Counterfactual Resimulation for Causal Analysis of Rule-Based Models
Authors: Jonathan Laurent, Jean Yang, Walter Fontana
IJCAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We introduce an algorithm to generate counterfactual traces and provide an efficient implementation for the Kappa language. In Appendix C, we provide a benchmark of our implementation on a scaled-up version of our toy model. The average slowdown per event compared to the Kappa simulator does not exceed 50% for a variety of interventions. |
| Researcher Affiliation | Academia | Jonathan Laurent1, Jean Yang1, Walter Fontana2 1 Carnegie Mellon University 2 Harvard Medical School {jonathan.laurent, jyang2+}@cs.cmu.edu, walter fontana@hms.harvard.edu |
| Pseudocode | Yes | We introduce Algorithm 1, a variation of the Doob-Gillespie algorithm, to sample a counterfactual trace efficiently given a reference trace τ and an intervention ι. We call it counterfactual resimulation, since it works by going through every event of τ, resimulating only those parts of τ that are affected by ι. |
| Open Source Code | No | The Kappa simulator ... is efficiently implemented for rule-based models in Kappa as described in [Danos et al., 2007b; Boutillier et al., 2017]. Our implementation is guaranteed to either produce or consume an event at each iteration. |
| Open Datasets | No | The paper does not mention using a publicly available or open dataset. It refers to a "toy model" and simulated data. |
| Dataset Splits | No | The paper does not provide specific details on train, validation, or test dataset splits. It describes simulation traces rather than traditional dataset partitioning. |
| Hardware Specification | Yes | We consider four kinds of intervention defined in terms of an event e0 = ((r0, ξ0), t0): 1. Singular block: ... Our experimental setup consists of the model in Figure 2 comprising 104 substrates and 104 kinases. Every agent starts out in an unbound and unphosphorylated state. The simulation is stopped as soon as the system contains more phosphorylated than unphosphorylated substrates. We then proceed as follows. (i) We first generate nτ = 10 reference traces and record the CPU time T it took the Kappa simulator to generate each of them on a personal computer with a 2.7GHz Intel Core i5 processor and 16GB of random-access memory. |
| Software Dependencies | No | The paper mentions the "Kappa language" but does not specify version numbers for Kappa or any other software libraries used in the implementation. |
| Experiment Setup | Yes | Our experimental setup consists of the model in Figure 2 comprising 104 substrates and 104 kinases. Every agent starts out in an unbound and unphosphorylated state. The simulation is stopped as soon as the system contains more phosphorylated than unphosphorylated substrates. |