coreason-inference

Causal Discovery, Representation Learning, Active Experimentation, & Trial Optimization

coreason-inference is the Causal Intelligence engine of the CoReason ecosystem. Unlike probabilistic models that predict correlation, this engine is designed to uncover Mechanism and Heterogeneity. It serves as the "Principal Investigator," discovering biological feedback loops, identifying latent confounders, and optimizing clinical trials through causal stratification and virtual simulation.

Features

• Cyclic Discovery (Dynamics Engine): Uses Neural ODEs to discover feedback loops and system dynamics in biological systems (Directed Cyclic Graphs).
• Latent Phenotyping (Latent Miner): Disentangles hidden confounders using Causal VAEs to identify unmeasured variables driving outcomes.
• Heterogeneous Stratification (Causal Estimator): Estimates Individual Treatment Effects (CATE) using Causal Forests to identify Super-Responders.
• Active Experimentation (Active Scientist): Proposes physical experiments (e.g., Gene Knockouts) to resolve causal ambiguity using Information Gain heuristics.
• Protocol Optimization (Rule Inductor): Translates CATE scores into human-readable clinical protocol rules to maximize Phase 3 Probability of Success (PoS).
• Virtual Trials (Virtual Simulator): Simulates Phase 3 trials in-silico using synthetic "Digital Twins" to predict efficacy and scan for safety risks.

Installation

pip install coreason-inference

Usage

Here is a quick example of how to initialize and use the InferenceEngine:

import pandas as pd
from coreason_inference.engine import InferenceEngine

# 1. Initialize the Engine
engine = InferenceEngine()

# 2. Load your data
# Data should contain time-series or observational data
data = pd.read_csv("patient_data.csv")
variable_cols = ["Glucose", "Insulin", "HbA1c"]
time_col = "Time"

# 3. Analyze: Discover Dynamics & Latents
result = engine.analyze(
    data=data,
    time_col=time_col,
    variable_cols=variable_cols,
    estimate_effect_for=("Insulin", "Glucose")
)

# 4. Inspect the Causal Graph
print(f"Discovered Graph Edges: {result.graph.edges}")
print(f"Detected Loops: {result.graph.loop_dynamics}")

# 5. Optimize for Heterogeneity (Identify Super-Responders)
# Estimate CATE for a specific treatment
engine.analyze_heterogeneity(
    treatment="Insulin",
    outcome="Glucose",
    confounders=["Age", "BMI"] + list(result.latents.columns)
)

# Induce rules to find the best subgroup
optimization_output = engine.induce_rules()

print("Recommended Protocol Criteria:")
for rule in optimization_output.new_criteria:
    print(f" - {rule.feature} {rule.operator} {rule.value} ({rule.rationale})")

print(f"Projected Probability of Success: {optimization_output.optimized_pos:.2f}")

Documentation

For detailed requirements and architectural philosophy, please refer to the Product Requirements Document.