2YRREVIEW

Large Science Models:

A General Framework for Generative Modeling of Microbiome-Metabolome systems

Ishanu Chattopadhyay, PhD

Assistant Professor of Biomedical Informatics & Computer Science

University of Kentucky

first wave

rule-based systems

second wave

Big Data / ML / Deep Learning

recognize patterns, make predictions, might improve over time, but struggle on tasks not trained for

third wave

contextual reasoning, generelizable models, stepping towards true intelligence

Control Systems

Screening for complex diseases
Digital Twins in biology & medicine
Modeling of human behavior
Large Science Models

Robotics

Self-organization of sensor networks

data smashing

inverse Gillespie

PhD

Postdoc

ZeDLAB

Mechanical Engineering MS, PhD

Mathematics MA

Computer Sc

Medicine

Career Trajectory

ZeDLAB

Biomedical Informatics

Collaborators

Gary Hunninghake, Pulmonary C, Harvard

Robert Gibbons, Bio-statistics

Peter Smith, Pediatrics

Michael Msall Pediatrics

Fernando Martinez, Pulmonary Critical Care, Weill Cornell

James Mastrianni, Neurology

James Evans, sociology

Erika Claud, Pediatrics

Andrew Limper Mayo Clinic

Department of Pediatrics

UChicago

Department of Neurology & The Memory Center

UChicago

Department of Psychiatry

UChicago

Pulmonary Critical Care, Weill Cornell

Department of Anesthesia and Critical Care

UChicago

Center for Health Statistics

UChicago

Pulmonary Critical Care, Harvard Medical School

Department of Psychiatry

UIC

Demon Network, Exeter, Alan Turing Institute, UK

Dalhousie University, Canada

Pritzker School of Molecular ENgineering

Social Science

UChicago

Collaborations

D3M (I2O)

PAI (DSO)

PREEMPT (BTO)

YFA (DSO)

NIA

~3.5M USD in 5 years

Pre-UK Funding

Publications

Impact

Nature Medicine

Nature Human Behavior

Nature Commun-ication

Science Advances

(3)

PNAS

JAMA

JAHA

JACC

Modeling & predicting complex social interactions

Point-of-care screening for complex diseases

Electronic Healthcare Record

IPF

ASD

ADRD

Research Thrusts

General framework for inferring digital twins in biology and medicine

Universal point-of-care screening via AI-driven pattern recognition
No new tests or blood-work
Uses routine data (EHR) already in patient file
No specific data demands
Generalizable in future to other targets beyond PF, ILAs

CKD

ILD

ZeBRA

ICD

Enable early diagnosis

Target PF/IPF or ILDs broadly

Seamless background integration with Epic workflows

Primary care

*Onishchenko, Dmytro, Robert J. Marlowe, Che G. Ngufor, Louis J. Faust, Andrew H. Limper, Gary M. Hunninghake, Fernando J. Martinez, and Ishanu Chattopadhyay. "Screening for idiopathic pulmonary fibrosis using comorbidity signatures in electronic health records." Nature Medicine 28, no. 10 (2022): 2107-2116.

Raising Flags before patient or their doctor notice symptoms

downstream care modulation

model published, retrospectively validated*

TimestampedDiagnostic procedural codes & prescriptions

SI/SA

ZeBRA: Method for Individualized Future-Risk State Inference and Patient-Journey Phenotyping From Routine Longitudinal Health Data

AI-driven Test-Free Prediction of ICU Admission, Insulin Dependence, and Exocrine Dysfunction after Acute Pancreatitis

\Lambda-OR \textrm{ Attribution}

ZeBRA Publications

AI-driven Test-Free Prediction of ICU Admission, Insulin Dependence, and Exocrine Dysfunction after Acute Pancreatitis

Predictive Performance (PF)*

High AUC across high and low risk sub-cohorts

Highlights:

1 yr out AUC ~88%
Positive Likelihood ration ~40

*Onishchenko, D., Marlowe, R.J., Ngufor, C.G. et al. Screening for idiopathic pulmonary fibrosis using comorbidity signatures in electronic health records. Nat Med 28, 2107–2116 (2022). https://doi.org/10.1038/s41591-022-02010-y

Model

Interstitial Lung Disease (ILD)

US prevalence: 1 in 500 (higher by 1.5X-2X in KY)

IPF prevalence: 10-25% of ILD

Age group: 50-85 years old

Observation window:

1+ years of records

Prediction window: 1 year

Used dataset size:

Case: 25.4k, Control: 15.1M

Performance (95% Specificity):

Males:

AUC 82.2% (82.0%, 82.5%)

Sensitivity 39.7% (39.3%, 40.1%)

Positive LR: 7.81 (7.85, 8.01)

Negative LR: 0.64 (0.64, 0.63)

Females:

AUC 82.1% (81.8%, 82.3%)

Sensitivity 39.1% (38.7%, 39.5%)

Positive LR: 7.77 (7.74, 7.90)

Negative LR: 0.64 (0.65, 0.64)

Current validation results (MarketScan)

ILD

Natural history of ILD progression: Predicting key events in patient journey

High AUCs.

High PPV: 1-2 false positive for every true flag

Interstitial Lung Disease (ILD)

US prevalence: 1 in 500

	95% specificity/39% sensitivity	99% specificity/17% sensitivity	99.5% specificity/12.5% sensitivity
Additional ILD diagnoses from ZeBRA	546	238	175
Total ILD diagnoses per year with ZeBRA	746	438	375
Additional transplant-eligible patients with ZeBRA	164	71	53
Expected False Positives	29,930	5,986	2,993

Net annual contribution margin*	$46,613,500	$12,706,700	$7,703,350

Case Study: University of Kentucky Heath-care (UKHC)

Patient population: 600K unique patients per year
Current ILD diagnoses: 200 per year

* diagnostic workup margin (CT+PFT): $950, lunng transplant contribution margin: $120,000, incremental program operating cost: -$1.5M

