2020
Artificial Intelligence
Trends To Watch
In 2020
2
CB Insights is a tech market intelligence
platform that analyzes millions of data
points on venture capital, startups, patents,
partnerships, and news mentions to help
you see tomorrow’s opportunities, today.
WHAT IS CB INSIGHTS?
CLICK HERE TO LEARN MORE
3
1. Commercial deepfakes will resurrect celebrities,
shake up retail, and revolutionize influencer
marketing
2. Next-gen hacking: AI trojans, voice spoofing,
and smart evasion evolve
3. AutoML: AI is the future of AI design
4. Federated learning will bring in a new data
partnership ecosystem
5. Alphabet will use AI to dominate smart city
contracts
6. AI will leave a massive carbon footprint —
and we’ll need AI to fix it
7. Doing more with less: Tackling small data
problems in AI will be a major focus
8. Quantum machine learning will take baby
steps to give traditional AI algorithms a boost
9. Natural language processing will help us
understand the building blocks of life
5

9
13
16
20
23
28
33
37
Table of Contents
4
AI in 2020
Artificial intelligence is fundamentally changing software as we
know it.
Corporations are moving past the hype around the technology to
discern how it can add practical value. Specifically, data for AI will
be a major theme in 2020, as new techniques that train AI with less
data or increase data privacy protections gain traction.
Energy efficient AI, quantum neural networks, and the role of
natural language processing in understanding amino acids and
proteins are some areas where we’ll see cutting-edge research this
year.
But as the tech matures, it’s also bringing along challenges that
didn’t exist a couple of years ago.
“Deepfakes” are becoming mainstream, making real and fake
media indiscernible. AI models can have an inherent bias, and
hackers are exploiting the model’s bias to fool it. Open-source AI
tools — which have been fundamental in democratizing AI — are
also easily available for use in malicious applications, like creating
the next generation of malware.
In this report, we dive into these and other trends we’ll be watching
closely in artificial intelligence this year.
Artificial Intelligence Trends To Watch In 2020
5
Artificial Intelligence Trends To Watch In 2020
Deepfakes are hyper-realistic AI-generated images and videos.
Media companies are ready to monetize the “benign” side of the
controversial tech.
Deepfakes are controversial, having already made their way into
political videos and morphed pornography. But they also provide a
great product opportunity for media companies determined to tap
into the “fun and goofy” side of the tech.
At the end of December 2019, Snapchat acquired AI Factory,
a Ukraine-based startup developing computer vision products,
for $166M.
Snap had previously worked with AI Factory to power Cameos,
a feature that enables users to insert selfies into GIFs to create
animated deepfakes. Bytedance-owned TikTok is working on a
similar feature.
Samsung published a paper on using neural nets to create realistic
“talking heads.” Below, the image on the left shows the source, and
the ones on the right are AI-generated.
Source: Samsung research
Commercial deepfakes will resurrect
celebrities, shake up retail, and
revolutionize influencer marketing
6
The Financial Times reported on a growing divide between traditional
computer generated graphics — which are often expensive and time-
consuming — and the recent rise in deepfake tech.
The report says that “deepfakery shaved several years off British
actor Bill Nighy in Pokémon Detective Pikachu,” if only for a few
frames in the film.
Although a full-fledged, deepfake-driven motion picture may
not yet be feasible, Hollywood is heading towards “digitally
resurrecting” actors from the ’50s and ’60s in films — a trend that
will benefit from advances in AI.
On the retail front, deepfakes will let brands hyper-personalize visual
marketing for consumers. Startup Superpersonal, for instance,
swaps out users’ faces in short video clips for virtual try-ons.

