Ishanu Chattopadhyay PRO
ML | Data Science Biomedical Informatics | Social Science | Assistant Professor
Pan-genomic Digital Twins for Predictive Diagnosis of Pulmonary Fibrosis
Ishanu Chattopadhyay, PhD
Assistant Professor of Internal Medicine
Institute of Biomedical Informatics
University of Kentucky
Data
19000 samples
1.2M SNPS, SOCIAL, CLINICAL
\{
11K SNPS, SOCIAL, CLINICAL
the variable of interest (among the ~11K) is the diagnosis state for PF: PFdx
Risk distribution for samples with positive and negative PF diagnoses
Looking ahead: OOS validation
Take oos samples, sensor PFdx and J84 clinical data, and calculate risk as defined in Eq. 1
Results
n=7000
Data
Digital Twin of Data
| x1 | x2 | x3 |
|---|
| xi |
|---|
\Phi_i:\prod_{j \neq i} \Sigma_j \rightarrow \mathcal{D}(\Sigma_i)
We infer a "recursive forest" that yields a generative model of the data
Predict the distribution of a variable (\(x_i\)) as function of all the other available variables
Captures the cross-talk and epistetic effects
\Phi_i:\prod_{j \neq i} \Sigma_j \rightarrow \mathcal{D}(\Sigma_i)
Viral genome example
\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"
$$J \textrm{ is the Jensen-Shannon divergence }$$
q-distance
a biologically informed, adaptive distance between samples
Smaller distances implies high probability of a "valid perturbation"
Metric Structure
Tangent Bundle
geometry
dynamics
\theta(x,y) \sim \log Pr(x \rightarrow y)
\theta
Digital Twin of Transgenome
(limited to 10K PF correlates)
parameters: 975050
variables: 11k
samples: 12k
PFdx
MUC5b
MUC5b
JHU_11.119702210_G
PFdx
Inferred path from MUC5b promoter to PFdx outcome
inferred network driving PF diagnosis
This is a fraction of the emergent network among our variables that "drives" the PFdx variable
inferred network driving PF diagnosis
Some qualitative match with what is know. Many "new" relationships.
Need to inspect for mechanistic meaning
Building the Classifier for PF dx
- Leverage the generative property
- The "training set" is imperfectly specified (not all PF positive are diagnosed, some clinical diagnoses might be "placeholders" or wrong)
- No-PF diagnoses does not mean "healthy"; it is a highly heterogeneous set
- PF diagnoses also does not mean the rest of the data is homogenous
We make "anchor" perturbation samples
PF dx
no PF dx
- set \(PFdx=X\) and perturb to generate "non-PF" samples
- set \(PFdx=J84111\) and perturb to generate "PF" samples
\rho(s) = \sum_{i=1}^{m}\frac{ \theta(s,H_i)}{\theta(s,J_i)}
Risk of PF diagnosis
H_i
J_i
(1)
Convex hulls in Metric embedding of perturbation vectors
We make two sets of "valid perturbation" samples
Risk distribution for samples with positive and negative PF diagnoses
Out-of-sample validation
Take oos samples, sensor PFdx and J84 clinical data, and calculate risk as defined in Eq. 1
Results
n=7000
PF-DT
By Ishanu Chattopadhyay
PF-DT
AI for medicine