Target	AUC
Frailty / Physical Debility	96.2%
Alzheimer's Disease and Related Dementia (ADRD)	93.4%
Chronic Fatigue Syndrome / ME	93.2%
Acute Pancreatitis: ICU Visit	92.3%
Chronic Pancreatitis: Exocrine Pancreatic Insufficiency	92.1%
Idiopathic Pulmonary Fibrosis (IPF)	91.6%
Sarcopenia	91.0%
Parkinson's Disease	87.9%
Dementia / Degenerative Neurologic Disease	87.8%
Acute Pancreatitis: Insulin Dependence	87.2%
Suicide Attempts / Suicidal Ideations (Males 50--75)	86.0%
Chronic Inflammation	85.9%
Heart Failure with Preserved Ejection Fraction (HFpEF)	84.9%
Suicide Attempts / Suicidal Ideations (Males 25--50)	84.0%
Interstitial Lung Diseases (ILD)	82.2%
Age-related Macular Degeneration	82.1%
Autism Spectrum Disorder (ASD)	81.8%
Chronic Kidney Disease (CKD)	81.8%
Cerebral Infarction	81.1%
Chronic Obstructive Pulmonary Disease (COPD)	81.0%
Major Depressive Disorder	80.5%
Myocardial Infarction / Cardiac Arrest post-arthroplasty	80.1%
CKD Progression to Stage 4+	80.1%
Prostate Cancer	80.0%
Osteoporosis	79.5%
Post-Traumatic Stress Disorder (PTSD)	78.1%
Hearing Loss	72.7%
Osteoarthritis	72.5%
Systemic Connective Tissue Disorders	72.0%

ZeBRA Model Family (expanding list)

Suicidal Ideation and Self-harm Attemps: Estimating risk from routine EHR

High AUCs.

High PPV: 1-2 false positive for every true flag

Suicidal Ideation and Self-harm Attemps: Estimating risk from routine EHR

High AUCs.

High PPV: 1-2 false positive for every true flag

~ 4yrs

current survival ~4yrs

~ 4yrs

current clinical DX

ZCoR screening

Onishchenko, D., Marlowe, R.J., Ngufor, C.G. et al. Screening for idiopathic pulmonary fibrosis using comorbidity signatures in electronic health records. Nat Med 28, 2107–2116 (2022). https://doi.org/10.1038/s41591-022-02010-y

n=~3M

AUC~90%

Likelihood ratio ~30

Alzheimer's Disease and Related Dementia*

* in press

>5 Million in US. >13 Million in next 10 years

Alzheimer's Disease and Related Dimentia

MOCA, Blood Tests

Current Practice:

state of art with EHR:

~67% AUC*

ZCoR: ~87%

Alzheimer's Disease and Related Dimentia

state of art with EHR:

~67% AUC*

ZCoR: ~87%

Preempting ADRD accurately upto a decade in future

Applicable To Screening for Mild Cognitive Impairment

Clinical Trial Participant Selection

Current screen-failure rate: 80-90%

Estimated rate with ZCoR:

40%

Research Direction II

Digital Twins

General framework for inferring digital twins in biology and medicine

Stamping Out the Next Pandemic **Before** The First Human Infection

BioNorad

Digital Twins for complex systems

Darkome

teomims

opinion dynamics

algorithmic lie detector

Mental health diagnosis

viral emergence

microbiome

Digital Twins for complex systems

Darkome

teomims

opinion dynamics

algorithmic lie detector

Mental health diagnosis

viral emergence

microbiome

Phase 1

Phase 2

PREPARE: Pioneering Research for Early Prediction of Alzheimer's and Related Dementias EUREKA Challenge

Algorithm for early diagnosis

Find Data for early prediction

Phase 1

Phase 2

Second Prize 40,000 USD

Lets give them:

1M patients clinical data diagnosed with ADRD/AD 60-80 years
1M African-American patients from Chicagoland
Open source - GNU public license

licensed patient data

digital twin

(generative AI)

teomims

(open cohort)

VeRITaAS

Can A Generative AI Tell if you Are Lying?

Vetting Response Integrity from
cross-Talk in Adversarial
Surveys

Q-Net

Hidden structure of cross-talk between responses to interview items

PTSD diagnostic interview

Number of possible responses

Minimum Performance (n=624)

Average Time: 3.5 min

No. of questions: 20

AUC > 0.95

PPV > 0.86

NPV > 0.92

At least 83.3% sensitivity at 94% specificity

Minimum AUC = $0.95 \pm 0.005$

Cannot be coached, or memorized

Datasets for training & validation

1. VA (n=294)

2. Prolific (n=300)

3. Psychiatrists (n=30)

10^{25}

Beat the test!

paraknowledge.ai/veritas

200 participants in

100 participants in

30 forensic psychiatrists

Can-You-Fake-PTSD Challenge Results

successful attempts

Future

Vision

Universal screening for IPF, ADRD, autism, rare cancers
Continuous monitoring of health
Bio-NORAD
Digital twins

Transform bio-surveillance

Democratize AI unleashing its power for social good

Transform early diagnosis

Transform modeling of complex systems

Impact on Popular Discourse on AI

Media Coverage

National Pop-culture Discourse

Interviews, Op-eds, and Forum Appearences

Joe Rogan Podcast
Walter Isaacson Interview
Speaker on Pritzker Forum on Global Cities
>150 News articles written on published papers

Rotaru, Victor, Yi Huang, Timmy Li, James Evans, and Ishanu Chattopadhyay. "Event-level prediction of urban crime reveals a signature of enforcement bias in US cities." Nature human behaviour 6, no. 8 (2022): 1056-1068.

"test-free" screening?