Source: Superpersonal
7
“Consider it you, but as your best fashion
model self. The whole process now takes
only three minutes (to give an idea of
how fast the cloud computing process
is moving, back in February it took 20
minutes).”
— KATIE BARON, FORBES
Deepfakes are also making a dent in influencer marketing. Startup
Synthesia used deepfake tech to make David Becham speak in 9
different languages in a campaign video for the NGO Malaria Must
Die. The startup has since raised $3M from LDV Capital, Mark
Cuban, and others.
Source: Malaria Must Die
8
IMPLICATIONS
• Personalization: Face swapping in retail is all about
doubling down on consumer experience. The tech will boost
e-commerce experience and virtual online try-ons.
• Hyper-targeted ads: As the tech becomes commoditized,
hyper-local advertising — such as instant dubbing in different
languages — would become low-hanging fruit.
• Automation in creative fields: The use of deepfake tech in the
TV and film industries could lead to proliferation in sequels,
spin-offs, and cultural adaptations of existing content. In other
fields, such as casting and modeling, AI-generated faces could
stunt demand for human influencers and models.
9
New-age hacks will evolve on two fronts: fooling AI systems and
leveraging AI to launch sophisticated attacks.
In 2019, researchers from Sydney-based Skylight Cyber found a way
to bypass the AI-based antivirus software created by Cylance — a
cyber AI unicorn Blackberry had acquired just a few months prior.
Skylight reported that it found an inherent bias in the AI model and
exploited it to create a universal bypass that allowed malware to
go undetected.
“...if you could truly understand how
a certain model works, and the type
of features it uses to reach a decision,
you would have the potential to fool it
consistently, creating a universal bypass.”
— SKYLIGHT CYBER
As cyber AI startups raise more funds to protect companies and
consumers with the tech, a new crop of hackers and malware that
target weaknesses unique to AI will emerge.
Next-gen hacking: AI trojans, voice
spoofing, and smart evasion evolve
10


Hackers can also fool AI through data poisoning. Corrupting
the training data of AI algorithms impacts how the AI eventually
classifies malicious and normal behavior in a network. In other
types of adversarial attacks, hackers can introduce small
perturbations, which are invisible to the human eye, to an image
to trick a neural network into misinterpreting it.
While cyber defenders grapple with these new-age threats that
increase with AI adoption, AI itself can be used to create more
sophisticated, hyper-targeted cyber attacks.
Media reports of AI-generated voice spoofing first emerged in
Europe in Q3’19. AI software was used to mimic the voices of CEOs
in phone calls to employees with instructions to transfer money,
followed by an email with details of the transfer. At least three
different companies, including a British energy firm, were targeted.
11
Source: Wall Street Journal
Deepfakes will likely also be misused in extortion scams.
Although real-world incidents have yet to emerge, IBM developed
a proof of concept for a deep learning-powered malware, dubbed
DeepLocker, back in 2018. Hackers can make hyper-targeted
malware by concealing it within AI code. DeepLocker only unlocks
the attack when it encounters a very specific trigger, such as
a particular person’s face or voice. Otherwise, the code sits
undetected in benign everyday applications.
Source: blackhat USA 2018
12
With open-source tools available, the barrier for entry into AI is low
for hackers.
“While a class of malware like DeepLocker
has not been seen in the wild to date,
these AI tools are publicly available... In
fact, we would not be surprised if this type
of attack were already being deployed.”
— SECURITY INTELLIGENCE
IMPLICATIONS
• Hackers are relentless and AI tools are more easily available
to everyone today than ever before.
• Cyber AI startups will face new attack vectors. Hackers have
proved it’s easy to exploit inherent biases of machine learning
models and trick the algorithm.
• Heavy industries are underprepared. In the last decade,
malware including Stuxnet, BlackEnergy, Havex, Troton, and
Industroyer have targeted industrial control systems, from
Iranian nuclear plants to Ukraninan power grids. Surveys
show asset-heavy industries have not kept up with evolving
cyber risks and are not prepared for more advanced threats
like AI malware.
13
AutoML, a suite of AI tools to automate design and training of
neural networks, will democratize the tech by lowering the barrier
to entry for enterprises with minimal AI expertise.
Designing or searching for the right neural network architecture
from thousands of available ones for a specific task is a time-
consuming process. It becomes harder when designing AI for more
complex situations like autonomous driving, where both speed and
accuracy are critical.
Neural architecture search (NAS) is a set of AI approaches to
automate the process of finding the best AI design for a given task.
Google officially coined the term “AutoML” for this in 2017.
“Typically, our machine learning models
are painstakingly designed by a team of
engineers and scientists...if we succeed
[with AutoML], we think this can inspire
new types of neural nets and make it
possible for non-experts to create neural
nets tailored to their particular needs…”
— GOOGLE AI BLOG
AutoML: AI is the future of AI design
14

