What it will take to cross the Valley of Death: �Translational Systems Biology, “True” Precision Medicine, Medical Digital Twins, Artificial Intelligence and In Silico Clinical trials

Gary An, MD, FACS

Department of Surgery

University of Vermont

Seminar Series

June 2, 2022

MSM Viral Pandemic Working Group

◤

◤

Translational Systems Biology

Primary Features:

• Dynamic Computational Modeling to Accelerate Hypothesis Testing
• Simulate Clinical Contexts => In Silico Clinical Trials
• Use abstraction to identify conserved functions across biology (species, individuals, etc…)

2008

2015

Name Motivated by the Zerhouni-Era NIH “Roadmap” 2003-5 =>

“Translational” Emphasis

◤

Axioms of “True” Precision Medicine

• Axiom 1: Patient A is not the same as Patient B (Personalization)
• Axiom 2: Patient A at Time X is not the same as Patient A at Time Y (Precision)
• Axiom 3: The goal of medicine is to treat, prognosis is not enough (Treatment).
• Axiom 4: Precision medicine should find effective therapies for every patient and not only to identify groups of patients that respond to medicine (Inclusiveness)

An and Day, “Precision Systems Medicine: A Control Discovery Problem”

Systems Medicine: Integrative, Qualitative and Computational Approaches, in press

◤

The Valley of Death

Most Impactful Failure Point in Drug Development

◤

Reasons for the Valley of Death

• Biological heterogeneity => intrinsic variability in how an organism responds to its environment (and changes thereof): Denominator Problem
• Current means of determining what is similar or conserved between one biological object and another are descriptive (do not account for how the dynamics/behavior of a system arise)

◤

The Denominator Problem

A

B

C1

C2

A’

◤

Current paths lead to the Valley of Death

• There is too much variation in the study population (signal-to-noise)
• Reducing variation/narrowing study population intrinsically reduces the power of a planned clinical trial
• The current means of narrowing a study population is ill-posed because there lacks a formal means of determining and characterizing what actually causes one patient to behave differently from another patient
• There is no process for “useful failure” in clinical trials

Can we learn from other disciplines?

◤

What other Sciences do = Engineering

• Engineering:
• Formal Process for identifying solutions given a set of system constraints
• Uses Math/Formal Representations

• Hard for what makes biology biology (i.e. behavior of cells)
• No first principles/natural laws
• Lack of “trustworthy” formal representations
• Reverse engineering difficult due to non-path uniqueness and intrinsic heterogeneity (epistemic limits of current methods = essentially Alchemy…)

◤

Failure to Translate: Medical Digital Twins

• A data structure for the real-world system
• Some process that links data together to form dynamics
• Some link to the real world that feeds back data into the data-propagation/generation process - Grieves 2019

• There exists a common formal specification that is shared by and links different individual examples within a type/class of objects.
• The goal of the Digital Twin is to examine the future state of the targeted specific physical object, meaning that accounting for dynamics and trajectories of behavior are necessary.
• The common specification leads to the ability to simulate the behavior of the object, meaning that dynamics and behavior are intrinsic to the specification.

Therefore

◤

Axioms of Precision Med + Medical Digital Twins

• A MDT of any individual means we are able to make a MDT of every individual (Axioms 1, 2, 4)
• Therefore, MDTs need to encompass heterogeneity (Axiom 1 and 2)
• Trajectories are implicit in MDT (Axiom 2) => mechanisms. And data-centric methods are strongly discouraged for linking time series (Wright and Davidson, 2022)
• “Medical” in MDT => treatment (Axiom 3)
• Epistemic uncertainty => MDTs that are “good enough” => Thus finding “good enough” must be overall goal (Axiom 4)

Therefore...MDTs ideally should further goal of developing effective control of disease for everyone

◤

Complex Control Discovery: ML and AI

Brief Diversion to 3 Things everyone should know about ML/AI:

• Universal Approximation Theorem: ANN can approximate any answer => Brittleness and Underspecification
• Curse of Dimensionality: Manifests as problem of data sparsity => synthetic data: but how to generate?
• Causal Hierarchy Theorem: While UAT means ANN can reconstitute any function, CHT says an ANN/SCM cannot predict the effect of an intervention on that function based on the data used to train the ANN alone

◤

”Exception” to CHT: Simulation-based Deep Reinforcement Learning (DRL)

• Simulation-based DRL = Game Playing AIs
• “Exception” because these are actually hybrid systems => incorporate mechanism-based simulations to generate synthetic data
• Proposal: Simulation-based DRL for complex control discovery for biomedical problems (proof-of-concept for sepsis)

◤

Populations to Populations of Individuals: In Silico Trials

• In Silico trials should overcome barriers in Valley of Death. But…
• How to encompass human heterogeneity in a clinical population with incomplete knowledge (“good enough”)?
• How to capture what is similar but able to generate heterogeneity (formal representation/DT specification)?

◤

One Approach: Clinical Heterogeneity = Parameter Landscape => Model Rule Matrix (MRM)

◤

Bridging The Valley of Death

• Augmented Throughput over the Valley of Death

◤

Grand Challenges (for Journal)

1. How to encompass and embrace the biological heterogeneity present in clinical populations in the design of in silico clinical trials? 
2. How to generate synthetic data that “obscures” the generative simulation model from ML and AI approaches? 
3. How to design and use Translational Systems Biology models/Medical Digital Twins to inform the next generation of clinical sensors and controllers to achieve True Precision Medicine?
4. How to integrate mechanism-based Translational Systems Biology simulation models with cutting edge developments in ML and AI to further control discovery and achieve the Axioms of Precision Medicine? 

◤

Grand Challenges (for Us and MSM)

• How to overcome bias against mechanism-based dynamic modeling?
• Classes of Bias:
• From Biology Side: Don’t appreciate abstraction
• A: You need abstraction to determine what is similar => engineering
• From Physics Side: “No First Principles/Natural Laws”
• A: Iterative refinement => Weather Modeling
• From Stats/ML Side: “We don’t need you, we just need data” (Endemic problem in current AI…)
• A: Understand limits of methods => bridge with synthetic data gen
• From Math Bio Side: “Models too complex”
• A: Complex problems require complex solutions => analogous to double-descent in Deep Learning (but really ¯\_(ツ)_/¯ )

◤

���Finis

◤