Autism
Idiopathic Pulmonary Fibrosis
Alzheimer's Disease and related dementia
Suicidality, PTSD
Perioperative Cardiac Event
Aggressive Melanoma
Uterine Cancer
Pancreatic Cancer

non-existent biomarkers

expensive, time-consuming diagnostic tests

Lack of Universal Screening at the point of care

Early diagnosis is difficult, late or missed diagnosis costs lives

We lack Universal Screening

for most diseases

Prognosis at Point-of-Diagnosis

Optimizing Management

Patient Journey

Continuous Risk Monitoring

Early Diagnosis

Universal Screening

Cohort Selection

Reduce screen failure rates

Holistic health surveillance

Predict antifibrotics continuation

improve outcomes

Interstitial Lung Disease / Pulmonary Fibrosis

Rapid Universal Point-of-care Screening for ILD/IPF Using Comorbidity Signatures in Electronic Health Records

Flag patients before they (or doctors) suspect

Primary Care

Pulmonologist

Zero-burden Co-morbid Risk Score (ZCoR)

Referral

shortness of breath

dry cough

doctor can hear velcro crackles

Non-specific Symptoms

>50 years old

more men than women

IPF

Rare disease

~5 in 10,000

Post-Dx

Survival

~4 years

Cannot always be seen on CXR

At least one misdiagnosis

~55%

Two or more misdiagnosis

38%

Initially attributed to age related symptoms:

72%

PCP workflow demands

Known Co-morbidities of PF

Are there more? Subtle footprints in the medical history that are more heterogeneous?

~ 4yrs

current survival ~4yrs

~ 4yrs

current clinical DX

ZCoR screening

n=~3M

AUC~90%

Likelihood ratio ~30

Conventional AI/ML attempts to model the physician

AI in IPF Research

Co-morbidity patterns
No data demands
Use whatever data is already on patient file

ICD administrative codes

IPF

ILD

target codes appear

Past medical history

No target codes appear

case

control

2yrs

prediction

target codes appear

Past medical history

No target codes appear

case

control

2yrs

IPF drugs prescribed

Signature of IPF diagnostic sequence

pirfenidone or nintedanib

age > 50 years
at least two IPF target codes identified at least 1 month apart
chest CT procedure (ICD-9-CM 87.41 and Current Procedural Terminology, 4th Edition, codes 71250, 71260 and 71270) before the first diagnostic claim for IPF
no claims for alternative ILD codes occurring on or after the first IPF claim

ICD Codes can be noisy

"cases" are not always true IPF

Truven MarketScan (IBM)
Commerical Claims & Encounters Database
2003-2018

>100M patients visible

>7B individual claims

>87K unique diagnostic codes

>7% Medicare data present

2,053,277 patients included in study

University of Chicago Medical Center 
2012-2021

68,658 patients

Random sample from Optumlabs Data Warehouse courtsey Mayo Clinic

861,280 patients

2,983,215 patients

Data: Onishchenko etal. Nat. Medicine 2022

patient A

patient B

patient C

Beyond "risk factors" to personalized risk patterns

Clinical Trial Cohort Selection

Current screen failure rate ~50-60%

ZCoR boosted screen failure rate ~20%

cohort size: 2000

initial cohort size: 5000

initial cohort size with ZCoR: 2500

Cost per patient for confirmatory tests: ~7k USD

Savings: ~20M USD

Clinical Trial Cohort Selection

Current screen failure rate ~50-60%

ZCoR boosted screen failure rate ~20%

cohort size: 2000

initial cohort size: 5000

initial cohort size with ZCoR: 2500

Cost per patient for confirmatory tests: ~7k USD

Savings: ~20M USD

Upto 4 year "signal" resolution

decreases risk

increases risk

Patient Journey: Tracking Risk over time

Autism

1 in 59

MCHAT/F

ZeD Lab: Predictive Screening from Comorbidity Footprints

CELL Reports

	ZCoR	Competition
Autism	>83%	"obvious"
Alzheimer's Disease	~90%	60-70%
Idiopathic Pulmonary Fibrosis	~90%	NA
MACE	~80%	~70%
Bipolar Disorder	~85%	NA
CKD	~85%	NA
Rare Cancers (Bladder, Uterus)	~75-80%	Low
Suicidality (with CAT-SS)	98% PPV	Low

Off-the-shelf AI does not suffice

How?

Odds ratios combined via ML

Data

cases

control

\vdots

odds ratios for all ICD codes

ML Model

odds-based risk estimator

\rho(X) = \zeta\left (\bigcup_i \bigg \{ \mathcal{O}(x_i) \bigg \}\right )

minimize generalization error by constraining model capacity

Conservation of complexity!

K(x) = K(S) + K(x \vert S_\star) + O(1)

for digital twins

K(x \vert S_\star) = O(1)

Research Direction II

Digital Twins

General framework for inferring digital twins in biology and medicine

Chattopadhyay, Ishanu, Kevin Wu, Jin Li, and Aaron Esser-Kahn. "Emergenet: Fast Scalable Pandemic Risk Assessment of Influenza A Strains Circulating In Non-human Hosts." (2023). Under Review in Nature

PREEMPT

Predicting Future Mutations for Viral Genomes in the Wild

predict future emergence risk

\Phi_i:\prod_{j \neq i} \Sigma_j \rightarrow \mathcal{D}(\Sigma_i)

Q-Net

recursive forest

q-distance

a biologically informed, adaptive distance between strains

\theta(x,y) \triangleq \\ \mathbf{E}_i \left ( \mathbb{J}^{\frac{1}{2}} \left (\Phi_i(x_{-i}) , \Phi_i(y_{-i})\right ) \right )

This distance is "special"

Smaller distances imply a quantitatively high probability of spontaneous jump

$$J \textrm{ is the Jensen-Shannon divergence }$$

Metric Structure

Tangent Bundle

geometry

dynamics

\theta(x,y) \sim \log Pr(x \rightarrow y)

\theta

Influenza Risk Assessment Tool (IRAT) scoring for animal strains

slow (months), quasi-subjective, expensive

*https://www.cdc.gov/flu/pandemic-resources/monitoring/irat-virus-summaries.htm

24 scores in 14 years

~10,000 strains collected annually

CDC

Emergenet time: 1 second

Stamping Out the Next Pandemic **Before** The First Human Infection

BioNorad

THE PROBLEM

Assuming a 1000 species ecosystem, and 1 successful experiment every day to discern a single two-way relationship, we would need 1,368 years to go through all possibilities.

