BGPT: Paper Review: Convergent and non-additive impact of schizophrenia risk genes in human neurons

Fuel Your Discoveries

Quick Explanation Copied

Core finding

Across human iPSC-derived glutamatergic neurons, schizophrenia-associated eGenes show shared (convergent) downstream transcriptional directionality, and combinatorial perturbations are often sub-additive (smaller-than-expected changes) rather than linear sums of single-gene effects—suggesting redundancy/saturation among functionally similar eGenes.

Evidence: pooled CRISPRa with single-cell readout identifies convergent downregulated gene sets, while arrayed combinatorial perturbations of 15 eGenes within synaptic/regulatory groups yield sub-additive non-additive transcription across ~16.8% and ~20.2% of the transcriptome, respectively.

Long Explanation

Paper: Convergent and non-additive impact of schizophrenia risk genes in human neurons

Preprint date (provided): March 30, 2022. DOI: 10.1101/2022.03.29.486286

Main methodological triad: (i) eGene prioritization via transcriptomic/epigenomic imputation and colocalization, (ii) pooled vs arrayed pooled CRISPRa/sh perturbations in iPSC-derived iGLUT glutamatergic neurons, (iii) convergence and non-additivity tested using meta-analysis and an explicit additive model comparison.

1) What the paper claims (mechanism-level map)

Hypothesis scaffold

Convergence: multiple schizophrenia eGenes produce downstream transcriptional changes with the same direction (e.g., consistent up vs down across perturbations).
Non-additivity: when functionally similar eGenes are perturbed together, observed combined transcription deviates from a linear additive expectation computed from single-gene perturbations.
Proposed bridge: if eGenes share downstream signatures (redundancy), then combined perturbations can saturate or compress the magnitude of directional gene-expression effects (sub-additivity).

2) Visual evidence from the paper (figures reconstructed from reported numeric summaries)

2A) Sub-additivity is strongest for synaptic + regulatory co-perturbations

2B) Convergence genes overlap strongly with non-additive genes (synaptic + regulatory sets)

3) Internal logic & analysis pipeline (what was done, and what it implies)

3A) Pooled CRISPRa single-cell screen: convergence definition & control for heterogeneity

Model system: iPSC-derived glutamatergic neurons (iGLUTs) are used for perturbations, and the pooled approach uses single-cell RNA-seq readouts with direct sgRNA detection.
Convergence metric: the authors use meta-analysis (METAL) and a heterogeneity test (Cochran’s Q) to define genes with shared effect direction across eGene perturbations and low heterogeneity.
QC / assignment: guide identity in cells is assigned using a weighted nearest neighbor framework (WNN), integrating sgRNA identity and gene expression to reduce misassignment risk.

Skeptical check

Even with WNN and heterogeneity filtering, convergence is still sensitive to (i) which eGenes are included, (ii) the directionality constraint (up-only CRISPRa vs both directions in arrayed screens), and (iii) the specific DEG calling thresholds and normalization strategy in single-cell RNA-seq.

4) How the additive model test is constructed (and where it can mislead)

4A) Expected additive vs measured combinatorial

The paper computes an expected additive differential expression for a given combinatorial set by summing single-eGene DE contrasts.
Non-additive effects are then identified by difference between measured combinatorial DE and the expected additive model.
The authors further interpret predominance of smaller-than-expected directional magnitudes as sub-additivity.

4B) Why convergence could cause non-additivity (plausible mechanism)

The paper’s logic is that if multiple eGenes push the same downstream transcriptional direction (convergence), then simultaneous perturbations can “hit” overlapping downstream programs. In that case, the marginal contribution of adding another eGene can shrink—producing sub-additive behavior in the additive-comparison framework.

Main skeptical vulnerabilities

Thresholding & power: “non-additive” is a statistical construct; if effect sizes are small or sample size is limited, the additive comparison can preferentially detect certain gene classes (e.g., those with larger variance or more stable expression).
Cell-state and maturation: combinatorial conditions may shift maturation timing or cell stress responses, changing transcriptomes in ways that mimic interaction-like deviations (the authors discuss activity-dependent maturation as a concern in pooled designs).
Model mismatch: an additive model assumes that the expected joint effect is the sum of independent single-gene contrasts; real gene perturbations can interact via chromatin accessibility, feedback loops, or network-level compensation, meaning “sub-additivity” might be caused by feedback saturation rather than redundancy of target pathways.

5) Drug-reversal module (prediction → in vitro opposition)

5A) Concept

The paper attempts to prioritize drugs by identifying small molecules predicted to reverse transcriptional signatures of convergent network node genes and then tests whether drug treatment opposes schizophrenia eGene perturbation effects in neurons.

Skeptical note on pharmacological inference

Signature reversal from gene-expression connectivity maps is not causal; it can reflect shared downstream pathway modulation but still not clarify which upstream eGene targets are functionally “reversed.” Also, the paper itself acknowledges experimental-context limitations of in vitro systems and interpretive caution when mapping GWAS/imputed targets to human physiology.

6) Evidence quality & limitations (critical, science-focused)

6A) What looks strong

Two complementary perturbation designs: pooled CRISPRa (single-cell sgRNA-aware readout) and arrayed combinatorial perturbations reduce the risk of drawing conclusions from a single perturbation/readout modality.
Explicit modeling: the additive expectation framework is formal and testable (not merely descriptive overlap).
Single-cell perturbation assignment: using direct sgRNA detection plus WNN-based assignment is a methodological attempt to reduce misassignment in pooled screens.

6B) Key limitations / blind spots the paper flags

Partial eGene coverage: only a small subset of schizophrenia eGenes is perturbed compared with the much larger polygenic architecture; therefore, “global non-additivity” across all schizophrenia risk variants is not directly established.
Context dependence: convergence and interactions could vary across neuronal/glial types and developmental stages; the authors argue future studies should expand across more cell types and stages.
In vitro vs in vivo mismatch: CRISPR perturbations in cultured neurons do not recreate the exact combination of inherited variants within individuals.
Non-linear dosage responses: the authors note uncertainty whether gradual changes in gene dosage yield linear vs non-linear transcriptional effects.

Additional critical blind spots (not fully resolved by what’s shown)

Stress/inflammation confounds: single-cell CRISPR in lentiviral systems can induce stress programs that would create broad convergence in downregulated housekeeping/secretory pathways; the paper reports no strong evidence of cessation of activity/survival but the strength of those negative controls is always a key question.
Batch/donor effects in combinatorial screens: the paper states donor status did not significantly impact eGene perturbation efficacy; however, transcriptomic synergy/sub-additivity can still be influenced by donor-specific baseline states and variance structure (a common issue in perturb-seq analyses).

7) What would most likely disprove the central conclusion?

Using a substantially larger and less preselected set of schizophrenia eGenes (including ones with more graded or different regulatory directions), combinatorial perturbations would still need to show: (i) convergence and (ii) systematic deviation from additive expectations specifically for functionally similar groups.
Alternate statistical formulations of additive/non-additive effects (e.g., different normalization, interaction terms, or modeling of cell-state trajectories) would need to recover the same direction: predominantly sub-additive behavior paired with convergence overlap.
Orthogonal experimental controls must show that convergence/non-additivity is not driven by shared lentiviral transduction burden or cell stress programs; stronger evidence would include matching perturbation burden across conditions with orthogonal readouts beyond RNA-seq.

8) Quick BGPT navigation (deeper dives)

Author Reviews on BGPT

Feedback:

Updated: April 22, 2026