Causal Inference in Genetics: Latest developments and New Directions

Séminaire organisé par Iuliana Ionita-Laza du 18 au 23 novembre 2024

Participants

Iuliana IONITA-LAZA (Columbia University – New York, États-Unis), Hugues ASCHARD (Institut Pasteur – Paris, France), Xihong LIN (Harvard University & Laboratoire SPHère – Boston, États-Unis), Matti PIRINEN (University of Helsinki – Helsinki, Finlande), Marie VERBANCK (Institut Curie – Paris, France), Chiara SABATTI (Stanford – Stanford, États-Unis), Hongzhe LEE (University of Pennsylvania – Philadelphia, États-Unis), Charles ASSAAD (INSERM – Paris, France), Malgorzata BOGDAN (University of Wroclaw – Wroclaw, Pologne), Judith ABÉCASSIS (Inria Saclay – Palaiseau, France), Xiang ZHOU (University of Michigan – Ann Arbor, États-Unis), Sriram SANKARARAMAN (UCLA – Los Angeles, États-Unis), Brielin C. BROWN (University of Pennsylvania – Philadelphia, États-Unis), Anne BRUUN ROVSING (New York Genome Center – New York City, États-Unis)

Résumé

L’identification de relations causales et biologiquement significatives entre les gènes et les phénotypes est un problème fondamental en génétique avec d’énormes implications pour la conception de stratégies de traitement et de prévention plus efficaces pour les maladies courantes. Il s’agit d’un problème particulièrement difficile en raison de la nature observationnelle des données. Afin d’établir des liens de causalité, une étape nécessaire consiste à éliminer les effets des facteurs de confusion potentiels. Il existe plusieurs facteurs de confusion dans les études génétiques, notamment les corrélations entre les sites du génome (déséquilibre de liaison ou LD), la structure de la population, les effets polygéniques, les facteurs environnementaux et expérimentaux, les effets pléiotropes, etc. Fidèle à son habitude, le séminaire a été l’occasion unique de réunir des participants de différents domaines (dont 5 participants juniors), notamment l’inférence causale, l’apprentissage statistique et automatique, et la génomique, qui n’ont normalement pas l’occasion de se rencontrer et de discuter ensemble à une même table. Les sujets abordés lors du séminaire comprenaient la cartographie statistique fine, la découverte causale et l’apprentissage automatique, les statistiques de grande dimension et les techniques de perturbation fonctionnelle et d’édition du génome. Durant le séminaire, nous avons discuté des défis propres à chaque domaine et mis en exergue des solutions possibles via des collaborations entre ces différents domaines. Nous tenons à remercier l’équipe de « La Fondation des Treilles » pour les excellentes conditions qui ont offert un cadre idéal à des échanges riches tout au long de la semaine.

Summary

Identifying causal, biologically meaningful relationships between genes and outcomes of interest is a fundamental problem in genetics with tremendous implications for designing more efficient treatment and preventative strategies for common diseases. It is a particularly challenging problem because of the observational nature of the data. In order to establish causality, a necessary step is to remove potential confounding effects. There are several sources of confounding in genetic studies, including correlations among sites in the genome (linkage disequilibrium or LD), population structure, polygenic effects, environmental and experimental factors, pleiotropic effects etc. True to form, the seminar provided the unique opportunity to bring together participants from different areas, including causal inference, statistical and machine learning, and genomics, who may not normally get the opportunity to meet and discuss together at the same table. The topics covered during the seminar included statistical fine-mapping, causal discovery and machine learning, high-dimensional statistics and functional perturbation and genome editing techniques. We discussed the unique challenges in each area, and how to move forward by discovering connections and opportunities to collaborate across different areas. We would like to thank the team at “La Fondation des Treilles” for the excellent conditions which provided an ideal setting for rich discussions throughout the week. 

Report

Fine-mapping and knockoff-statistics

From a statistical perspective, causality can be inferred by performing conditional inference. LD confounding is a major impediment to inferring causality following genome-wide associations studies (GWAS). Fine-mapping and knockoff-based inference are two main classes of methods that can contribute to claiming causality in genetic studies by accounting for the confounding effect of LD. While fine-mapping involves a two-step approach to identify potentially causal variants at GWAS loci by jointly modeling all variants at a locus, knockoff-based inference provides a one-step framework for performing genome-wide conditional testing while controlling for LD. Matti Pirinen (University of Helsinki) provided an overview of the progress in fine-mapping over the past 10-15 years and discussed the unique challenges that large meta-analyses pose for fine-mapping due to increased heterogeneity in effects and LD structures in such studies. Chiara Sabatti (Stanford University) discussed an alternative approach based on constructing knockoff (synthetic) control variables that offer the opportunity to analyze high-dimensional data with sophisticated algorithms that capture signal, but at the same time provide precise statistical inference controlling false discovery rate. She discussed new directions of research in this area that allow for adaptive identification of potential gene-by-environment interactions, limiting the number of tests that one has to perform and increasing power.

Causal inference with graphical models

Methods for disentangling correlation from causation have been developed in statistics and machine learning with varied biological and health science applications. An intuitive way to represent causality in high-dimensional systems is by using directed acyclic graphs (DAGs), where nodes correspond to variables (e.g. variants or genes) and directed edges between nodes represent causal relationships. Causal discovery is the area of inferring these causal DAGs from observational data under specific assumptions that can be informed by domain knowledge. However, this is a challenging problem due to the combinatorial nature of the search space, the high-dimensionality of the data and the need to account for unmeasured confounding.

