Skip to content

How the valproic acid and lithium "GxT in a dish" paper fits dismech

Written 2026-10-05 against dismech main at commit 681ac38851. Revised 2026-10-09 against main at bd855d0346, after the paper was curated, and again against 4cca8cef13 to correct what it said about the issue tracker.

Status of this report

The paper is now in dismech. Fetal_Valproate_Syndrome.yaml was added in #13604, merged 2026-10-06, and cites it 13 times. The five gaps below were written before that, and this revision records what each one became.

Gap State
1. Fetal valproate syndrome has no entry Closed by #13604
2. Autism valproate link has no intermediates Open, reframed as a cross-reference. Issue #13779
3. Bipolar Disorder has no model sections Open. Issue #13600
4. Neural tube defect entries contradict the new entry Open. Issue #13778
5. Schizophrenia and Epilepsy have no model sections Dropped as not worth the work

The schema gaps at the end of this report were confirmed by how #13604 had to record the model, and are the part of this report that has not been acted on.

The paper

Valone JM, Le BD, Matoba N, Mory JT, Wolter JM, Love MI, Stein JL. "Assessing molecular gene by treatment interactions using a population of neural progenitors exposed to valproic acid and lithium." Molecular Psychiatry, 2026.

When this report was written dismech did not cite the paper at all. It is now cited 13 times, all in Fetal_Valproate_Syndrome.yaml, and is cached in references_cache/.

This review covers the main text only. The 21 supplementary tables were not read.

What the paper offers as a model

The model is a primary culture of human neural progenitor cells from fetal dorsal telencephalon (14 to 21 gestational weeks). The cells come from 83 genotyped donors described as neurotypical. Each line was exposed for 48 hours to 1 mM valproic acid (VPA), 1.5 mM lithium chloride, or vehicle, then profiled by ATAC-seq, RNA-seq and an EdU proliferation assay.

In dismech terms this is an ExperimentalModel of type PRIMARY_CELL_CULTURE. It models an exposure and a treatment response. It does not model a disease genotype.

Main results:

  • VPA changed chromatin accessibility at 65% of peaks, more often closing than opening. Lithium changed 18%, more often opening.
  • VPA shifted progenitors from proliferation toward differentiation and reduced the S-phase fraction in every donor at 1 mM. Lithium increased proliferation at 0.75 to 1.5 mM and decreased it at higher concentrations.
  • VPA-responsive regions were enriched for the heritability of autism, intelligence, educational attainment and bipolar disorder. Lithium-responsive regions were enriched for bipolar disorder heritability.
  • Genetic variants changed the molecular response to the drugs: 779 response caQTLs and 214 response eQTLs for VPA, and 15 of each for lithium.
  • Under VPA only, a transcriptome-wide association study linked folate and cobalamin metabolism genes (DHFR2, MTFMT, MTHFD1, MMACHC, MTRR, MMUT) to educational attainment.

Gaps in disorder entries the paper could fill

1. Fetal valproate syndrome has no entry and no stub

Closed. Fetal_Valproate_Syndrome.yaml was added in #13604: 4,165 lines, 17 pathophysiology nodes, five experimental models and four animal models. The paper is curated there as the model this section anticipated, under experimental_models:

- name: Genotyped human neural progenitor population exposed to valproic acid
  experimental_model_type: PRIMARY_CELL_CULTURE
  publication: PMID:41935183

linked to mechanism nodes with MEASURES and RECAPITULATES relationships at MODERATE fidelity. What follows is the original analysis, kept because the schema limits it predicted are what the merged entry ran into.

MONDO:0012275 appeared nowhere in kb/ or stubs/. This is the disorder the model most directly represents.

The paper can be used for an entry, with limits. An ExperimentalModel could hold:

  • the model itself: type, organism, cell type, tissue and publication;
  • links to mechanism nodes through modeled_mechanisms, each with a relationship, fidelity, divergences and evidence;
  • structured readouts on each link, such as chromatin accessibility changed and S-phase fraction decreased, with a direction and a quoted snippet.

Three things would fall back to prose (see "Gaps in the schema" below): the exposure dose and duration, the donor population, and the gene-by-exposure results. The last is the paper's main novelty.

The paper cannot found the entry on its own. It would supply the experimental model and evidence for one or two mechanism nodes, such as chromatin change and the shift from proliferation to differentiation. Phenotypes, prevalence, the clinical exposure-outcome evidence and the mechanism nodes themselves have to come from the clinical and teratology literature.

Open, but reframed. The chain this section asked for now exists in Fetal_Valproate_Syndrome.yaml, and no entry references that file: a git grep for its name across kb/ matches only inside it. So the work is to point the autism entry's existing valproate factor at that chain with a cross-entry reference, not to add a model to the autism entry. Issue #13779.

The "Prenatal valproate exposure" factor is linked to the node "Heterogeneous neurodevelopmental molecular effects" as INDIRECT_UNKNOWN_INTERMEDIATES. Its description says "No cited sentence follows valproate to any molecular step".

The paper supplies human in-vitro intermediates. It also shows that VPA-responsive regions are enriched for autism heritability and VPA-responsive genes for autism-associated genes.

The entry has no valproate model of any kind. Its two animal models are a Cntnap2 knockout with maternal immune activation and a prenatal interferon-alpha rat.

