using deep learning for product innovation in the personalized care space

Artificial Intelligence in Health

Shalini Ananda, PhD [email protected]

http://quantifiedskin.com1061 Market St. #511 San Francisco, CA 94103

mailto:[email protected]

http://quantifiedskin.com


Deep Learning Teaches Computers to Think

Our brain’s neocortex utilizes layers to create representation from low level inputs to transform them into meaningful assertions.

Deep learning enables computers to do the same.


Deep Learning Assists in Feature Engineering

Raw DataFeature

ExtractorFeatures

Algorithm (machine learning)

(train or predict)


Deep Learning Assists in Feature Engineering

Raw DataFeature

ExtractorFeatures

Algorithm (machine learning)

(train or predict)

Deep learning


Easy to Tell These Are Cookies


Necessary to Have a Deep Understanding of Material Science


What If a Computer Could Think Like a Researcher?

Published experiments contain magnitudes of errors, but there are meaningful patterns if you know what to look for


We Can Teach Computers to Learn the Relevant Scientific Data

Ever increasing number of experiments

Experimentation steps


Humans Use a Logical Approach to Experimentation

1

Elucidating the role of a material or molecular

system

i.e. Targeting and sustained

release

2

Estimating the effect of a material system in the

state of action

i.e. Biodistribution and half-life

3

Design of material

Run experiment

Prepare for next iteration

Human’s can’t physically test all possibilities


Need for Reducing Irrelevant Experiments

Experiments must factor characteristics such as:

•Targeting modalities •Charge modifications •Shape, and more…

If there are 10 nanoparticles and 6 possible parameters to change and 43 variants; that’s over:

10258 experiments

Not enough time to run every experiment.Smart materials

Lets use nanoparticles as an example


Our Validation Approach

1

We outreached to researchers working on

smart materials.

The researchers provided us characterization input data.

Three step process

We conducted 152 blind experiments with 31 institutions.

A few of the institutions we worked with


Our Validation Approach

1

We outreached to researchers working on

smart materials.

The researchers provided us characterization input data.

2

Our team ran the characterization inputs

through our deep learning engine, NuSilico™.

We predicted biodistribution and half-life outputs.

3

Researchers ran their experiments

(avg. length 4 months).

We compared NuSilico™’s outputs to the researcher’s

experimental data.

Three step process


Our Results - Our Predictions vs Observed Empirical Experiments

From 152 experiments with 31 institutions

R2 = 0.9370 R2 = 0.9656


Deep Learning Is Better for Feature Selection than SVM

recall: 61.40

precision: 66.80

AUC: 84.01

SVM polynomial kernel deep learning

recall: 76.80

precision: 88.10

AUC: 93.70

In terms of sensitivity/recall, ROC, and precision where the laplacian score was utilized as the feature selection method for building the model. We compare SVM polynomial kernel to deep learning


Our Simulations Produced Accurate Models in a Fraction of the Time

We have repeatably demonstrated an order of magnitude time improvement using deep learning

Over 39 months

Just 1 month

Current researching methods

What we’ve demonstrated with deep learning


Building Personalized Therapies - Systematically

1

Determine the target cells.

What are you trying to achieve physically?

Three step process

2

Understand the building blocks.

Train the computer to test outcomes of experiments.

3

Use deep learning to simulate and rank the best

designs.

Test orders of magnitude of experiments to find the most optimal design path.

Thank You

using deep learning for product innovation in the personalized care space

Technology

comdeep learning assists

deep learning engine

researchers experimental

material system

published experiments

blind experiments

comour validation approach1we

characterization input