Several participants addressed this difficult problem from different perspectives. Charles Assaad (Inserm and Sorbonne Université) introduced Summary Causal Graphs (SCGs), an abstraction of DAGs that are easier to construct but allow for cycles and vertices that may not correspond directly to individual random variables. He described novel tools for causal inference using SCGs, even in the presence of cycles and hidden confounding. Gosia Bogdan (University of Wroclaw) described a technique for identifying the graphical model structure by regularizing the partial correlation matrix. She showed that this technique has a much larger ability to properly identify hub structures than methods that regularize the precision matrix.

Hub structures in graphical models are important for identifying key biological processes that when dysregulated can have widespread effects on the system leading to disease.

Interventional data can help learn the causal structure by improving the identifiability of causal models and by attenuating the bias due to unmeasured confounding. Sriram Sankaraman (University of California, Los Angeles) discussed the problem of inferring gene regulatory networks (GRNs – which can be represented as DAGs) motivated by new biological assays like Perturb-seq. He introduced dotears, a continuous optimization framework which leverages both observational and interventional data to infer GRNs that leads to consistent estimation of the true underlying DAG and scales to high-dimensional perturbation screens. Similarly, Brielin Brown (University of Pennsylvania) discussed a novel approach to learning causal networks that leverages large-scale intervention-response data, and showed how it can be used to study the regulatory architecture of blood traits. He also discussed many remaining challenges in causal structure reconstruction, including the need for uncertainty quantification in the estimated DAG structure.

Mendelian randomization and mediation analyses

Another statistical approach to infer causal relationships is via mendelian randomization (MR) which is an instrumental variable (IV) approach that uses genetic variants as instruments and attempts to reduce the bias induced by unmeasured confounding. The validity of MR results is heavily dependent on several  assumptions that can be violated in practice, including the need for strong and independent genetic variants as IVs, and the exclusion restriction assumption (no pleiotropy). Xihong Lin (Harvard University) discussed the Debiased Estimating Equation Method (DEEM), an MR approach that incorporates a large number of correlated weak-effect and invalid variants. DEEM effectively eliminates the weak IV bias but also improves the statistical efficiency of the causal effect estimation by leveraging information from many correlated variants. Sriram Sankaraman (UCLA) presented MR-Twin, a MR method that leverages family data to test for causal relationships while being provably robust to population stratification. Xiang Zhou (University of Michigan) presented a method, MR with automated instrument determination (MRAID), that explicitly models both uncorrelated and correlated horizontal pleiotropy.

A related methodology is mediation analysis which aims to understand the mechanism of action of a treatment on an outcome and identify potential intermediate intervention targets. For example, a typical question in mediation analysis is whether the smoking effect on lung function is mediated through the methylation of a gene. To answer this question, causal mediation analysis decomposes the total causal effect into direct vs. indirect effects, i.e. through a mediator. A particular challenge in this area is developing methods in the context of high-dimensional mediation analysis. Judith Abecasis (INRIA) discussed parametric and non-parametric estimation approaches in the context of high-dimensional mediators such as imaging biomarkers. Xihong Lin (Harvard University) discussed the high-dimensional mediator setting as well by delving into the complexities of testing mediation effects, especially addressing the challenge of testing a large number of composite null hypotheses and the blessing of the large number of mediators in this case. Hongzhe Lee (University of Pennsylvania) also discussed the case where the set of potential mediators is high-dimensional by assuming that the causal structure involving genetic variants, confounders, mediators and outcome is described by a (possibly unknown) DAG. He presented a method inspired by IDA (Interventional calculus when the DAG is Absent), for estimating mediation effects in high-dimensional observational data.

Marie Verbank (Institut Curie) presented a method that leverages pleiotropy to pinpoint direct effects and build causal networks for top GWAS variants. Using data on 813,290 significant variant-trait effects from GWAS on 60 traits in UK Biobank, she showed that the resulting causal networks are consistent with established biological mechanisms, and that only a small proportion of genetic effects are direct effects.

Functional perturbation and genome editing techniques

CRISPR-based perturbation experiments have emerged as transformative tools for understanding genome function and genetic regulation. There is clearly a need to validate predictions from statistical and machine learning models as described above with experimental data such as CRISPR-CAS functional perturbation and genome editing. For fine-mapping, Anne Bruun Rovsing (New York Genome Center) reported findings from a large-scale study integrating biobank-scale GWAS, massively parallel CRISPR screens, and single-cell sequencing, aiming to connect noncoding GWAS variants with their target genes. This was achieved using a novel approach, STING-seq, where statistical fine-mapping was performed on GWAS of blood traits to identify candidate cis-regulatory elements (CREs), which were systematically silenced to uncover their target genes. These findings demonstrate the power of CRISPR-based technologies to address the variant-to-function challenge in GWAS interpretation, providing critical insights into the molecular mechanisms underlying complex traits and diseases. Similarly, for approaches that try to identify specific mechanisms such as causal mediation or causal discovery using DAGs, experimental techniques provide opportunities to improve DAG reconstruction (as discussed in talks by Sriram Sankaraman and Brielin Brown) and also to validate the most promising findings by performing interventions such as gene knockouts along specific biological pathways.

Conclusions

Gaining mechanistic understanding in genetics requires collaborations across diverse areas including causal inference, statistical learning and genomics. Progress requires a combination of statistical, algorithmic and technological developments. This seminar provided a great opportunity to discuss avenues to address unique challenges in developing methods and applications for performing causal inference in genetics.