Life Sciences Practice

How artificial
intelligence can power
clinical development
Artificial intelligence is accelerating drug discovery. If clinical
development fails to keep pace, the benefits to patients will be delayed.
This article is a collaborative effort by Chris Anagnostopoulos, David Champagne, Thomas Devenyns,
Alex Devereson, and Heikki Tarkkila, representing views from McKinsey’s Life Sciences Practice.

November 2023
Clinical development and the randomized on average, 40 percent more assets per indication
clinical trial (RCT) remain largely unaffected than in 2006,2 raising the need for greater
by the unprecedented wave of innovation in differentiation. And clinicians and patients alike
pharmaceutical R&D fueled by developments in need more evidence to make decisions about the
artificial intelligence (AI), including generative AI best available treatment at a given time.
(gen AI) and foundational models. However, they are
under pressure from the rise of precision medicine All this comes as researchers accelerate drug
and a more competitive development landscape. discovery by making greater use of AI and gen AI,
So far, the adoption of AI for clinical development such as DeepMind’s AI-enabled platform AlphaFold,
has emphasized operational excellence and which can predict the 3-D structure of molecules3
acceleration, but advances in scientific AI have leading to a pipeline of more and better-designed
made this the time to leverage modern analytical preclinical assets with a better target validation.
tools and novel sources of data to design more Companies are also striving to improve their
precise, efficient trials with greater success rates. operational processes with a view to speeding up
the time it takes to apply for a first-in-human study.
The introduction of RCTs in the mid-20th century But if clinical development fails to keep pace, the
ushered in the modern era of evidence-based drug benefits to patients of faster drug discovery will
development. Their rigidity and simplicity were inevitably be delayed, and delivering on the promise
welcome defenses against an overreliance on of AI-enabled acceleration may flounder.
anecdotes and case studies and have unarguably
served patients well, supporting the introduction While the RCT will remain a central pillar of clinical
of countless safe, effective therapies. Yet RCTs are development, the tide is changing. Regulators have
starting to be seen as bottlenecks that lengthen the recently issued guidance on appropriate use of real-
time for therapies to gain approval and can increase world data (RWD).4 The healthcare data ecosystem
costs.1 Meanwhile, as gen AI breaks milestone after is booming, leveraging privacy-respecting
milestone, patients eagerly await the translation technology to make patient-level data safely
of such unprecedented technological progress available for research. And evidence generation
into pharmaceutical R&D that could deliver faster from multiple data sources is now possible using
access to better treatments. causal machine learning, which aims to distinguish
correlation from causation via a combination of
In addition to rising expectations about timelines, biostatistics with machine learning (ML) and,
clinical development is also facing increasing increasingly, gen AI and foundational models.
demands to generate targeted and meaningful
data. As precision medicine becomes mainstream, Despite this level of promise, only a handful of
RCTs often must prove not only the general efficacy established companies are deploying AI and
of a treatment but also whether it will benefit a data-driven approaches systematically in their
specific segment of the patient population. The clinical development. The focus so far has been on
smaller that segment, the harder it can be to enroll improving operational excellence and increasing
enough patients in trials. In addition, the bar for acceleration rather than helping to inform trial
what constitutes a clinically meaningful treatment design strategy. The remainder of this white paper
effect is creeping higher to meet standards set dives deeper into the context, gives tangible
by regulators and payers as well as competitive example use cases and associated impact, and
pressure. In most therapeutic areas today, there are, identifies challenges companies face when

1
Donald A Berry, “The brave new world of clinical cancer research: Adaptive biomarker-driven trials integrating clinical practice with clinical
research,” Molecular Oncology, 2015, Volume 9, Issue 5; Michael Baumann et al., “How much does it cost to research and develop a new drug?
A systematic review and assessment,” PharmacoEconomics, 2021, Volume 39, Issue 11.
2
McKinsey research on data from EvaluatePharma, March 2019.
3
“AI in biopharma research: A time to focus and scale,” McKinsey, October 10, 2022.
4
See, for example, the US Food and Drug Administration’s real-world evidence activities responding to the 21st Century Cures Act.

