Search bioRxiv⌕ Search
Home/Archive/bioRxiv

bioRxiv · 10.1101/2022.07.18.500549

Causal feature selection using a knowledge graph combining structured knowledge from the biomedical literature and ontologies: a use case studying depression as a risk factor for Alzheimer's disease

Malec, S. A.·Taneja, S. B.·Albert, S. M.·Shaaban, C. E.·Karim, H. T.·Callahan, T. J.·Munro, P. W.·Levine, A. S.·Boyce, R. D.
View original source ↗Export metadata ↓

Abstract

BackgroundCausal feature selection is essential for estimating effects from observational data. Identifying confounders is a crucial step in this process. Traditionally, researchers employ content-matter expertise and literature review to identify confounders. Uncontrolled confounding from unidentified confounders threatens validity, conditioning on intermediate variables (mediators) weakens estimates, and conditioning on common effects (colliders) induces bias. Additionally, without special treatment, erroneous conditioning on variables combining roles introduces bias. However, the vast literature is growing exponentially, making it infeasible to assimilate this knowledge. To address these challenges, we introduce a novel knowledge graph (KG) application enabling causal feature selection by combining computable literature-derived knowledge with biomedical ontologies. We present a use case of our approach specifying a causal model for estimating the total causal effect of depression on the risk of developing Alzheimers disease (AD) from observational data. MethodsWe extracted computable knowledge from a literature corpus using three machine reading systems and inferred missing knowledge using logical closure operations. Using a KG framework, we mapped the output to target terminologies and combined it with ontology-grounded resources. We translated epidemiological definitions of confounder, collider, and mediator into queries for searching the KG and summarized the roles played by the identified variables. We compared the results with output from a complementary method and published observational studies and examined a selection of confounding and combined role variables in-depth. ResultsOur search identified 128 confounders, including 58 phenotypes, 47 drugs, 35 genes, 23 collider, and 16 mediator phenotypes. However, only 31 of the 58 confounder phenotypes were found to behave exclusively as confounders, while the remaining 27 phenotypes played other roles. Obstructive sleep apnea emerged as a potential novel confounder for depression and AD. Anemia exemplified a variable playing combined roles. ConclusionOur findings suggest combining machine reading and KG could augment human expertise for causal feature selection. However, the complexity of causal feature selection for depression with AD highlights the need for standardized field-specific databases of causal variables. Further work is needed to optimize KG search and transform the output for human consumption. HighlightsO_LIKnowledge of causal variables and their roles is essential for causal inference. C_LIO_LIWe show how to search a knowledge graph (KG) for causal variables and their roles. C_LIO_LIThe KG combines literature-derived knowledge with ontology-grounded knowledge. C_LIO_LIWe design queries to search the KG for confounder, collider, and mediator roles. C_LIO_LIKG search reveals variables in these roles for depression and Alzheimers disease. C_LI Graphical abstract O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=127 SRC="FIGDIR/small/500549v4_ufig1.gif" ALT="Figure 1"> View larger version (36K): org.highwire.dtl.DTLVardef@1c3f417org.highwire.dtl.DTLVardef@1ce1703org.highwire.dtl.DTLVardef@148d645org.highwire.dtl.DTLVardef@eb55e7_HPS_FORMAT_FIGEXP M_FIG C_FIG Statement of SignificanceO_ST_ABSProblemC_ST_ABSExtensive knowledge is required to identify confounders for estimating total effects in observational settings. The literature is too vast for humans to process. Bias may remain from not adjusting on unknown confounders or erroneously adjusting on a collider or mediator. What is already knownStructured literature-derived knowledge is often useful but is noisy. What this papers addsWe present a biomedical knowledge graph linking literature-derived and ontology-ground knowledge of biochemical processes with definitions of clinical disease. We search the KG to distill a model for estimating the (total) effect of depression on Alzheimers disease from observational data.

