NLCD Extends Gene Causal Discovery Beyond Linear Relations
PLOS Computational Biology published NLCD on July 31, a causal-discovery method that adds nonlinear regression and conditional feature-importance testing to an established genetics framework. In the paper's simulated nonlinear benchmarks, NLCD achieved an average AUPRC of 0.76, compared with 0.60 for CIT and 0.56 for Findr; the authors also released code and reproducibility materials on GitHub.
PLOS Computational Biology published NLCD, a method for identifying nonlinear causal relationships between genes or other traits from matched genetic and observational data, on July 31. Aravind Easwar and Manikandan Narayanan designed the method to address a limitation of several causal-discovery approaches that assume the relationship between measured traits is linear.
NLCD extends the Causal Inference Test, or CIT, with nonlinear regression and a conditional feature-importance score. Its input is a triplet: a genetic variant used as an instrumental variable and two measured traits, such as the expression levels of two genes. The method applies four statistical tests and uses their maximum p-value to decide whether the data support a causal, reactive, independent, or inconclusive relationship.
Performance across nonlinear and biological data
In simulated linear benchmarks, the paper reports average area under the precision-recall curve, or AUPRC, of 0.94 for NLCD, CIT, and Findr, while MRPC reached 0.99. The separation appeared in nonlinear sine- and sawtooth-type benchmarks: NLCD averaged 0.76, compared with 0.60 for CIT, 0.56 for Findr, and 0.73 for MRPC.
The authors also tested the methods on a nonlinear subset of yeast genomic data built around known transcription-factor relationships. NLCD reached an average AUPRC of 0.82, close to CIT's 0.81 and above Findr's 0.71 and MRPC's 0.54. These results show a clearer advantage in the simulated nonlinear setting than in the yeast benchmark, so the paper does not establish that NLCD will dominate every biological dataset.
Applied to human GTEx data, NLCD identified two muscle-tissue relationships, IRF1 to PSME1 and HLA-C to HLA-T, after the paper's filtering steps. Those are method-derived causal hypotheses from observational data, not experimental confirmation that changing one gene will produce the predicted effect in vivo.
Reproducibility and practical limits
The public BIRDS Group repository contains scripts, data, figures, and result artifacts used in the study, along with instructions for running NLCD on individual triplets or batches. The implementation supports kernel ridge regression, support vector regression, and an artificial neural network as nonlinear models.
For practitioners, the release is most useful when a linear relationship assumption is difficult to defend and a suitable genetic instrument is available. Its conclusions still depend on Mendelian-randomization-like assumptions, sample size, regression quality, permutation testing, and filters for mapping artifacts. The paper's human-data results are therefore a starting point for validation and follow-up experiments rather than a substitute for intervention-based evidence.
Key Points
- 1NLCD adds nonlinear regression and conditional feature-importance testing to the CIT causal-discovery framework for genetic and trait triplets.
- 2The paper reports average AUPRC of 0.76 for NLCD on simulated nonlinear benchmarks, compared with 0.60 for CIT, 0.56 for Findr, and 0.73 for MRPC.
- 3The authors released code and reproducibility materials, while the GTEx muscle-tissue findings remain observational hypotheses requiring experimental validation.
Scoring Rationale
NLCD is a peer-reviewed, reproducible causal-discovery method with a measurable advantage on simulated nonlinear benchmarks and direct relevance to computational genomics. Its biological evidence is narrower and observational, so the result is methodologically useful without yet demonstrating broad production or clinical impact.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
