Perturbation-MMKPNN: Interpretable Modeling of Single-Cell Perturbation Responses

Interpretable modeling of how cells respond to genetic and chemical perturbations.

---

Overview

Perturbation-MMKPNN extends the Multimodal Knowledge-Primed Neural Network (MM-KPNN) to predict single-cell perturbation responses while keeping reasoning transparent through a pathway/TF concept bottleneck.

Black-box predictors such as scGen and CPA achieve accuracy but fail to reveal which pathways and subnetworks mediate response or provide stable explanations across datasets. Perturbation-MMKPNN constrains learning through curated biological priors (Reactome, DoRothEA, MSigDB), transforming predictions into an interpretable mapping: perturbation → regulatory program → transcriptional outcome.

Validated across scPerturb, Perturb-seq, L1000, and DrugComb, it benchmarks against scGen, CPA, and linear baselines, emphasizing attribution stability and cross-dataset generalization. This approach moves perturbation modeling from prediction to mechanistic discovery, identifying regulators of resistance and synthetic lethality with reproducibility and biological credibility.

---

Architecture

The architecture combines:

• Inputs: single-cell expression profiles and perturbation metadata
• Encoder: feature projection and dimensionality reduction
• Concept bottleneck: pathways and transcription-factor modules that mediate predictions
• Prediction head: outputs perturbed expression state or classification label
• Interpretability outputs: pathway attributions, counterfactual edits, and driver ranking

---

Workflow

1. Datasets: scPerturb, Perturb-seq, L1000, and DrugComb
2. Modeling: MM-KPNN backbone adapted for perturbation inputs
3. Evaluation: predictive accuracy, attribution stability, cross-dataset validation
4. Outputs: pathway maps, regulatory modules, and candidate resistance drivers

---

Methodology

Inputs consist of single-cell expression and perturbation metadata. An encoder projects cells into a latent space, constrained by a biological bottleneck of pathway and transcription-factor modules. The prediction head outputs either a perturbed signature or a categorical response label. Interpretability uses Integrated Gradients and Gradient×Input to obtain pathway-level attributions, and counterfactual experiments silence or boost bottleneck nodes to test causal relevance. Evaluation measures predictive accuracy, attribution stability, and transferability across datasets, perturbations, and donors.

---

Project Roadmap

Objective: Extend Perturbation-MMKPNN into a reproducible framework for analyzing single-cell perturbation responses with interpretable pathway-level outputs.

• Preprocessing: Add harmonized loaders for scPerturb, L1000, DrugComb, and Perturb-seq datasets.
• Modeling: Finalize MM-KPNN implementation with pathway and TF bottlenecks; include CPA and scGen as baselines.
• Attribution Stability: Quantify pathway-level overlap across seeds and datasets; test counterfactual pathway edits.
• Case Study: Demonstrate interpretability on a defined resistance example (e.g., EGFRi).
• Reproducibility: Package scripts and example notebooks for end-to-end execution.

---

Limitations & Ongoing Work

Sensitive to label/guide imbalance and batch effects; ongoing efforts include seed/noise robustness, leave-drug/gene-out validation, and expanded baseline comparisons (CPA, scGen, and diffusion/backbone variants).

---

Extensibility

The current MM-KPNN backbone can be replaced with transformer-based encoders or pretrained embeddings from large-scale biology models such as scGPT or Geneformer, enabling scalability while maintaining interpretability through pathway, TF, and ligand–receptor bottlenecks.

---

References

1. Norman TM et al. Perturb-seq (2019).
2. Replogle JM et al. Combinatorial single-cell CRISPR screens. Nature (2022).
3. Subramanian A et al. Connectivity Map. Science (2017).
4. Liu Q et al. DrugCombDB. Nucleic Acids Research (2020).
5. Yepes S. MM-KPNN. GitHub Repository (2025).

---

Citation

Yepes, S. Perturbation-MMKPNN: Interpretable Modeling of Single-Cell Perturbation Responses. GitHub, 2025.
DOI: 10.5281/zenodo.17189224

---

Perturbation-MMKPNN is part of the MM-KPNN framework family, extending interpretable modeling from multimodal single-cell data to perturbation-response analysis and regulatory mechanism discovery.

---