Source connections

Explore related subjects

systems biologyPreprint

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Malec, S. A., Taneja, S. B., Albert, S. M., Shaaban, C. E., Karim, H. T., Callahan, T. J., Munro, P. W., Levine, A. S., Boyce, R. D.. 2022-07-20. Causal feature selection using a knowledge graph combining structured knowledge from the biomedical literature and ontologies: a use case studying depression as a risk factor for Alzheimer's disease. https://doi.org/10.1101/2022.07.18.500549

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related preprints

bioRxiv2026

Homogeneous network inference for subtype-stratified master regulator analysis in cancer

Master-regulator analysis of 34 TCGA tumour cohorts (cancer x molecular subtype; n between 27 and 562 tumours) requires one regulatory network per cohort. Mutual-information (MI) networks inferred with ARACNe3 are the usual choice, but they are not available for the smallest cohorts; filling the gap with correlation networks would mix two methods inside the very comparison between subtypes that the analysis is meant to make. We therefore infer every network with a single procedure, fixed-size Spearman regulons (FSR): rank correlation between regulator and target, a fixed effect-size threshold rho >= 0.2, the per-target data-processing-inequality (DPI) pruning implemented in corto, bootstrap support, and truncation to the 100 best-supported targets per regulator. We quantify what this choice concedes and what it provides, labelling each claim as a standard result, a heuristic, or an empirical result. Rank correlation is more powerful than the MI estimators we simulated for n < 200 but cannot see non-monotone dependence, which in a synthetic worst case costs at most about 0.05-0.10 of the master regulators recovered; the per-target DPI is a heuristic with a measurable disagreement rate with the MI-based one; and pairwise dependence is not direct interaction, for MI and correlation alike. Where both methods are feasible (n >= 95) they overlap in 0.7 of the top-20 regulators, the same as two filterings of the same ARACNe3 network, although the rank correlation of the full score vector shows a measurable method effect in the tail. Subsampling large real cohorts indicates that networks inferred from fewer than 60 tumours do not reliably reproduce their own cohort's master regulators, that 60-100 is a grey zone, and that a cancer-wide network does not rescue small subtypes. We present FSR as an adequate and uniform choice for multi-cohort analysis, not as a statistically superior one.

systems biology↗
bioRxiv2026

Low-fidelity simulator selection for multi-fidelity inference of epithelial vertex models

Vertex models link cell division, neighbour exchange, and intercellular forces to epithelial tissue form through parameters such as junctional line tension and perimeter contractility. Bayesian inference of these parameters from data quantifies their uncertainty, but for stochastic models of proliferating tissues it requires many costly simulations. Multi-fidelity simulation-based inference reduces this cost by training on many runs of a cheaper, less accurate ('low-fidelity') simulator before refining with a few full ('high-fidelity') runs. Low-fidelity versions can be constructed in several ways, for example by stopping simulations early, coarsening the timestep, or averaging fewer replicates, but it is unclear which to choose before committing to costly training. To address this, we introduce PRISM (Paired Ranking Index for Simulator Matching), a screen that compares each candidate with the full simulator on paired runs, scoring agreement in cell-packing statistics against cost. For a vertex model of a proliferating tissue, PRISM flags stopping early as unreliable before training; once trained, this choice can produce confidently wrong parameter estimates. By contrast, a coarser timestep, within the range tested, is cheap and accurate; while fewer replicates are accurate but too costly. Using a screened low-fidelity simulator, multi-fidelity inference outperforms approximate Bayesian computation and neural-network inference trained on full simulations alone, at matched cost and with well-calibrated uncertainty. The same pattern recurs for a second vertex model fitted to published Drosophila pupal-wing imaging data, although there stopping halfway proves safe. Together, our findings suggest that which approximations are safe depends on the model, and that screening them before training makes multi-fidelity inference of tissue mechanics practical.

systems biology↗
bioRxiv2026

Calibrating developmental dynamics from single-cell atlases: lineage paths outrun least-action routes

Single-cell atlases infer developmental dynamics from destructive snapshots, yet how far realized lineage paths depart from such inferred dynamics has never been measured with controls. We introduce LADD (Least-Action Developmental Dynamics), a measurement framework that learns a stochastic developmental vector field from time-series single-cell RNA-seq and computes minimum-action and minimum-time paths, together with a three-layer calibration protocol that separates genuine path deviation from the generic cost of centroid measurement, and we compare its forecasting component with optimal-transport and flow-matching methods. In the lineage-resolved Caenorhabditis elegans atlas, the raw action ratio of observed lineage-centroid trajectories to optimal paths is 3.79, yet forward-simulated ensembles of the learned model itself score 2.07, leaving a genuine excess of about two-fold. In zebrafish (Danio rerio), the observed ratio sits entirely at the model's floor. Three mechanisms that could explain the residual deviation (detours, flow-opposition, and fate-map complexity) are each excluded; what remains is speed, with lineages moving 30-43% faster along the learned flow. We report validity boundaries for temporal resolution, cross-species comparison, and energetic interpretation, delimiting where action-based trajectory comparisons are meaningful.

systems biology↗