Source: Google AI
Since then, adoption of AutoML tools for AI design — including data
preparation, training, model search, and feature engineering — has
been gradually increasing.
Waymo, for example, recently partnered with Google to automate
the process of finding the best neural network architecture to
enable autonomous vehicles to identify trees, pedestrians, and
vehicles from lidar (light detection and ranging) data.
This adoption is fueled by cloud computing giants like Google
making AutoML available to its existing client base. Google Cloud
AutoML can be used for computer vision, video processing,
translation, and NLP tasks.
Startups are also offering plug-and-play solutions to enterprises.
Databricks, a unicorn in the data management and analytics space,
introduced AutoML last year. DataRobot, H2O, and RapidMiner are
some other startups that offer AutoML solutions to enterprise clients.
15
IMPLICATIONS
AutoML helps to viably scale AI for two reasons:
• Talent shortage: There’s an acute shortage of AI experts, and
AutoML will democratize the tech for enterprises with minimal
AI expertise.
• Cost and complexity: Designing neural nets is a time-
consuming and manual process, even for experts. AutoML
can create better solutions and cut down computation costs
associated with a trial-and-error approach. We wrote about
how AutoML helped Waymo design better AI for perception
tasks here.
16
Federated learning shows promise in training AI in industries
with sensitive and siloed data. In 2020, it will enable a new data
partnership model without requiring users to actually share the
raw data.
Federated learning, which allows for increased data privacy while
still improving the AI model, is used for applications like Google’s
text prediction software and URL searches in Firefox.
Google initially debuted the tech for its Android keyboard, Gboard, to
predict what a user will type next. In the company’s Q2’19 earnings
call, Sundar Pichai stressed federated learning and other privacy
control measures being major focus areas for the tech company.
“On user privacy and control, it’s always
been a big focus for us...Initiative is
underway for example like federated
learning for almost three years...I think
it’s one of the most important areas we
are working on.”
— SUNDAR PICHAI
Federated learning will bring in a new data
partnership ecosystem
17
As depicted below, federated learning allows the Gboard software
to improve its AI model without sending raw personal data back
to Google.
Source: Google
In this case, data stays on your phone instead of being sent to or
stored in a central cloud server. A cloud server sends the most
updated version of an AI algorithm — called the “global state” of
the algorithm — to a random selection of user devices.
Your phone makes improvements and updates to the AI model
based on your localized data. Only the update — not the data used
to make those updates — is sent back to the cloud to improve the
“global state” and the process repeats itself. (Read about federated
learning, and how it differs from other distributed learning
approaches here.)
The ability to protect user data while still improving AI algorithms
make federated learning a viable option for industries such as
healthcare and banking, where regulatory concerns over data
sharing are higher.
18
For example, Nvidia’s AI-powered hardware and software
framework for healthcare, called Clara, now supports federated
learning. Initial users of the tech include American College of
Radiology, MGH and BWH Center for Clinical Data Science, and
UCLA Health.

Source: Nvidia
Nvidia also partnered with healthcare startup Owkin, which uses
federated learning to predict patients’ resistance to cancer drugs.
The chipmaker is extending this application to the auto and
transportation sectors, enabling cross-country partnerships for
accelerating autonomous vehicle research.
In finance, China-based digital bank WeBank is partnering with
parent company Tencent’s cloud division and Canadian AI
research institution Mila for federated learning research.
19
IMPLICATIONS
• Global model, local data: With federated learning, users
can train AI on data stored locally, and only share AI model
updates with a cloud server. The “global model” will then
benefit all partners in the network to improve their local
AI applications.
•
Increase data diversity: Federated learning will enable
more cross-institutional or cross-country partnerships,
eventually allowing the global AI model to benefit from
diverse local datasets.
20
The tech giant is joining forces with local governments to create
new city blocks, redrawing competitive lines in sectors ranging
from real estate to energy utility to transportation, and more.
Alphabet, a $1T AI powerhouse, is making inroads into urban
development and smart city planning with IoT and machine learning.
Alphabet subsidiary Sidewalk Labs released a 1,500 page document
in Q2’19 highlighting its plans for a $1.3B smart city development
project in Toronto with government and private partners.
This has significant implications for the use of AI in government and
urban planning.