How artificial intelligence can power clinical development 2

adopting AI for clinical development. Rather than Four use cases that demonstrate
staying at a high level, it aims to bring the topic to life the potential of AI
through relevant details and case examples. Companies that are deploying AI combined with
RWD are seeing impactful results. While RWD can
be valuable supporting evidence in health authority
Expanding adoption of AI and submission packages, the most successful
RWD for clinical development companies focus their use of AI and RWD to make
Today AI can draw upon growing volumes of RWD. better and more informed decisions to support the
Electronic health records and claims data from success of clinical development programs in every
certain geographies are widely available, and novel step, from asset and portfolio strategy to protocol
data sources—including biobanks, data omics and trial design:
panels, population-wide genomic studies, patient
registries, and imaging and digital pathology— — At the stage of defining an asset strategy, AI
are increasing in number and diversity. All these and RWD can reveal which indications are most
sources reflect a notable change in patients’ ability promising indications to pursue for novel assets.
to share their data for the purpose of advancing Several leading biopharma companies identified
research and treatments in privacy-respecting multiple new indications for existing assets in
ways, including technology that enables the this way, and an early-stage biotech used AI and
training of ML models in a manner that respects RWD to assess whether to shift its indication
patients’ privacy. selection strategy regarding a novel asset.

In addition, new tools can systematically capture — AI and RWD can support decisions about the
knowledge from unstructured data. Large target patient population of a clinical trial, with
language models (for example, BioGPT) are able subgroup discovery, refine trial eligibility criteria
to convert unstructured physician notes into high- and help remove patients that are highly unlikely
quality structured data. Similarly, these models can to benefit from the treatment, and shorten the
search the vast corpus of published literature and length of trials. One early-stage biotech could
identify connections between biological entities better characterize “super-progressors” (that is,
on a large scale, generating high-quality input patients whose disease was likely to progress
for knowledge graphs that better represent the faster within the time frame of a clinical trial) and
totality of available evidence on a given indication design a trial with similar expected benefits in a
across domains, including genes, targets, proteins, faster time frame.
pathways, and phenotypes.
— For decisions about portfolio strategy, AI and
Most pharma companies using AI and RWD RWD can help companies identify the right
in clinical development tend to do so only in combination of drugs for an indication or for
isolated use cases, and they rarely deploy both the right patients. One biopharma company
in combination. For example, they might use leveraged a prospective observational data
machine learning to select trial sites or to predict set to generate evidence supportive of earlier
patient enrollment rates. They might use RWD to positioning of its third-line treatment. Another
measure disease prevalence—the size of an eligible was able to identify “super-responders” for
population—to understand the natural course of a several drugs in its portfolio—insights that
disease among untreated patients, or to construct helped the company position its assets optimally
an external control arm for regulatory purposes. in a crowded indication.
While these use cases deliver value, a great deal
more is now possible. Similarly, while the level of — At the step of selecting and optimizing the end
funding and investment going into AI-enabled drug points of a clinical trial, AI and RWD can help a
discovery companies has tripled in the last five company identify patient attributes that closely
years, that same trend is not true for equivalent track the primary end point over time. One
companies in the clinical development space. biopharma company replaced a rare disease’s

