For the creation of new drugs, understanding how genes interact with one another is crucial. Researchers can find new potential drugs that could be utilised to treat diseases by looking at gene-gene interactions.
 Scale-based research on gene-gene interactions proved challenging in the past. It was necessary to measure the expression of thousands or even millions of genes in each individual cell. Recently, high-throughput sequencing technology has made it possible to detect gene expression at this level. These advances have led to the development of new methods for inferring causal gene-gene interactions. These methods use single-cell gene expression data to identify genes that are statistically associated with each other. However, it is difficult to ensure that these associations are causal, rather than simply correlated. So, the CausalBench Challenge seeks to improve our ability to understand the causal relationships between genes by advancing the state-of-the-art in inferring gene–gene networks from large-scale real-world perturbational single-cell datasets. This information can be used to develop new drugs and treatments for diseases. The main goal of this challenge is to improve one of two existing methods for inferring gene-gene networks from large-scale real-world perturbational single-cell datasets: GRNBoost or Causal Discovery from Interventional Data (DCDI). This paper will describe three small improvements to the DCDI baseline.

@CausalBench Challenge 2023 - Minor Improvements to the Differentiable Causal Discovery from Interventional Data Model

This brief report describes an approach to modify the graph inference function for the causal bench challenge which is based on the dataset described in (Chevalley et al., 2022).

Challenge Report

We present a novel directed acyclic graph (DAG) learning method for data generated by a linear structural equation model (SEM) and apply it to learn from gene expression data. In prior work, linear SEMs can be viewed as a linear transformation of a dense input vector of random valued root causes (as we define). In our novel setting we further impose the assumption that the output data are generated via a sparse input vector, or equivalently few root causes. Interestingly, this assumption can be viewed as a form of Fourier sparsity based on a recently proposed theory of causal Fourier analysis. Our setting is identifiable and the true DAG is the global minimizer of the $L^0$-norm of the vector of root causes. Application to the CausalBench Challenge shows superior performance over the provided baselines.

Learning Gene Regulatory Networks under Few Root Causes assumption.

Part of the difficulty of learning a gene-regulatory network from expression data is related to the fact that edges in such a network represent different interactions with a different effect size. Therefore modeling gene associations requires learning an individual function for each pair of interacting genes. This may greatly inflate the number of parameters in a model and lead to insufficient generalization. In this paper we propose a method for gene regulatory network inference, called CaTran (Causal Transformer), which avoids explicitly learning pairwise relation between genes, which allows it to significantly reduce the size of the model. The key feature of this approach is learning for each gene a low dimensional embedding and then using a self-attention mechanism to estimate its relation to other genes. Our method is applicable for both observational data and data with interventions. For the latter it implements a differentiable gene importance test and forces attention values to be in accordance with it. Because the gene regulatory network in CaTran is learned as a soft adjacency matrix, it allows sampling graphs with arbitrary number of edges based on a set threshold. Comparison of these graphs with the gene networks from databases showed that even for large graphs the edges are predicted with high precision.

CaTran: ultra-light neural network for predicting gene-gene interactions from single-cell data

The introduction of large-scale, genome-wide, single-cell perturbation datasets provides the chance to learn a full gene regulatory network in the relevant cell line. However, existing gene regulatory network inference methods either fail to scale or do not explicitly leverage the interventional nature of this data. In this work, we propose an algorithm that builds upon GRNBoost by adding an additional step that complements its performance in the presence of labeled, single-gene interventional data.
Applying BetterBoost to the CausalBench Challenge, we demonstrate its superiority over the baseline methods in inferring gene regulatory networks from large-scale single-cell perturbation datasets. Notably, BetterBoost exhibits significantly improved performance when non-zero fractions of labeled interventions are available, highlighting the effectiveness of our approach in leveraging interventional data for accurate gene regulatory network inference.

BetterBoost - Inference of Gene Regulatory Networks with Perturbation Data

In this challenge, we transformed the task of detecting gene pairs with causal relationships into a supervised learning problem. We constructed a dataset for all gene pairs, with initial labels determined by gene expression correlations. A LightGBM model was trained and applied to the same data for prediction. The top 1001 pairs with the highest prediction scores were selected. In local experiments, this solution achieved a 0.3779 AUC score in the RPE1 data and a 0.3265 score in the K562 data.

A Supervised LightGBM-Based Approach to the GSK.ai CausalBench Challenge (ICLR 2023)

In this write-up, we describe our solution to the 2023 CausalBench Challenge. We describe our approaches to preprocessing the data, parameterizations of DCDI and GRNBoost, and modifications to the baseline algorithms.