Science 28Feb2020

INSIGHTS | PERSPECTIVES

sciencemag.org SCIENCE

MEDICINE

Artificial intelligence in cancer therapy

Artificial intelligence can optimize cancer drug discovery, development, and administration

By Dean Ho

A

rtificial intelligence (AI) approaches
have the potential to affect several
facets of cancer therapy. These in-
clude drug discovery and devel-
opment and how these drugs are
clinically validated and ultimately
administered at the point of care, among
others. Currently, these processes are ex-
pensive and time-consuming. Moreover,
therapies often result in variable treatment
outcomes between patients. The conver-
gence of AI and cancer therapy has resulted
in multiple solutions to address these chal-
lenges. AI platforms ranging from machine
learning to neural networks can accelerate
drug discovery, harness biomarkers to ac-
curately match patients to clinical trials,
and truly personalize cancer therapy using
only a patient’s own data. These advances
are indicators that practice-changing can-
cer therapy empowered by AI may be on
the horizon.
AI is already making progress in acceler-
ating drug discovery and has successfully
predicted drug behavior by using genomic
and chemical data in lieu of large-scale
screening assays, which may accelerate
drug repurposing (that is, using drugs indi-
cated for one disease to treat other diseases)
( 1 ). Reinforcement learning, which uses re-
ward and punishment to train algorithms
to achieve a desired drug structure, success-
fully designed a new compound in just 21
days versus conventional timelines of ap-
proximately 1 year. Subsequently observed
pharmacokinetic properties suggested that
threshold drug exposure and efficacy could
be attained, supporting further evaluation
of the lead compound ( 2 ). In this study, the
generative tensorial reinforcement learning
(GENTRL) platform was trained by using a
dataset of chemical structures that target
the tyrosine kinase discoidin domain re-
ceptor 1 (DDR1), which is involved in the
progression of multiple cancers. Computer
simulations predicted lead compound and
receptor binding, aiming to minimize tar-
geting of other tyrosine kinases and en-
hance DDR1 targeting. Although additional
compound optimization is required, this is

an important step toward AI-accelerated

cancer drug discovery.

AI-enhanced drug discovery represents

one stage of a pipeline of processes needed

to optimize cancer therapy (see the figure).

During traditional drug development, ap-

proved and investigational compounds

are often codelivered in combination pre-

clinically and clinically to address multiple

drug targets, improving treatment efficacy.

Subsequent clinical dose escalation studies

identify the doses that achieve drug syn-

ergy, in which the drug combination has a

greater effect together than individually.

Unfortunately, issues such as off-target ef-

fects can preclude drug approval because

of unforeseen toxicity ( 3 ). Additionally,

well-designed compounds delivered at sub-

optimal doses may limit efficacy. Therefore,

optimized combination therapy design

simultaneously identifies the best drugs

for combination and doses that address

the right targets while minimizing toxic-

ity. Testing all possible drug combinations

at multiple doses for each drug is virtually

impossible. However, AI can overcome this

challenge by markedly reducing the number

of experiments needed to resolve drug and

dose parameters, optimizing combination

therapy development.

Achieving drug synergy is a key objective

of traditional approaches to design combina-

tion therapies to improve efficacy. However,

patient responses to combination therapy

are highly variable. Computational modeling

showed that effective combination therapy

can be realized without drug additivity or

synergy; a combination of drugs that act

independently with favorable efficacy may

improve treatment outcomes over synergy-

driven combinations ( 4 ). Modeling tumor

growth kinetics and drug sensitivity predic-

tions by using data from clinical trials that

recruited large cohorts of patients with dif-

ferent cancer types demonstrated that maxi-

mizing independent drug efficacy is a major

determinant of treatment response.

AI will play a vital role in designing drug

combinations without relying on synergy-

based modeling or predicted synergy be-

tween different drug targets and pathways.

This may markedly increase the pool of

drugs considered for treatment and identify

unexpected combinations that outperform

standards of care. When the dose of each

candidate drug is also considered, the pos-

sible drug and dose combinations are too

extensive for comprehensive validation.

However, AI can rapidly resolve large drug

and dose parameter spaces. For example,

the quadratic phenotypic optimization plat-

form (QPOP) uses quadratic relationships

represented by parabolas to visually corre-

late a set of inputs (e.g., drugs and doses)

with optimal outputs (e.g., preclinical tu-

mor reduction with minimal toxicity). This

correlation markedly reduces the number

of experiments and data needed to identify

the drugs and doses that optimize combina-

tion therapy design. Furthermore, QPOP is

agnostic to disease mechanisms, drug tar-

gets, and drug synergy.

The QPOP platform evaluated 14 chemo-

therapy drugs to treat multiple myeloma

(MM) with combination therapy ( 5 ). The

resulting drug combinations—such as

decitabine and mitomycin C, which were

unexpectedly and agonistically identified

by QPOP—markedly improved outcomes

in a MM mouse model compared with the

clinical standard-of-care drug combina-

tions ( 5 ). Importantly, neither decitabine

nor mitomycin c monotherapy mediated

efficacy alone, but their co-administration

optimally and synergistically reduced tu-

mor burden. Expanding QPOP to build

combinations of targeted therapies and

immunotherapies will be an important

next step toward treatment strategies be-

yond chemotherapy.

As oncology drug candidates move into

clinical validation, their regulatory ap-

proval rates have been reported to be as

low as 3.4% ( 6 ). Recent advances in trial

design have used biomarkers, such as ge-

nomic alterations, because of their poten-

tial as predictive indicators of treatment

response to stratify patient recruitment.

Including biomarkers in study recruitment

has improved patient outcomes compared

with traditional stratification information

such as pathology or responses to prior

treatments ( 3 , 7 ). Combining patient bio-

marker data and electronic health records

(EHRs) for AI analysis may further affect

trial outcomes ( 8 ). In the SYNERGY-AI

study (NCT03452774), virtual tumor

boards, which enable clinician teams to

provide remote treatment guidance and

diagnosis input, and EHRs pair oncol-

ogy patients with suitable clinical trials.

Multiple disease factors outlined by EHRs

The N.1 Institute for Health (N.1), Institute for Digital
Medicine (WisDM), Department of Biomedical Engineering,
and Department of Pharmacology, National University of
Singapore, Singapore. Email: [email protected]

982 28 FEBRUARY 2020 • VOL 367 ISSUE 6481

Published by AAAS