How artificial intelligence can power clinical development 3

 existing end point, which was an infrequent journeys. Foundational models that treat medical
event, with end points that occurred with events as words and patient medical histories as
greater frequency or could be measured with documents can uncover the semantic similarity
blood tests. This cut the length of trials by 15 to of different events, including diagnoses. Each
30 percent. indication, from what is likely to be a sizable list of
those with biological proximity, is scored according
The following discussion looks in greater depth at to its similarity to one or more anchor references—
four detailed use cases, illustrating the wealth of perhaps the indication for which the asset was
information that AI can unlock. originally approved or, in the case of a novel asset,
one for which there is strong preclinical evidence of
Indication selection for asset strategy efficacy of the asset’s mechanism of action (MoA).
Selecting which indications to target with a specific
molecule is one of the most important decisions The scoring can also incorporate information
a biopharma company makes. This decision is from molecular knowledge graphs that show new
often informed by a combination of input from connections—for example, between entities such
key opinion leaders, literature reviews, omics as proteins or human biological pathways that
analysis (for example, genome-wide association have already been identified in literature or public
studies), RCT data, and competitor decisions. data. Exhibit 1 shows a potential outcome of this
Such decisions rarely are fully data driven, typically approach: it can reveal the indications most strongly
prove hard to integrate, and cover only part of the connected to each of the reference indications,
available evidence base, resulting in a subjective serving as an anchor into the existing evidence
and suboptimal synthesis. In contrast, strategies base. The indications with the most connections to
informed by RWD and AI are objective and these references are further prioritized by unmet
comprehensive in their use of multiple data sets and medical need, strategic fit, and technical feasibility,
foundational models built on clinical data continually resulting in an evidence-based indication selection.
redefining the art of the possible.
These analytical approaches can identify novel
For already-approved therapies, companies can use indications that can be rapidly validated via in vitro
RWD to understand the therapies’ likely efficacy on or animal models, can increase the confidence
alternative indications. One approach is examining in selecting indications with a high probability
the outcomes of patients who were prescribed the of success, and can derisk resource allocation
drug on a spontaneous basis by their physician; accordingly. The analysis provides a clearer
another is determining a drug’s average effect and more holistic evidence base for investors,
on patients exposed to the treatment because of shareholders, and R&D leaders, and it can reduce
incidental comorbidities. AI techniques can then the opportunity cost of blind alleys by helping new
extrapolate any findings to a set of patients whose treatments reach patients faster.
characteristics may be distributed differently, by
leveraging observed correlations between patient Subgroup discovery for trial design
characteristics and outcomes. Once an asset has been matched with an indication,
pharmaceutical companies pursue an all-comer
Where companies are looking to expand the clinical development and trial design strategy—
indication of an approved asset or identify for that is, one that includes all patients except those
a novel asset, RWD and AI can estimate the deemed high-risk based on input from key opinion
biological proximity of one indication to another leaders and conventional wisdom.5 As a result,
from a patient and clinical perspective—that is, patients unlikely to respond to a treatment may
whether the patient experiences similar symptoms, be included in the trial. A notable exception to this
comorbidities, lab characteristics, and treatment method is in precision oncology, where researchers

5
The use of the all-comer strategy may reflect a low level of confidence in existing hypotheses that currently seek to explain response
heterogeneity in patient subpopulations.

How artificial intelligence can power clinical development 4

Web <2023>
<AdvancedDigitalAnalyticsClinical>
Exhibit
Exhibit <1>1 of <3>

Indications resulting from the RWD AI pipeline can be visualized in relation to

the different reference indications.
How the indication-finding network chart works
The network chart displays the
Neurodermatitis top-ranking indications (small nodes)
resulting from the RWD AI pipeline for
each reference indication (large nodes).
Ulcerative colitis Only the connections are important;
the distances between nodes do not
convey any information. The more
Neurodermatitis connections a node has, the more it
has surfaced for different reference
indications.
Rheumatoid arthritis

Reference indication
Top-ranking indications

McKinsey & Company

use specific biomarkers discovered preclinically the trial was designed in a manner that focused on
(such as genetic mutations) to stratify patients patients without access to effective treatments.
according to their probability of progression or to
predict a patient’s response to different treatments. AI for subgroup discovery can also directly
This approach is revolutionizing oncology support the strategic trade-off between the
treatments, but even it can identify only a fraction of size of the eligible population and the level of
possible patient subpopulations. the treatment effect in a smaller population, in
the form of an efficient frontier (Exhibit 2). The
In contrast, AI and omics-rich RWD can examine different subpopulations that are marked as blue
thousands of genetic and/or phenotypic attributes on the exhibit are those for which it is impossible
to pinpoint the combinations most likely to influence to simultaneously increase their breadth without
prognostic or predictive scores, explain response reducing the drug’s expected effect. This approach
heterogeneity, and improve a trial’s technical widens the overlap between patients with access to
probability of success. Even in situations where the this new treatment and patients who are most likely
patient cohorts are modestly sized, foundational to benefit significantly from it.
models trained on broad RWD can be used as a
starting point for more bespoke modeling at the Subgroup discovery and comparative
indication level. efficacy for portfolio optimization
Not only can AI be used to discover subgroups to
One large biopharmaceutical company used RWD inform the design of a trial, it also can contribute
to remove likely nonresponders from a trial and to to the wider portfolio strategy. The product
reduce its expected duration by between 5 and pipelines and portfolios of many biopharmaceutical
10 percent without compromising its probability of companies contain multiple assets aimed to treat
success, which would allow the treatment to reach a similar set of patients, often because companies
patients faster. Another company leveraged hospital lack sufficient evidence to inform the portfolio
episode RWD to identify super-responders to its strategy differently, especially for new MoAs.
comparator arm for an asset in Phase I, ensuring However, the systematic analysis of all available
data sources can predict the response of different

How artificial intelligence can power clinical development 5

Web <2023>
<AdvancedDigitalAnalyticsClinical>
Exhibit <2> of <3>
Exhibit 2

Artificial intelligence can support selection of optimal subpopulations.