Digital Twin for the Maturing Human Microbiome

Forecast microbiome maturation trajectories

Predict neurodevelopmental deficits

Boston U

U Chicago

Two centers

Ability to "fill in" missing data is equivalent to making trajectory forecasts

predicting neurodevelopmental deficits

forecasting ecosystem trajectories

Which entities are most predictive

of neurodevelopmental deficit

entity X timestamp

SHAP value

No transplantation is guaranteed to work reliably

Just add those microbes back to reduce risk?

No!

Bacterial transplantation must be personalized

Future task:

Explicit supplantation profiles that are tuned to individual ecosystems

No transplantation is guaranteed to work reliably

Just add those microbes back to reduce risk?

No!

Bacterial transplantation must be personalized

Future task:

Explicit supplantation profiles that are tuned to individual ecosystems

Phase 1

Phase 2

Uncorrelated, yet indistinguishable !!

What is a healthy Microbiome/ Metabolome?

Can gut microbiota pre-empt developmental deficits?

High Abundance Classes

Complex Interdependecies Exist over time and taxonomic strata

quantized output levels

Purely wet-lab investigations are insufficient to understand complex ecosystems

We need to scale up!

We need a digital twin

of the dataset

Learn from uncurated data

understand non-trvial cross-dependencies between entities, without prior knowledge

can explore valid perturbations of the overall system

can identify deviations from "normal" behavior

Can model out-of-box targeted and untargeted metabolites (or uncharacterized microbes)

Can microbial assay from gut actionably

pre-empt developmental deficit?

Study Set-up

Predicting neurodevelopmental deficits

Forecasting ecosystem trajectories

Build classifiers

Ability to "fill in" missing data is equivalent to making trajectory forecasts

Forecast ecosystem fluctuations

Large Science Models

Consider a large number ($n$) of coupled observables
- We dont know potential couplings a priori

Compute all $x_i \vert x_1,\cdots,x_{i-1},x_{i+1},\cdots, x_n$ conditionals

Estimate the joint distribution of all variables

Brook’s lemma*

*Brook, D. (1964). On the distinction between the conditional probability and the joint probability distribution. Journal of the Royal Statistical Society. Series B (Methodological), 26(2), 295–307.

perfect inference of these conditionals gives us a unique joint

The central problem of ML is computing joint distributions or estimates thereof

hard!

How Complex Are These Models?

Hundreds of thousands to 10s of millions of features

The Goal: Create a digital twin which can reveal valid perturbations

\psi_0

\psi_t

Completely uninformative state

Observed state

\rho(x) =\frac{\theta_\mathcal{H}(\psi_0,x)}{\theta_\mathcal{D}(\psi_0,x)}

Risk

Geometric Interpretation

LSM-based Risk

Infer LSM for typical development
Infer LSM for disease trajectories

\mathcal{H}

\mathcal{D}

How different are the individual estimators for typical and dysbiotic models?

Bacilli 30

typical

deficit

Coriobacteria 32

typical

deficit

Gammaproteobacteria 32

typical

deficit

All Patients

Feeding Variables added

Forecasting of class abundance variations both

in-sample ($ R^2 \ge 95\%$)

and out-of sample ($R^2\ge 72\%$)

Forecasting of class abundance variations both

in-sample ($ R^2 \ge 95\%$)

and out-of sample ($R^2\ge 72\%$)

Building classifier based on LSM metric

moving the top drivers impact risk of deficit maximally

Understanding Key Drivers via LSM perturbations

Just add those microbes back?

No! The LSM indicates that supplantations need to be patient specific

No transplantation is guaranteed to work reliably

Predicted to reduce

risk reliably

Predicted to reduce

risk reliably

Network Interpretations? We see clear differences between two cases

Typical

Deficit

Integrating Clinical Features

IBD Metabolomics: Preliminary Results

Dataset from Metabolomics Workbench

Study ID	ST000923
Study Title	Longitudinal Metabolomics of the Human Microbiome in Inflammatory Bowel Disease

Institute	Broad Institute of MIT and Harvard
Last Name	Avila-Pacheco
First Name	Julian
Submit Date	2017-11-14
Num Groups	3
Total Subjects	546
Num Males	276
Num Females	270
Analysis Type Detail	LC-MS

State-of-art microbiome based Classification (~10 species) *

IBD vs UC	0.82
IBD vs CD	0.76

*Zheng, J., et al. (2024). Noninvasive, microbiome-based diagnosis of inflammatory bowel disease. Nature Medicine, 30(12), 3555–3567. https://doi.org/10.1038/s41591-024-03280-4

IBD vs non IBD

0.85

Gut-Metabolome based Classification (~36 metabolites) *

Application 2

Large Science Models

1. How will proposer form and maintain a computationally tractable LSM tree structure given, as proposed, hundreds to thousands of observable variables?

$\checkmark$

Each predictor is inferred independently
Can scale up to thousands of variables in Python implementation
Further scale-up $10^6 - 10^8$ needs C/C++ implementation

https://34.66.189.202/data/trees_mbol_UC

Full Example of Hyperlinked Trees

Metabolomics

LSM model

No. parameters: 70 million
Out of sample n=150
Uses all names and unnamed metabolites (>81K features)

	AUC (out of sample)
Healthy vs IBD	96.1%
Healthy vs UC	92%
UC vs CD	99%
Healthy vs CD	99%

Metabolomics

LSM model

No. parameters: 70 million
Out of sample n=150
Uses all names and unnamed metabolites (>81K features)

	AUC (out of sample)
Healthy vs IBD	96.1%
Healthy vs UC	92%
UC vs CD	99%
Healthy vs CD	99%

Metabolomics

Insight: the discriminating hypersurface is 2d (almost 1d)

Number of metabolites	5,280	85% untargeted metabolites
Number of parameters	4,002,306
Average Tree Depth	38.62
Number of constraints inferred	108,884	~83% involve untargeted metabolites
Number of samples used	180	(no clinical phenotype information used)

Using all samples of metabolite profiles
Not using clinical phenotypes and replicates

Application 3

ASD Metabolics: Preliminary Results

R_{\mathcal A}(y) := \|L(y)\|_2 = \left( \sum_{k=1}^K \bigl[-\log \Pr(a_k\to y)\bigr]^2 \right)^{1/2}

ASD Metabolics: Preliminary Results

ASD sample profile

ASD samples from same patient

sample profile of new patient $y$

Pr(a \rightarrow y)

Need one patient!