Source: Sidewalk Labs
Alphabet will use AI to dominate smart
city contracts
21
Smart city planning is a broad concept encompassing smart
health, smart mobility, surveillance, and data infrastructure, among
other things, with a number of different use cases for AI and
machine learning.
Alphabet’s reach and AI initiatives in each of those industries make it
a formidable competitor for public contractors in sectors ranging from
real estate to energy utility to transportation to consulting services.
For example, Sidewalk Labs has been spinning off smaller
companies like Replica and Coord to tackle specific urban
development challenges with AI.
Replica uses machine learning to model commuter behavior and
answer questions about what influences a commuter’s choices,
such as taking public transport. The Portland government will pay
Replica over $450K for a one-year initial service period, and Illinois
has signed a $3.6M deal with the startup for a 3-year period.
22
Coord uses machine learning to map curbside assets. The company
is inviting cities to apply to its 2020 Digital Curb Challenge. Its
technology will be offered to winners free of cost, and the pilot will
give Coord the opportunity to fine tune its platform and strategies.
Apart from this, Sidewalk Labs’ pilot project in Toronto is
emphasizing building smart cities with drastically reduced
greenhouse emissions and smarter resource management.
The company is hiring ML engineers to analyze data from sensors
and building management systems to build recommendation
engines and predictive models related to energy consumption
and sustainability.
IMPLICATIONS
• AI expertise makes Alphabet a formidable competitor for
government contracts: Sidewalk Labs has the advantage
of pooling in resources and expertise from other Alphabet
subsidiaries, including DeepMind, Waymo, and the moonshot
lab X.
• End-to-end solutions: While a number of vendors offer
services like curbside information to cities, Alphabet is
well-positioned to offer everything from ML-based urban
development tools to autonomous vehicles to building energy
management.
• Financial risk undertaking: Google is also able to share risk
and invest upfront. Sidewalk Labs announced it would bear a
“disproportionate share of the cost of upfront innovation” in its
public-private partnership model, and receive compensation
when it meets performance metrics in later stages. This
increases the chances that municipalities and governments
will experiment with the tech.
23
Given how computationally intensive AI is, the tech will not only
necessitate smarter, sustainable solutions — it will also help meet
a growing global energy demand.
One of the reasons many advances in AI have been top-down (i.e.,
tech giants pioneering AI R&D and open-sourcing tools to others)
is how computationally intensive AI research is.
To put this in perspective, Fast Company reported that Google
used up “the equivalent of the electricity that the average
American household uses in just under six months” for its BigGAN
experiment in 2018 to create hyper-realistic images of dogs,
butterflies, and burgers.
AI for energy will be a major theme in 2020, from tech giants and
automakers to oil & gas companies looking to cut costs, improve
efficiencies, and meet global power consumption.
AI will leave a massive carbon footprint —
and we’ll need AI to fix it
24
Two distinct trends are emerging within AI and energy:
• Energy-efficient AI devices
• AI tools for large-scale energy management
First, energy efficiency will take center stage as AI comes to more
edge devices, such as phones and cameras, since edge computing
does not have the same power and resources as cloud computing.
For instance, Kneron is one company that recently announced low-
power AI processors for edge devices.
As another example, Apple acquired Xnor.ai, a startup that makes
low-power edge AI tools, in Q1’20.
“...our hardware engineering and machine
learning teams asked the audacious
question, ‘can we create a hardware,
machine learning architecture capable of
running deep learning models without a
battery? That can be so low-power they
can harvest ambient energy from the
sun?’”
— XNOR.AI
25
Xnor.ai was working on ultra-low powered cameras that can run AI
algorithms. The acquisition was an obvious move for Apple, which
is doubling down on AI chips and VR apps for iPhones.
The second trend to watch for in this area is AI-drive energy
management and forecasting for large-scale power plants
and utilities.
Google has been making a major push towards purchasing 100%
renewable energy for its data centers — and it’s leveraging AI to
help do this. The tech giant partnered with DeepMind to use neural
networks to improve wind energy output.
Source: Google AI research
DeepMind’s neural nets were able to predict wind power output 36
hours in advance based on weather forecasts and wind turbine data.
26
“Based on these predictions, our model
recommends how to make optimal hourly
delivery commitments to the power grid
a full day in advance. This is important,
because energy sources that can be
scheduled (i.e. can deliver a set amount
of electricity at a set time) are often more
valuable to the grid.”
— DEEPMIND
Early-stage startups like Bill Gates-backed Heliogen are
already experimenting with niche applications, like using AI to
algorithmically control the position of heliostats (mirrors) that
capture sunlight.
27
IMPLICATIONS
• Hardware companies will focus on “ultra-low power” devices
for ML: Energy efficiency is a main consideration for AI at the
edge (running on devices such as smartphones, smart home
cameras, etc).
• AI for utility-scale energy production: More cloud giants
will transition to using sustainable energy, leveraging AI for
increasingly renewable energy output and for streamlining
datacenter operations.
• Streamlining operations for power plants and oil & gas:
Artificial intelligence can predict renewable energy output,
automate grid management, aid in precision drilling of oil
wells, and power sustainable energy management solutions in
smart homes and commercial buildings.
28
There are two workarounds if you don’t have sufficient data to train
data-hungry deep learning algorithms: generate synthetic data or
develop AI models that work well with small data.
Deep learning is extremely data hungry — models are trained on
huge sets of labelled data, like millions of images with animals
identified and tagged to teach the AI to identify them — and
massive amounts labeled data are not available for certain
applications. In such cases, training an AI model from scratch
is difficult, if not impossible.
One potential solution is to augment real datasets with synthetic
data. This has particularly taken off in autonomous driving, where
AVs drive millions of miles in photo-realistic simulated environments
that can recreate situations like snowstorms and unusual pedestrian
behavior, where acquiring real-world data is hard.
Niche synthetic datasets, like the fake MRIs from Nvidia shown
below, are also emerging to augment lack sufficient real-world data
for rare diseases.