Illustrative examples

Subpopulation analysis Subpopulation scenarios

Nonoptimal subpopulation Optimal subpopulation
Expected
High Time to trial completion, months efficacy,¹ %

Efficient frontier
Inclusion 20 90
criterion A

Inclusion
Subpopulation 13 60
criterion B
size

Inclusion
10 55
criterion C

Inclusion
3 40
Low criterion D
Low High
Effect size
¹Plus or minus 5%–10%.

McKinsey & Company

patient subgroups (or disease endotypes), action mechanisms with the novel treatment. This
which could help companies hone their portfolio exercise can be repeated across multiple disease
strategy—in some cases, resulting in a double- endotypes or patient subgroups, identified through
digit percentage improvement in net present value. a combination of patient-level attributes (Exhibit 3).
This strategic stance can support targeted trials
that produce the type of precision evidence that End point optimization in clinical development
clinicians need to make the best possible treatment In many trials, there is no single, established end
decisions for their patients amid a proliferation of point. In others, it can be hard to know which of
novel assets. several would best measure the intended action
of the drug. Some end points may take a long time
In addition, several analytical approaches and data to manifest, leading to extended development
sources can be combined to determine the efficacy timelines. Others may detect progression that
of novel MoAs in the portfolio relative to each other is more reliable for some types of patients than
or to approved treatments of different patient for others or that may be particularly invasive,
subgroups. Enriching RWD with data from molecular increasing the trial’s burden on patients.
knowledge graphs can enable foundational-
model-powered representations of treatments that AI used on rich, patient-level longitudinal data sets
represent their associated biological pathways. can be deployed to identify patient attributes that
From this, the company can estimate the efficacy closely track the primary end point over time. Such
of a novel treatment by observing the outcomes of an application can control multiple confounding
approved treatments that share similar biological factors to ensure the association persists in future

How artificial intelligence can power clinical development 6

Web <2023>
<AdvancedDigitalAnalyticsClinical>
Exhibit
Exhibit 3of <3>
<3>

A systematic AI application can predict which subpopulations will respond to a

novel treatment.

Approach to understanding the efficacy of novel treatments

Step 1 Population
Discover subpopulations that
score differently on selected
end points for different
approved treatments. End
points of interest are Subpopulations
determined in terms of
response, safety, or mixed.

Step 2 Treatments Subpopulations Outcomes

Simulate outcomes of different
approved treatments for each A
subpopulation.

Step 3 Novel treatment Subpopulations Outcomes

Simulate outcomes of novel
treatments for each
subpopulation, using data
collected from observed
treatment outcomes and D
knowledge graphs to enable
foundational-model-powered
representations.

McKinsey & Company

trials where the patient population may differ, altogether novel, deeply hidden visual signatures
thereby establishing potential novel end points. of disease activity and severity with better
predictive characteristics.
Here, X-rays, CT and MRI scans, and other imaging
techniques are often valuable sources of data for In all cases, the association between novel end
monitoring biomarkers in a noninvasive manner points and desired patient outcomes such as
(as is the case with using MRIs to help detect survival or quality of life must persist even under
Alzheimer’s disease6). Foundational models using treatment. The clinical data and the information
medical imaging are increasingly able to extract present in RWD must be combined with a biological
imaging biomarkers curated by human expertise understanding of the mechanisms affected by the
from raw images at scale. They can even discover considered treatment.

6
Ruth Stephen et al., “Change in CAIDE dementia risk score and neuroimaging biomarkers during a 2-year multidomain lifestyle randomized
controlled trial: Results of a post-hoc subgroup analysis,” Journals of Gerontology: Series A, 2021, Volume 76, Issue 8.