3. Bipolar Disorder has no model sections

Open. Issue #13600 proposes the lithium model described here.

The entry has neither experimental_models nor animal_models. The patient iPSC neuron work is cited only inside pathophysiology nodes.

The paper would add a lithium model with three findings:

  • lithium increases progenitor proliferation at 0.75 to 1.5 mM;
  • lithium-responsive regions are enriched for bipolar disorder heritability;
  • 105 genes associate with bipolar disorder under lithium, including FADS1, TRANK1 and BDNF.

There is no neurogenesis or progenitor-proliferation node to link to. "Neuroplasticity Alterations" is the closest. The companion paper on lithium-induced proliferation and GNL3 (PMID:36307327) is also absent.

4. Neural tube defects: the folate question is open or contradicted

Open, and now a contradiction between entries rather than a gap. Issue

13778.

  • Anencephaly records the valproate link as indirect because the responsible step is "unsettled between folate antagonism and histone deacetylase inhibition".
  • Spina Bifida Cystica says VPA acts independently of maternal folate status. It has no model sections.

Fetal_Valproate_Syndrome.yaml curates folate receptor antagonism as a mechanism, quoting that valproate-exposed cells take up less folate, while recording separately that the evidence for folic acid preventing valproate teratogenesis is conflicting. Those are two different claims, and the spina bifida wording conflates them: "independent of maternal folate status" is defensible about rescue and is contradicted about mechanism. The three entries cannot all be right.

The paper's own folate result bears on this only indirectly, since its outcome is cognition (educational attainment) rather than neural tube closure.

5. Schizophrenia and Epilepsy have no model sections

Dropped on 2026-10-09 as not worth the work: one colocalization and a "modest" enrichment, against the cost of opening model sections in two large entries. No issue filed.

The contribution here is thinner.

  • Schizophrenia: a VPA-responsive eQTL for AS3MT (rs7096169) colocalizes with a schizophrenia GWAS signal. AS3MT is not among the entry's genes.
  • Epilepsy: the paper describes the heritability contribution of VPA-responsive regions as modest.

Gaps in the schema

Confirmed by the merged entry, and not acted on. #13604 hit all three rows below. It recorded the donor panel as cell_source: Primary human neural progenitor cells from 83 genotyped multi-ancestry donors and the exposure as culture_system: Monolayer, 48-hour exposure to 1 mM valproic acid, so both are prose in slots not meant for them. The genetic half of the paper was dropped: the entry has no mention of a response QTL, of MTHFD1, or of educational attainment, which is the paper's own headline result.

ExperimentalModel can record this model, but three parts of the design have no structured home.

Part of the design What the schema offers Consequence
Exposure, dose and duration (1 mM VPA, 48 h) The free-text conditions slot. ExperimentalPerturbation exists only on Experiment, which holds proposed experiments "All models exposed to valproate" cannot be queried
Donor population (83 genotyped, unaffected donors) The free-text cell_source slot A population-scale donor panel cannot be told apart from a single line
Gene-by-exposure interaction (response QTLs) Nothing. EnvironmentalMechanismTarget has no genetic-context field, and TreatmentEffectModifier covers patient subgroups The result can only be mentioned in findings or notes

No disorder file mentions a response QTL, and one mentions a caQTL.

The first two rows are a loss of queryability, not of content. The third is a loss of content.

The authors' stated limitations fit the existing ModelDivergence kinds: a single cell type is a BOUNDARY_OMISSION, and acute 48-hour exposure is a TEMPORAL_SCOPE divergence.

Caveats for curation

  • The findings are associative: heritability enrichment, colocalization and transcriptome-wide association. They belong as IN_VITRO evidence on model links, not as support for human phenotypes.
  • The donors are unaffected, so the model says how typical progenitors respond to the drugs. It does not say how patient cells respond.
  • Snippets can be checked against the PMC full text, which is open access and now cached in references_cache/.

What is left

Gaps 2, 3 and 4 are filed as issues and need a curator.

One of the three schema gaps is filed, which this report originally said none were. #13768 covers the exposure and perturbation row: it asks for a structured perturbation on the model classes, and notes that ExperimentalModel.conditions is a list of strings.

No issue proposes a fix for the other two rows. The donor-population row is recorded in #13602's list of what the schema cannot hold today — cell_source is free text, so a genotyped donor panel cannot be told apart from a single cell line — but nothing asks for the change. For the gene-by-exposure row, searching the tracker for QTL, gene-by-exposure, cell_source and donor panel turned up nothing beyond those two issues. That row is the paper's own headline result — common variants change how neural progenitors respond to these drugs — and it was dropped when the paper was curated.

Whether that last result belongs in dismech is an open question rather than a gap to close. A response QTL describes variance across a donor panel, not a step in any one patient's disease, and dismech is organized around a single disease with a reasonably conserved pathograph. #13602, which sorts models by what ties them to the disease, is where that question is being worked out; its bottom row already places a looser case, model-responsive regions overlapping GWAS signal, outside what a model link should carry. A response QTL is not that case — it is measured in the dish, on the model's own donors — but it raises the same question about scope.

One candidate home is EnviroMech, a separate knowledge base in the same family, which records the exposure-outcome association itself rather than hanging it on a curated disease. A Mech scoped to pharmacogenomic and gene-by-environment response is another. Neither is settled.