B. (2015) Classifying aging
Digital Twin for a Patient
effort to find longevity interventions and as a disease in the context of ICD-11. Front. Genet. 6,
Multimodal aging clocks obscure the dif- 1–16
complementary biomarkers that can be 8. Stubbs, T.M. et al. (2017) Multi-tissue DNA methylation
ference between aging and disease sta- age predictor in mouse. Genome Biol. 18, 68
used to evaluate the effectiveness of such
tus, essentially turning the many aging 9. Petkovich, D.A. et al. (2017) Using DNA methylation profil-
therapies in a clinical setting. ing to evaluate biological age and longevity interventions.
clocks into a marker of the health status Cell Metab. 25, 954–960
of an individual. Because all living beings Disclaimer Statement 10. Putin, E. et al. (2016) Deep biomarkers of human aging:
application of deep neural networks to biomarker develop-
change over time, multimodal aging P.M. and A.Z. are employees of Insilico Medicine, a for- ment. Aging (Albany) 8, 1–13
clocks and clock ensembles trained on all profit longevity biotechnology company developing an 11. Mamoshina, P. et al. (2018) Population-specific bio-
markers of human aging: a big data study using South
accessible data types can act as a digital end-to-end target identification and drug discovery Korean, Canadian, and Eastern European patient popula-
twin for a patient. This likeness can be pipeline for a broad spectrum of age-related diseases. tions. J. Gerontol. A Biol. Sci. Med. Sci. 73, 1482–1490
12. Bobrov, E. et al. (2018) PhotoAgeClock: deep learning al-
moved forward and backward in time Insilico Medicine has applied for multiple patents in the gorithms for the development of non-invasive visual bio-
using GANs with multiple defined area of deep aging clocks and GAN-generated small markers of aging. Aging (Albany) 10, 3249–3259
molecules. 13. Sager, M. et al. (2015) Transcriptomics in cancer diagnos-
generation conditions, including lifestyle tics: developments in technology, clinical research
choices and interventions. These clocks Resources and commercialization. Expert Rev. Mol. Diagn. 15,
may also be embedded into field- i www.mouseage.org 1589–1603
14. Mamoshina, P. et al. (2018) Machine learning on human
trainable mobile devices that learn on the ii https://haut.ai/ muscle transcriptomic data for biomarker discovery and
tissue-specific drug target identification. Front. Genet. 9,
individual and help to maintain an optimal iii www.technologyreview.com/the-download/609963/ 242
biological age. y-combinator-will-give-you-1-million-to-try-to-cure- 15. Pyrkov, T.V. et al. (2018) Extracting biological age from
aging/ biomedical data via deep learning: too much of a good
iv thing? Sci. Rep. 8, 5210
Concluding Remarks www.juvenescence.ai/ 16. Goodfellow, I.J. et al. (2014) Generative adversarial net-
v
In this article we highlight the convergence of https://lvf.vc/ works. arXiv Published online June 10, 2014. https://
vi
http://lifebiosciences.com/ arxiv.org/abs/1406.2661
AI with aging research and review some of 17. Zhavoronkov, A. (2018) Artificial intelligence for drug dis-
the deep aging clocks that have been covery, biomarker development, and generation of novel
1
Insilico Medicine, Hong Kong Science and Technology Park, chemistry. Mol. Pharm. 15, 4311–4313
developed in the recent past. We also lay 18. Polykovskiy, D. et al. (2018) Entangled conditional adver-
Hong Kong, China
out the potential utilities of these clocks in 2
The Buck Institute for Research on Aging, Novato, CA, USA sarial autoencoder for de novo drug discovery. Mol.
3
The Biogerontology Research Foundation, London, UK Pharm. 15, 4398–4405
the pharmaceutical industry. In the coming 4 19. Zhavoronkov, A. et al. (2018) Artificial intelligence for aging
Computer Science Department, University of Oxford, Oxford, UK
years we expect the convergence of AI and longevity research: recent advances and perspec-
*Correspondence: tives. Ageing Res. Rev. 49, 49–66
and aging research to accelerate, given 20. Zhavoronkov, A. and Cantor, C.R. (2011) Methods for
alex@insilico.com (A. Zhavoronkov).
the emergence of longevity biotechnology https://doi.org/10.1016/j.tips.2019.05.004
structuring scientific knowledge from many areas related
to aging research. PLoS One 6, e22597
as a standalone industry [21] and the 21. de Magalhães, J.P. et al. (2017) The business of anti-aging
many players who are entering the field, © 2019 The Author(s). Published by Elsevier Ltd. This is an open
access article under the CC BY-NC-ND license (http://
science. Trends Biotechnol. 35, 1062–1073

from universities and non-profits to large creativecommons.org/licenses/by-nc-nd/4.0/).

4 Trends in Pharmacological Sciences, Month 2019, Vol. xx, No. xx

S-ar putea să vă placă și