How artificial intelligence can power clinical development 7

Tackling the organizational challenges the biopharmaceutical industry is not typically their
Inevitably, supplementing a century-long, RCT- first choice. Measures that can help include offering
dependent approach to clinical development with competitive remuneration packages, building an
a new, analytics-driven one using novel data poses innovative employer brand, and locating in key
organizational challenges. Here are some ways to talent hubs. So, too, can a focus on the industry’s
tackle them. value proposition—offering employees the chance
to help people lead longer and healthier lives.
Embed AI into clinical development
The governance processes and incentive Develop a targeted data strategy
structures in place in most biopharmaceutical With so much data becoming available, companies
companies encourage default solutions in clinical should consider formulating a detailed and iterative
development at the asset team level and beyond. data strategy for each disease area. The purchase
One consequence is the prevalence of all-comer of certain commercial data sets, such as electronic
trials, which target the broadest possible population health records, claims, and sometimes omics, can
without leveraging evidence about nonresponders be relatively straightforward, even in therapeutic
and super-responders—a method that can increase areas such as oncology, where biomarker-
trials’ duration and decrease their probability of rich data are required. But other types of data
success, resulting in longer waits for patients who sets—registries, biobanks, and clinical trials from
would benefit from the treatment. other pharma companies—can involve complex
negotiations and contracts, and companies will have
R&D leaders can help change that culture by to partner with a range of different data providers.
encouraging the use of analytical methods to
integrate all sources of available evidence when In addition, each data set is likely to be structured
making key decisions. Individual incentives can help. differently, with varying levels of quality and
A stronger step is to strengthen the development completeness. This means companies will likely
governance model to require that all key decisions need data curation and analytical capabilities
be supported by AI. for tasks such as combining different data sets
to plug gaps in evidence and meeting regulatory
Develop internal capabilities and talent expectations for the use of real-world data.
External providers can supply off-the-shelf AI
solutions to support some use cases. But with so Build scalable products
many new solutions emerging, no provider can yet AI’s power lies in its application across the clinical
support them all. Even if they could, companies development portfolio, not restricted to isolated use
would be unwise to become dependent upon them cases. Achieving this kind of scale can be hard, due
for critical strategic decisions. Providers with to limitations of the code base, data platforms, the
such rich intellectual property might eventually company’s data engineering capabilities, and the
become competitors. analytical stack needed to collect, combine, and
analyze data. As a result, pilots often remain just
Biopharmaceutical companies could therefore that—never reused or deployed against other assets
consider building their own AI capabilities and in development.
products. However, significant investment is
needed to acquire the relevant data sets; build up To counter this, analytical use cases should be built
the necessary data science, data engineering, and as products, meaning solutions built in a manner
business translator expertise; and establish agile that makes them easy to reuse and scale for
product development processes that ensure use different use cases, even at the proof-of-concept
cases deliver business value in a timely and cost- phase. The code base should be well annotated, for
efficient manner. example, and deploy common, reusable analytical
components that are either developed internally,
Companies will also have to work hard to attract and open-source, or bought off the shelf. Shared ML
retain top data scientists and data engineers, as development standards and coding frameworks

How artificial intelligence can power clinical development 8

 can help by speeding up development work across process, from formulating asset strategy to
Find more content like this on the
different use cases because data scientists and designing the protocol and trial planning. For
McKinsey Insights App
engineers are familiar with them. Companies also patients, the results can speed up access to new
need a mature and flexible ML operations stack— treatments, thanks to an accelerated development
that is, the operational components needed to timeline, and make treatments more likely to elicit
streamline the process of taking ML models to the strongest response.
production and maintaining and monitoring them.
High-impact solutions are already being
implemented, and more will follow. Companies can
benefit from beginning to build the capabilities they
Scan • Download • Personalize
By harnessing novel data and the power of AI, need and scaling them across the portfolio and their
biopharmaceutical companies can move beyond assets’ life cycles. The clinical development and trial
RCTs to vastly improve clinical development. The design system is ripe for change, and the value to
benefits accrue at each stage of the development patients will be vast.

Chris Anagnostopoulos is an associate partner in McKinsey’s Athens office, David Champagne and Alex Devereson are
partners in the London office, Thomas Devenyns is an associate partner in the Geneva office, and Heikki Tarkkila is an
associate partner in the Helsinki office.

Designed by McKinsey Global Publishing

Copyright © 2023 McKinsey & Company. All rights reserved.

How artificial intelligence can power clinical development 9