Predictive Performance

	AUC	Sensitivity at 95% spec
LSM	92.7%	74%
MCHAT/F	67%	39%
ADOS-2	90-97%	85%

getting close to the gold standard

1 false positive

1 false negative

10% flag in TBD (expected 8.3% positives)

Predictive Performance

Top risk drivers mapped to known pathways

P_x(m_j) = \frac{\partial R(x)}{\partial m_j}

Patient specific driver profile

\displaystyle P(m_j) = \frac{1}{N} \sum_x P_x(m_j)

Average driver profile

Top 30 risk drivers (targeted met.)

Tissue and Serum Metabolomics (Targeted and Untargeted) in Fibrotic Interstitial Lung Disease (F-ILD)

Uncovering latent geometry separating disease phenotypes

Application 4

Tissue and Serum Metabolomics (Targeted and Untargeted) in Fibrotic Interstitial Lung Disease (F-ILD)

Learn LSM with tissue samples -> Can disambiguate serum samples

LSM model for healthy profiles

What is a healthy Microbiome/ Metabolome?

\mathcal{H}

Any profile generated by $\mathcal{H}$ is a healthy profile, while they might be different from one another

A Universal Risk Index

\theta_\mathcal{H}(\psi_\star,x)

average healthy profile

A New Paradigm of AI-driven Discovery in Metabolome Biology

Supplantation MUST be bacteroidia

Supplantation MUST be Actinobacteria

No risk-decreasing supplantation

*Hothorn, Torsten, Kurt Hornik, and Achim Zeileis. "Unbiased recursive partitioning: A conditional inference framework." Journal of Computational and Graphical statistics 15, no. 3 (2006): 651-674.

LSM Forest of Conditional Inference Trees*

Revealing Emergent Cross-talk between mutations in a viral protein (Influenza A HA)

Component predictor (Conditional Inference Tree*)

Example: Influenza A HA protein

Large Science Models: Properties

LSM-Distance Metric*

\theta(\psi,\psi') \triangleq \frac{1}{N}\sum_{i=1}^{N} \sqrt{D_{JS}\Bigl(\phi^i(\psi^{-i}) \vert \vert \phi^i(\psi'^{-i})\Bigr)}

where $D_{JS}(P\vert \vert Q)$ is the Jensen-Shannon divergence.

g_{ij}(\psi) \;=\; \frac{1}{2}\,\frac{\partial^2}{\partial \psi^i\,\partial \psi^j}\,\theta^2(\psi,\psi')\Biggr|_{\psi'=\psi}

\left \lvert \ln \frac{\Pr(\psi\to \psi')}{\Pr(\psi' \rightarrow \psi')}\right \rvert \le \beta\,\theta(\psi,\psi')

Large Deviation Bound*

Induced Riemannian metric tensor

This bound connects ``closeness'' of samples to the odds of perturbing from one to the other, bridging geometry to dynamics

Ergodic Projection

\psi_\star \triangleq \bigotimes_{i=1}^N\phi^i\left (\prod_{1}^{N-1}\varnothing\right )

(Sanov's Theorem, Pinkser's Inequality)

$\psi$

$\psi'$

$\theta$

"spatial average": average of all plausible worldviews or states

* Sizemore, Nicholas, Kaitlyn Oliphant, Ruolin Zheng, Camilia R. Martin, Erika C. Claud, and Ishanu Chattopadhyay. "A digital twin of the infant microbiome to predict neurodevelopmental deficits." Science Advances 10, no. 15 (2024): eadj0400. https://www.science.org/doi/full/10.1126/sciadv.adj0400

persistence probability

Ergodic dispersion

\Psi_\star = \theta(\psi,\psi_\star)

Central to Model Drift Quantification

Start with opinion vector with all entries missing

This is a standard Physics construct, quantifying curvature of the underlying latent geometry

Pr(\psi \rightarrow \psi')

Easily computable in LSM framework!

Apply $\phi^i$

Random variable quantifying dispersion around the spatial average of worlviews

const. scaling as $N^2$

Digital Twin Inference

Consider ecosystem of many inbitants
The abundunce of each entity is a variable
The variables interact in complex ways
We are going to model this cross-talk

The LSM Framework

\Phi_i:\prod_{j \neq i} \Sigma_j \rightarrow \mathcal{D}(\Sigma_i)

Q-Net

recursive forest

Proposed Concept

Develop Foundation models of complex systems with
- hundreds to thousands of evolving variables with apriori unknown cross-talk
- no governing equations are know a priori
- reflexivity: system changes if observed
Learn intrinsic system geometry from data
Derive equations of motion with variational principles (stationary action on Lagrangian).
Inference under data sparsity
Detect data (in)sufficiency, adapt to model drift
Support forward simulation and perturbation analysis

Data inference boundaries & limitations

Alignment validation

Complex phenomena

Adaptation to model obsolence

Precise validation protocols to assess process drift triggering re-calibration/training

Built-in flexibility for changing contexts and non-ergodicity

Scalable to thousands to millions of variables, intrinsic reflexivity

Component LSM predictors enforce statistical significance of splits in recursive partitioning, ensuring precise uncertainty quantification

*Hothorn, Torsten, Kurt Hornik, and Achim Zeileis. "Unbiased recursive partitioning: A conditional inference framework." Journal of Computational and Graphical statistics 15, no. 3 (2006): 651-674.

emergent macro-structure

Component predictor (Conditional Inference Tree*)