Source: Nvidia
Doing more with less: Tackling the small
data problem in AI will be a major focus
29
Another way to work around the data issue is to develop AI models
that can learn from small data sets.
One approach that has taken off in computer vision tasks is
transfer learning. This means taking an AI algorithm that’s pre-
trained for a different task where there is ample labeled data
available (for example, identifying cars in images), and transferring
that knowledge to a different application for which there is little
data (like identifying trucks).

Taking a pre-trained model is like taking a readymade pizza crust
and customizing it instead of making the dough from scratch. While
it has taken off in computer vision, pre-training was challenging in
natural language processing (NLP) given the lack of labeled data in
general, until now.
30
An approach called self-supervised pre-training is becoming
popular in NLP.
AI is pre-trained on the enormous amount of text readily available
on the Internet. For example, OpenAI pre-trained an AI model with
8M pages of internet text — an extremely compute-intensive task.
During training, the model’s task was to predict the next word in a
sentence based on preceding words.
This is called self-supervised learning because there are no
“labels” involved here: the AI learns about language by predicting a
hidden word based on other words in a sentence.
Researcher Jeremy Howard explains why these self-supervised
language models are so important in an excerpt from fast.ai:
“We are not necessarily interested in
the language model itself, but it turns
out that the model which can complete
this task must learn about the nature of
language and even a bit about the world
in the process of its training. When we’d
then take this pretrained language model,
and fine tune it for another task, such
as sentiment analysis, it turns out that
we can very quickly get state-of-the-art
results with very little data.”
— JEREMY HOWARD, FAST.AI
31
Another popular example is BERT by Google, where the AI language
model not only predicts a word based on the preceding words, but
also the succeeding ones (bi-directional understanding of context).
Source: Google
Facebook’s AI division, led by Yann LeCun, has been bullish on
self-supervision. One example is pre-training a language model
similar to the above examples and fine-tuning it for applications
like identifying hate speech.
Source: Facebook
32
Facebook recently open-sourced its work on self-supervised
learning for speech recognition, bypassing the need for manually
annotated transcripts by smaller research projects. Facebook
open-sourced the code, wav2vec, which is particularly useful for
speech recognition in non-English languages, where annotated
training data is sparse.
IMPLICATIONS
• NLP will hog the limelight in 2020 as a result of self-
supervised techniques. We will finally see better downstream
NLP applications like chatbots, advanced machine translation,
human-like writing, etc.
• Big tech players leading the way: Research on AI models for
small data is top-down given how compute intensive it is to
develop these pre-trained language models. Tech companies
are open-sourcing their research so that other researchers
can use it for downstream applications.
• Synthetic data and tools that generate realistic fake data level
the playing field for smaller companies that don’t have access
to massive datasets that tech giants do.
33
Hybrid models, which combine classical machine learning algorithms
with quantum AI, will see some practical applications soon.
Quantum computers require specialized data preparation, quantum
algorithms, and quantum programming. In short, the way we interact
with classical computers won’t work with quantum computers.
Quantum machine learning borrows from the principles of
traditional machine learning, but the algorithms are designed to
run on quantum processors. This makes them faster than typical
neural nets and overcomes hardware constraints that limit current
AI research on massive datasets.
Quantum machine learning will take
baby steps to give traditional AI
algorithms a boost
A quick refresher: Unlike binary computing, where
information is stored either as 0 or 1, quantum computers
are based on qubits. Qubits can be any value from 0 to 1,
or have properties of both of these values simultaneously.
Right away, there are many more possibilities for
performing computations.
Despite the hype, today’s most powerful quantum
computers, including those being developed by Google, are
capable of harnessing the power of 50 to 100 qubits. (To
put this in perspective, for quantum computers to have a
significant commercial impact, researchers say we require
at least a few thousand qubits.)
34
Research on quantum neural networks (QNN) is very nascent.
So given the current hardware limitations of quantum processors,
how can QNN algorithms solve any real-world problems?
“Traditional machine learning took many
years from its inception until a general
framework for supervised learning was
established. We are at the exploratory
stage in the design of quantum neural
networks.”

