Science 28Feb2020

SCIENCE sciencemag.org

GRAPHIC: A. KITTERMAN/

SCIENCE

will be correlated with progression-free

survival and overall survival to assess the

effectiveness of the trial-matching plat-

form. Although many types of data may be

useful for stratification, AI will ultimately

need both population-scale and individu-

alized data to ensure that patients given

therapies, including those designed by AI,

have a high likelihood of responding.

AI will also play a pivotal role in how

cancer therapy is administered. Maximum

tolerated dosing eliminates drug-sensitive

tumor cells. However, drug-resistant cells

can eventually cause treatment failure.

Game theory is being explored to address

this challenge, with dose-reduction algo-

rithms competing against the tumor to

prevent drug-resistant cells (which have

high energy costs) from outnumbering

drug-sensitive cells ( 9 , 10 ). This is known

as adaptive therapy and may prolong treat-

ment efficacy by maintaining threshold

drug-sensitive cell populations in a tumor

to combat drug-resistant cell prolifera-

tion. Adaptive therapy was recently used

to modify paclitaxel dosing to treat mouse

models of breast cancer. Specifically, an

adaptive therapy algorithm (AT-1) itera-

tively reduced paclitaxel dosing after ob-

served tumor size reductions. AT-1 was

compared with a fixed-dosage algorithm

that halted drug dosing after observed

tumor reduction (AT-2) and high-dose

standard therapy ( 11 ). The AT-1 algorithm

improved tumor control and survival com-

pared with AT-2 and standard therapy,

demonstrating that game theory–driven

dosing can improve treatment outcomes

over established therapies at high doses.

Adaptive therapy was translated into a

pilot study for prostate cancer patients on

abiraterone hormone therapy ( 12 ). The adap-

tive therapy cohort received, on average, 47%

of the standard abiraterone dose, and three

of the patients received less than 25% of the

conventional dose. At the time of reporting,

1 of 11 adaptive therapy participants expe-

rienced tumor progression. This led to an

estimated median time to progression, us-

ing the biomarker prostate-specific antigen

(PSA) and radiographic imaging, of no less

than 27 months in this cohort, which was su-

perior to 11.1 months (PSA) and 16.5 months

(radiographic) for patients on continuous

abiraterone therapy ( 12 ). To maximize the

benefits of this platform, personalizing adap-

tive therapy by use of each patient’s response

to therapy versus population-based dose ad-

justment rules will likely be needed.

To further personalize patient-specific

combination therapy administration with

AI, CURATE.AI, a neural network–derived

platform, modulated multidrug dosing

using only a patient’s data by means of a

second-order algebraic algorithm that dy-

namically related dosing to optimal tumor

reduction and safety at any treatment time

point ( 13 , 14 ). In a patient treated with

enzalutamide hormone therapy and an in-

vestigational bromodomain and extrater-

minal domain (BET) inhibitor, data such as

physician-guided dose variations and corre-

sponding amounts of PSA were used to rec-

ommend a 50% reduction of the BET inhibi-

tor dose to increase efficacy. Subsequent

dynamic dosing of both drugs resulted in a

durable response and halted tumor progres-

sion, which was confirmed by computed

tomography (CT) imaging ( 13 ). This study

successfully used AI to modulate experi-

mental therapy dosing, demonstrating that

dosage guidance, without requiring big data

and complex genetic information, markedly

enhanced treatment efficacy compared with

fixed- and high-dose chemotherapy. Future

studies will need to evaluate whether this

platform can be implemented with addi-

tional classes of disease markers. Also, be-

cause of the individualized nature of this

platform, it will need to be further evalu-

ated in larger patient cohorts. As such, it is

being tested in a clinical trial that involves

a large pool of patients with hematological

malignancies (NCT03759093).

AI represents a gateway to the next fron-

tier of cancer therapy, reconciling extraor-

dinary amounts and different types of data

into actionable therapy. Among its barriers

to deployment are its siloed use in narrow

segments of the cancer therapy workflow.

For example, AI-optimized compounds

that are suboptimally combined with other

therapies or dosed incorrectly are unlikely

to substantially improve patient outcomes.

Overcoming this challenge in oncology will

require the seamless implementation of AI

across the spectrum of discovery, develop-

ment, and administration. Its potential

downstream applications include increas-

ing the resolution of personalized care by

tailoring bespoke regimens that integrate

multiple therapeutic strategies. For exam-

ple, AI-optimized radiation therapy dosing,

which maintained robust tumor size con-

trol, can potentially be combined with AI-

driven drug administration ( 15 ). Ultimately,

comprehensively adapting AI into clinical

oncology practice may improve drug acces-

sibility and reduce health-care costs. As AI

continues to be validated and a path toward

widespread practice is identified, its poten-

tial to redefine the clinical standards of can-

cer therapy is becoming evident. j

REFERENCES AND NOTES

1. M. P. Menden et al., PLOS ONE 8 , e61318 (2013).


2. A. Zhavoronkov et al., Nat. Biotechnol. 37 , 1038 (2019).


3. A. Lin et al., Sci. Transl. Med. 11 , eaaw8412 (2019).


4. A. C. Palmer, P. K. Sorger, Cell 171 , 1678 (2017).


5. M. B. M. A. Rashid et al., Sci. Transl. Med. 10 , eaan0941
   (2018).


6. C. H. Wong et al., Biostatistics 20 , 273 (2019).


7. S. Harrer et al., Trends Pharmacol. Sci. 40 , 577 (2019).


8. A. Rajkomar et al., NPJ Digit. Med. 1 , 18 (2018).


9. J. Chmielecki et al., Sci. Transl. Med. 3 , 90ra59 (2011).


10. C. Jedeszko et al., Sci. Transl. Med. 7 , 282ra250 (2015).


11. P. M. Enriquez-Navas et al., Sci. Transl. Med. 8 , 327ra24
    (2016).


12. J. Zhang et al., Nat. Commun. 8 , 1816 (2017).


13. A. J. Pantuck et al., Adv. Ther. 1 , 1800104 (2018).


14. A. Zarrinpar et al., Sci. Transl. Med. 8 , 333ra49 (2016).


15. B. Lou et al., Lancet Digit. Health 1 , e136 (2019).
    ACKNOWLEDGMENTS
    D.H. holds pending patents on AI-based personalized
    medicine. D.H. acknowledges support from AI Singapore
    (NRF), the National Medical Research Council, and the
    Ministry of Education.
    10.1126/science.aaz3023

Tar g e t

identifcation

Combination

therapy design

Patient-therapy

matching

Dose

Time

Tar g e t

Discovery

Lead

Lead generation

and optimization

Preclinical

development

Clinical trials

phase I–III

Personalized

cancer therapy

Opportunities

Minimize of-target efects and toxicity

Enhance drug exposure

Challenges

Identifying optimal targets

Properly validating AI-designed drugs

Development

Opportunities

Optimize drug and dose selection

Match patients to therapies and trials

Challenges

Improving trial outcomes

Stratifcation with the right patient data

Administration

Opportunities

Sustained dose optimization

Overcoming resistance with

game theory

Challenges

More clinical validation needed

Use in more cancer types

28 FEBRUARY 2020 • VOL 367 ISSUE 6481 983

Improving multiple aspects of cancer therapy

Cancer therapy involves different stages, including drug discovery, development, and administration.

Artificial intelligence (AI) is poised to benefit each stage but is also confronted by challenges that, when

overcome, may lead to practice-changing cancer treatment.

Published by AAAS