Example: Influenza A HA protein

Recursive

LSM

forest

LSM Forest of Conditional Inference Trees*

Revealing Emergent Cross-talk

Large Science Models: Mathematical Framework

\begin{aligned} \text{Observables:} \quad & \color{yellow}X = \{x^1, \ldots, x^N\}, \overbrace{x^i \in \Sigma^i}^{\text{finite alphabet}} \\ {\color{gray}\text{Notation:} }\quad & \color{gray} x^{-i} = \{x^j : j \ne i\}\\ \text{Crosstalk:} \quad & \forall i \ P(x^i) = \color{red} f_i(x^{-i}) \\ \text{System state:} \quad & \color{Cyan} \psi = \bigotimes_{i=1}^N \psi^i, \quad \psi^i \in \mathscr{D}(\Sigma^i) \cup \varnothing \\ %\textbf{Degenerate case:} \quad & \psi^i \text{ is a delta distribution (fully observed)} \\ {\color{gray}\text{Notation:} }\quad & \color{gray}\psi^{-i} = \bigotimes_{j \ne i} \psi^j \end{aligned}

	reliten	gunlaw	abany	---	grass
Person 1
Person 2
---
Person m

observables

samples

Distributions over alphabet $\Sigma^i$

\phi = \bigotimes_{i=1}^N \phi^i, \quad \phi^i(\psi^{-i}) \in \mathscr{D}(\Sigma^i) \\

Individual Predictor (CIT)

cross-talk

\phi(\psi) \vert \vert \psi

Tension between predicted and observed distribution drives change

Example

GSS topic: There should be more gun-control

$\psi^i$

strongly agree

agree

neutral

disagree

strongly disagree

\Sigma^i

Digital Twin

\phi^i(\psi^{-i}) \sim \widetilde{\psi}^i

$\phi$ estimates $\psi$

Examples: GSS, ANES, WVS, ESS, Eurobarometer, Afrobarometer, Asian Barometer etc

group

individual

estimate is always a non-empty non-degenerate distribution

missing observation

Large Science Models: Properties

LSM-Distance Metric*

\theta(\psi,\psi') \triangleq \frac{1}{N}\sum_{i=1}^{N} \sqrt{D_{JS}\Bigl(\phi^i(\psi^{-i}) \vert \vert \phi^i(\psi'^{-i})\Bigr)}

where $D_{JS}(P\vert \vert Q)$ is the Jensen-Shannon divergence.

g_{ij}(\psi) \;=\; \frac{1}{2}\,\frac{\partial^2}{\partial \psi^i\,\partial \psi^j}\,\theta^2(\psi,\psi')\Biggr|_{\psi'=\psi}

\left \lvert \ln \frac{\Pr(\psi\to \psi')}{\Pr(\psi' \rightarrow \psi')}\right \rvert \le \beta\,\theta(\psi,\psi')

Large Deviation Bound*

Induced Riemannian metric tensor

This bound connects ``closeness'' of samples to the odds of perturbing from one to the other, bridging geometry to dynamics

Ergodic Projection

\psi_\star \triangleq \bigotimes_{i=1}^N\phi^i\left (\prod_{1}^{N-1}\varnothing\right )

(Sanov's Theorem, Pinkser's Inequality)

$\psi$

$\psi'$

$\theta$

"spatial average": average of all plausible worldviews or states

persistence probability

Ergodic dispersion

\Psi_\star = \theta(\psi,\psi_\star)

Central to Model Drift Quantification

Start with opinion vector with all entries missing

This is a standard Physics construct, quantifying curvature of the underlying latent geometry

Pr(\psi \rightarrow \psi')

Easily computable in LSM framework!

Apply $\phi^i$

Random variable quantifying dispersion around the spatial average of worlviews

const. scaling as $N^2$

Digital Twin & Fidelity of Simulation

\mathcal{N}_\epsilon(\psi) \triangleq \big\{ \psi': {\color{red}\forall i \ \psi'_i \sim \phi^i\left ( \psi^{-i}\right )} \wedge {\color{yellow} \theta(\psi,\psi') \leqq \epsilon }\big \}

Sample predicted distributions

perturbed state within $\epsilon$ of $\psi$

Digital Twin

-Neighborhood of state $\psi$

\epsilon

Definition

Sample neighborhood to impute missing data

\psi

\epsilon

LSM sampling: sampling the $\epsilon$-neighborhood of a state or worldview allows reconstruction of censored opinions

Predictive ability of LSM quantified as ability to reconstruct censored out-of-sample observations

{\color{Tomato}\psi_\star }\rightarrow \psi \rightarrow \cdots \rightarrow \psi'

Null state (all missing observations)

Valid perturbations/ simulations

LSM sampling allows simulating opinion perturbations

Global Emergent Structure via Clusters & Poles

2018 GSS

\theta_t(\psi_+,\psi_-)

Polar separation over time

2016 Presidential Election Vote Prediction

2004

abany	no	yes
abdefctw	always wrong	not wrong at all
abdefect	no	yes
abhlth	no	yes
abnomore	no	yes
abpoor	no	yes
abpoorw	always wrong	not wrong at all
abrape	no	yes
absingle	no	yes
bible	inspired word	book of fables
colcom	fired	not fired
colmil	not fired	not allowed
comfort	strongly agree	strongly disagree
conlabor	hardly any	a great deal
godchnge	believe now, always have	don't believe now, never have
grass	not legal	legal
gunlaw	oppose	favor
intmil	very interested	not at all interested
libcom	remove	not remove
libmil	not remove	remove
maboygrl	true	false
owngun	yes	no
pillok	agree	strongly agree
pilloky	strongly disagree	strongly agree
polabuse	no	yes
pray	several times a day	never
prayer	disapprove	approve
prayfreq	several times a day	never
religcon	strongly disagree	strongly agree
religint	strongly disagree	strongly agree
reliten	strong	no religion
rowngun	yes	no
shotgun	yes	no
spkcom	not allowed	allowed
spkmil	allowed	not allowed
taxrich	about right	much too low

conservative pole

\psi_+

liberal pole

\psi_-

Clustering LSM distance $\theta(x,y)$ between out-of-sample individuals

conservative

liberal

poles:

partial states aligning with extreme opposing worldviews

Compare across time and different GSS surveys
Derived features for individuals (ideology index)

I(x) = \frac{\theta(x,\psi_+) - \theta(x,\psi_-)}{\theta(\psi_+,\psi_-)}

Predict 2016 votes using ideology index

Emergent global structure

Reflexivity and State Collapse on Observation

Emergent Equations of Motion

L \triangleq \frac{1}{2} \sum_i g_{kl} P^k_p \dot{\psi}^p_i P^l_n \dot{\psi}^n_i - \theta(\psi, \phi)