— GOOGLE
Tech giants and quantum startups are looking at a hybrid
approach, where part of the task is done by traditional neural
networks that run on normal computers and another part of it is
augmented by QNNs.
Startup Xanadu applied hybrid classical-quantum AI to transfer
learning (see the previous trend on small data for more about
transfer learning). The results were promising for image
classification tasks.
35
Source: Xanadu research paper, arxiv.org
Google’s AI team has been focusing on writing algorithms for
quantum computers since 2013. The immediate goal, similar
to Xanadu’s, is to develop “hybrid quantum-classical machine
learning techniques on near-term quantum devices.”
“While the current work [on QNN] is
primarily theoretical, their structure
facilitates implementation and testing
on quantum computers in the
immediate future.”

— GOOGLE AI BLOG
36
In two research papers published on the topic, Google explores
peculiar ways to train QNNs compared to classical neural net
training methods, and tests a QNN’s ability to perform simple
image classification tasks in simulation.
IMPLICATIONS
• We will start seeing two of the world’s most powerful computing
paradigms, quantum computing and AI, solve practical problems
initially in conjunction with classic computers.
• Quantum cloud computing is the latest frontier in cloud wars,
with all major providers — AWS, Google, IBM, and Microsoft—
doubling down on efforts or venturing into it. This means
quantum computers will work in tandem with traditional GPUs
and CPUs to add value to cloud clients, as highlighted in a
2020 paper published by Rigetti, Microsoft, and OpenAI. We
will see cloud AI algorithms running on such hybrid hardware
platforms.
37
Both natural language processing and genomes are comprised of
sequential data. AI algorithms that do well in one field are proving to
be useful in the other in surprising ways.
In the self-supervised learning example discussed earlier in this
report, researchers masked specific words in a sentence and had
the algorithm guess the missing word to learn about languages
more broadly.
Just as sentences are a sequence of words, proteins are a
sequence of amino acids in a specific order. Researchers at
Facebook AI and NYU used the same concept of self-supervision
on massive datasets of protein sequences.
Instead of hidden words, AI has to predict what the hidden amino
acid is.


Source: Biorxiv
Natural language processing will help us
understand the building blocks of life
38
Researchers from Germany tapped into a similar concept of self-
supervised language models for classifying proteins.
The most popular recent development was in genomic modeling.
DeepMind developed an algorithm, Alphafold, to understand protein
folding — one of the most complex challenges in genomics — to
determine the 3D structure of proteins.
“...it would take longer than the age of the
known universe to randomly enumerate
all possible configurations of a typical
protein before reaching the true 3D
structure — yet proteins themselves fold
spontaneously, within milliseconds.”

— DEEPMIND

Source: DeepMind
39
Although AlphaFold uses a hybrid method, it borrows concepts
from natural language processing to predict distance and angle
between amino acids.
IMPLICATIONS
• Better drug design: Several drug candidates today target
proteins, but proteins dynamically change structures based
on environmental factors. Understanding their structure and
how they fold presents the opportunity to develop drugs for
previously unknown targets. Companies like Relay Therapeutics
are focused on understanding how proteins move in order to
model them, which will aid in new drug discovery.
• AI algorithms can help model proteins and understand their
structure without requiring in-depth domain knowledge.
•
It would be possible to create or optimize new protein designs
for specific functions in both healthcare and material science.
The CB Insights platform
has the underlying data
included in this report
WHERE IS ALL THIS DATA FROM?
CLICK HERE TO SIGN UP FOR FREE