Skip to content

algorithm.md

Latest commit

 

History

History
48 lines (37 loc) · 4.5 KB

algorithm.md

File metadata and controls

48 lines (37 loc) · 4.5 KB

C-STDP: Temporal Event Mining of Host Immune Response

Overview

Understanding the host immune response to infection requires more than a snapshot of differentially expressed genes; it demands a temporal map of activation events. This project pivots the Causal Spike-Timing Dependent Plasticity (C-STDP) algorithm—originally biologically inspired—into a purely statistical tool for mining temporal precedence in gene expression. The goal is to reconstruct the "infection cascade": the precise order in which immune factors activate, allowing us to distinguish early drivers from downstream effectors.

Input Data

The pipeline accepts bulk RNA-seq expression matrices (LogCPM or TPM). Crucially, the algorithm treats the sample axis as a pseudo-time dimension, assuming that the samples represent a progression of the infection state or a time-series experiment.

Immune-Aware Preprocessing

To focus the analysis on relevant biological signals and reduce computational noise, we implement a strict immune-aware filtering step:

  1. Harmonization: Gene identifiers (Ensembl) are stripped of version suffixes and harmonized to standard Gene Symbols using mygene.
  2. Universe Construction: We define an "Immune Universe" by aggregating pathways from KEGG, Reactome, and GO Biological Process that contain keywords such as "cytokine", "interferon", "inflammation", and "viral defense".
  3. Filtering: Only genes present in this universe are retained for analysis. This reduces the feature space (from ~25k to ~4k genes) to a biologically coherent subset, improving the signal-to-noise ratio for detecting regulatory interactions.

Sudden Onset Detection

Traditional amplitude-based thresholding often fails in log-transformed RNA-seq data due to dynamic range compression. To address this, we implemented a Derivative-based Z-score Encoding:

  1. We compute the temporal derivative (rate of change) of gene expression.
  2. We apply gene-wise Z-score normalization to these derivatives.
  3. A "Sudden Onset Event" is defined as a normalized derivative exceeding a strict threshold (Sigma > 1.5).

This approach identifies genes that undergo a rapid acceleration in expression—interpreted biologically as the moment of transcriptional activation or "burst"—rather than just high absolute abundance.

C-STDP Learning Rule

The core learning algorithm leverages the principle of temporal precedence. For any pair of genes (A and B), we analyze the relative timing of their onset events:

  • If Gene A consistently activates before Gene B within a short time window, the connection A $\rightarrow$ B is strengthened (Potentiation).
  • If Gene A activates after Gene B, the connection is weakened (Depression).

This creates a directed weight matrix where the magnitude of a link represents the strength and consistency of the temporal delay between the two genes.

Network Extraction

From the learned weight matrix, we extract a Gene Regulatory Network (GRN):

  • Regulators: Defined as genes with high "Out-Degree" (sum of outgoing weights). These represent nodes that consistently precede a large number of other immune factors.
  • Sparsity: We apply a strict sparsity constraint (e.g., retaining only the top 10-25% of strongest links or subsetting to the most active genes) to filter out spurious correlations and focus on the backbone of the immune response.

Validation Strategy

The algorithm's reliability is bolstered by:

  • Synthetic Validation: Recovery of ground-truth networks in simulated data.
  • Permutation Tests: Verifying that the inferred structure vanishes when temporal order is shuffled.
  • Negative Controls: Ensuring unrelated housekeeping genes do not appear as top regulators.

Interpretation Boundaries

While powerful, this method infers temporal precedence, which is a proxy for, but not identical to, biological causality.

  • It can identify: Likely upstream regulators, activation sequences, and temporal modules.
  • It cannot claim: Direct physical binding (e.g., transcription factor binding) or therapeutic efficacy without experimental validation.
  • Results are dependent on the resolution of the input time-series; very fast molecular events occurring between sampling points cannot be resolved.

Summary

By rigorously filtering for immune-relevant genes and focusing on the rate of change rather than absolute levels, C-STDP provides a novel lens for infectious disease data. It moves beyond "what changed" to "when it changed," enabling the reconstruction of the infection cascade from transcriptomic data.