Define Lagrangian*

\frac{d}{dt} \left( \frac{\partial L}{\partial \dot{\psi}^m_i} \right) - \frac{\partial L}{\partial \psi^m_i} = 0

Via the Euler-Lagrange Equations$^\dag$:

\ddot{\psi}^m_i = -g^{km} P^k_m \frac{1}{2N} \sum_j \frac{1}{\sqrt{D_{JS}(\psi^m_j \| \phi^m_j)}} \left[ \ln\left( \frac{2e\psi^m_j}{\psi^m_j + \phi^m_j} \right) - \frac{1}{2(\psi^m_j + \phi^m_j)} \right]

Over-damped Gradient flow Equation*

where $-g^{km}$ is the inverse metric tensor

kinetic energy

state collapse

strongly agree

agree

neutral

disagree

strongly disagree

strongly agree

agree

neutral

disagree

strongly disagree

Query/

Observation

$X_i$

Non-local Influence propagation on measurement/observation (QM-like)

\phi^i(\psi^{-i})

potential energy

* Einstein notation used

Goldstein, Herbert, et al. Classical Mechanics. 3rd ed., Pearson, 2002.

$^\dag$

Principle of stationary action

Dynamics

Local potential field eqn

Local Potential Fields

Stable

(captured by local extrema)

Free to move locally towards extrema

GSS 2018 individuals and neighborhoods

Influenza C : strains and their neighborhoods

Observation: This lineage (Mississippi lineage) is now extinct since 2022/23

stable lineage

Local potential fields can be computed given the LSM and dynamical considerations, which reveal future evolution

Data Sufficiency via Conservation of Complexity

%K(x) = K(S) + K(x \vert S_\star) + O(1) = K(S') + K(x \vert S'_\star) +O(1) K(x \vert S_\star) = O(1) = K(S \vert x_\star)

The No-cheating Thorem: Generative models cannot cheat on complexity

Kolmogorov Complexity

Optimal Generative Model

compressed data representation

compressed model representation

Theorem

K(\textrm{data}) = K(\textrm{LSM}) +O(1)

Conservation Law arising from the continuous symmetry of typicality*

\mu_0(X) \triangleq \frac{\delta(\vert \langle S(X) \rangle \vert)}{\delta(\vert \langle X \rangle \vert)} \leq 1

Saturation relation:

Data Sufficiency Statistic $\mu_0$

We need LSM-sampling to calculate this

*Noether's Theorem

For every continuous symmetry of a physical system, there exists a corresponding conserved quantity

\vert \langle X' \rangle \vert \approx \max\{1,\mu_0(X)\} \vert \langle X \rangle \vert

How much more data do we need?

Data saturation

Data deficient

Needed

Current

Empirical Validation

Model Drift Quantification

Ergodic dispersion

\Delta_\star = \theta(\Psi,\psi_\star)

z(\Delta_\star) = \frac{\Delta_\star^{[t]} - \langle \Delta_\star^{[t]} \rangle}{\sigma(\Delta_\star^{[t]} )}

Z-value of dispersion

Do new samples (survey respondents) still conform to the model?

GSS Model drift

ergodic projection (all missing values)

A random belief state (with possibly missing entries)

random variable

normal variate

\zeta(M) = \vert z(\Delta_\star^0) - z(\Delta_\star^{[t]}) \vert

Model drift stochastic process ($\zeta$)

\mathbf{E}(\zeta(M) )

assess if $\zeta$ is stationary: if not then new samples are not conforming to model

Example for GSS LSM inferred for year 2000

Large Science Models & Ergodicity

$\checkmark$ 4. Address whether your approach makes assumptions regarding ergodicity, and if so, how these assumptions affect the model's applicability to non-ergodic systems.

No Convergence

(~50% belief mismatch between pairs)

2018 GSS survey belief vectors simulated via LSM sampling

No ergodicity assumption: LSMs are built for non-ergodic systems
Sampling and simuation "remembers" the start point (No convergence), demonstrating non-ergodic learned structure
Local potential fields vary across the space
Potential wells may arise, driven by the dynamics at hand, not via assumptions
"change" is driven by non-equilibrium (dissonance)

Embedded Social Theories in LSM

When applied to Social Modeling and Opinion Dynamics

Belief about topic $i$ is expected to align with beliefs about other topics $\displaystyle\psi^{-i}$.
Deviations are exponentially improbable $\Rightarrow $ people/groups seek internal coherence.
Theory Link:
- Cognitive consistency theory – Abelson et al. (1968)
- Constraint satisfaction in beliefs – Read & Marcus-Newhall (1993)

Beliefs evolve to minimize tension between actual state and “expected” state.
Reflexive gradient flow — system reduces internal contradiction.
Theory Link:
- Cognitive Dissonance Theory – Festinger (1957)
- Homeostatic belief adjustment – Gawronski & Strack (2004)

Observing a belief changes it and affects all conditionals.
Direct encoding of feedback loops central to human systems.
Theory Link:
- Reflexivity in social systems – Giddens (1984), Soros (1994)
- Theory of mind / mutual modeling – Premack & Woodruff (1978)

Validation of Social Theory Questions:

Perception changes reality, which changes perception
The Constitution of Society
The Alchemy of Finance
Does a chimpanzee have a theory of mind?

Our system “wants” to reach a low-energy (low-dissonance) state — a direct computational analog of Festinger’s theory.

People strive to align beliefs and attitudes across related domains. Inconsistencies create cognitive discomfort, prompting adjustments across belief clusters to restore harmony.

Exploratory: Belief systems react measurably to exogenous events and shocks

Exploratory: Cross-dependencies between beliefs have observable effects on societal resilience.

Is Polarization an Inevitable Attractor?

Social Identity Theory vs. Belief Proximity

Large Science Models: Broader Applications

A General Framework for modeling Complex Systems

Genomic database: Missing heritability problem

Personalized Clinical Digital Twin, Virtual Patients

Any structured interview, PTSD fabrication

Assess sysmptom data and co-pathologies

Predict future mutations; which animal strain is closest to jumping to humans

Mental health diagnosis

Microbiome Analysis**

Algorithmic lie detector

Viral emergence

Teomims

Opinion Dynamics

Darkome

Generative model of complex microbial ecosystems, and their impact on health and disease

Data requirements

Tabular data
Potentially large number of features/covariates ($10^2 - 10^8 $)
Sufficient number of samples ($10^3 - 10^6$)
Small number of longitudinal samples (currently, $ < 100$)

Limitation	Mitigation / Response
Conventional time series is currently out-of-scope	Focus on cross-sectional interdependencies and belief geometry; time handled via drift
LSMs model statistical interdependence, not causal mechanisms	Use perturbation-based simulations to infer plausible influence pathways
Limited by observed belief variables	Integrate multiple surveys; use latent proxies and test sensitivity of digital twins
Social theory connections and interpretability may be challenging	Anchor dynamics with theory-driven constructs (e.g., ToM, cognitive dissonance)

LSMs for complex systems

**preliminary study published (https://www.science.org/doi/10.1126/sciadv.adj0400)

[
    {
        "patient_id": "P000038",
        "sex": "F",
        "birth_date": "01-01-2006",
        "DX_record": [
            {"date": "07-31-2006", "code": "Z38.00"},
            {"date": "08-07-2006", "code": "P59.9"},
            {"date": "08-29-2016", "code": "J01.90"},
            {"date": "09-10-2016", "code": "J01.90"},
            {"date": "11-14-2016", "code": "J01.91"}
        ],
        "RX_record": [
            {"date": "10-29-2011", "code": "rxLDA017"},
            {"date": "05-16-2015", "code": "rxIDG004"},
            {"date": "08-08-2015", "code": "rxIDG004"},
            {"date": "06-04-2016", "code": "rxIDD013"}
        ],
        "PROC_record": [
            {"date": "02-05-2007", "code": "90723"},
            {"date": "11-05-2007", "code": "J1100"}
        ]
    }
]

{
  "predictions": [
    {
      "error_code": "",
      "patient_id": "P000012",
      "predicted_risk": 0.005794344620009157,
      "probability": 0.8253881317184486
    }
  ],
  "target": "TARGET"
}

Data Out

Data In

Current state: Fully functional API *

user-specific API
Custom models targeting range of disorders
Scalable

*Documentation: https://github.com/zeroknowledgediscovery/paraknowledgedoc

Model ready to deploy behind UK firewall

>5 Million in US. >13 Million in next 10 years

Alzheimer's Disease and Related Dimentia

MOCA, Blood Tests

Current Practice:

state of art with EHR:

~67% AUC*

ZCoR: ~87%

HrEF

ILD

TimestampedDiagnostic procedural codes & prescriptions

ICD

Primary care

Large Science Models:

A General Framework for Generative Modeling of Microbiome-Metabolome systems

ZeBRA: Method for Individualized Future-Risk State Inference and Patient-Journey Phenotyping From Routine Longitudinal Health Data

ZeBRA Publications

Predictive Performance (PF)*

High AUC across high and low risk sub-cohorts

Interstitial Lung Disease (ILD)

US prevalence: 1 in 500 (higher by 1.5X-2X in KY)

ILD

Natural history of ILD progression: Predicting key events in patient journey

High AUCs.

High PPV: 1-2 false positive for every true flag

Interstitial Lung Disease (ILD)

US prevalence: 1 in 500

Case Study: University of Kentucky Heath-care (UKHC)

ZeBRA Model Family (expanding list)

Suicidal Ideation and Self-harm Attemps: Estimating risk from routine EHR

High AUCs.

High PPV: 1-2 false positive for every true flag

Suicidal Ideation and Self-harm Attemps: Estimating risk from routine EHR

Suicidal Ideation and Self-harm Attemps: Estimating risk from routine EHR

High AUCs.

High PPV: 1-2 false positive for every true flag

How?

This distance is "special"

What is a healthy Microbiome/ Metabolome?

Can gut microbiota pre-empt developmental deficits?

High Abundance Classes

Complex Interdependecies Exist over time and taxonomic strata

Purely wet-lab investigations are insufficient to understand complex ecosystems

We need to scale up!

We need a digital twin

of the dataset

Learn from uncurated data

understand non-trvial cross-dependencies between entities, without prior knowledge

can explore valid perturbations of the overall system

can identify deviations from "normal" behavior

Can model out-of-box targeted and untargeted metabolites (or uncharacterized microbes)

Can microbial assay from gut actionably

pre-empt developmental deficit?

Study Set-up

Predicting neurodevelopmental deficits

Forecasting ecosystem trajectories

Build classifiers

Ability to "fill in" missing data is equivalent to making trajectory forecasts

Forecast ecosystem fluctuations

Large Science Models

Estimate the joint distribution of all variables

Brook’s lemma*

perfect inference of these conditionals gives us a unique joint

The central problem of ML is computing joint distributions or estimates thereof

hard!

How Complex Are These Models?

Risk

Geometric Interpretation

LSM-based Risk

How different are the individual estimators for typical and dysbiotic models?

Forecasting of class abundance variations both

in-sample (\( R^2 \ge 95\%\))

and out-of sample (\(R^2\ge 72\%\))

Forecasting of class abundance variations both

in-sample (\( R^2 \ge 95\%\))

and out-of sample (\(R^2\ge 72\%\))

moving the top drivers impact risk of deficit maximally

Understanding Key Drivers via LSM perturbations

Just add those microbes back?

Network Interpretations? We see clear differences between two cases

Integrating Clinical Features

IBD Metabolomics: Preliminary Results

Large Science Models

Full Example of Hyperlinked Trees

Metabolomics

Metabolomics

Metabolomics

Insight: the discriminating hypersurface is 2d (almost 1d)

Using all samples of metabolite profiles

Not using clinical phenotypes and replicates

ASD Metabolics: Preliminary Results

ASD Metabolics: Preliminary Results

ASD sample profile