Benchmark whether receptor signaling forecasts held-out mouse behavioral responses

Build a reproducible benchmark comparing simple receptor-signaling models against a mean-only baseline for published mouse head-twitch response magnitude. Test both held-out compounds and transfer between the study's two compound panels. The useful outcome is a bounded decision about what evidence a computational neuropharmacology forecast should require, including a finding that no tested model improves reliably on the baseline.

The OpenLabs Nootropics project asks what should count as a credible forecast beyond target binding. This task tests one measurable intermediate step: whether measured Gq or beta-arrestin2 efficacy helps forecast mouse head-twitch magnitude outside a model's training compounds. It uses the actual published source data from Wallach et al. (2023); it does not train a human-effect predictor, propose new compounds, or establish therapeutic value.

Use the Fig 5 worksheet of the Source Data workbook. Panels A/B contain the same 14 numbered 25N compounds; panels C/D contain the same 24 named comparison compounds. Join A to B and C to D by the explicit compound identifier, retaining panel membership. Gq efficacy comes from A/C, beta-arrestin2 efficacy from B/D, and the repeated HTR magnitude must agree within each joined record. These are 38 compound summaries, not 76 independent observations. The response is maximum observed head-twitch counts per minute, not ED50, a binary psychedelic label, or a human outcome. Study the paper's assay and response definitions before interpreting a benchmark.

Compare exactly five prespecified models using ordinary least squares with an intercept: a training-response mean; Gq alone; arrestin alone; both linear predictors; and Gq plus Gq squared. No tuning or model selection uses held-out outcomes. Standardize predictors using training means and population standard deviations before constructing the quadratic term; record conventions and handle any zero variance explicitly. Use a numerically stable least-squares solver. Do not clip negative predictions: flag and discuss them as model limitations.

Run leave-one-compound-out predictions within the 14-compound panel, within the 24-compound panel, and across the pooled 38 compounds. Separately train on all 14 and predict all 24, then train on all 24 and predict all 14. For each setting report MAE, RMSE and mean signed error, plus each model's change in MAE and RMSE relative to that setting's training-mean baseline. Report training-fit errors separately to show optimism. Do not describe the pooled split as a scaffold-held-out or prospective validation.

For each nonbaseline model, quantify influence by removing each compound in turn and recalculating the complete within-panel and pooled leave-one-out procedures on the reduced dataset. Report the range of baseline-relative MAE changes and any reversal of the model-versus-baseline conclusion. This is a finite influence sensitivity analysis, not a confidence interval. For cross-panel transfer, flag test predictor values outside each training predictor's observed range, summarize how many test records are outside at least one range, and show their errors separately. No statistical significance claim is required.

• The exact Source Data workbook, provenance.json and compound-data.csv: all 38 records with source sheet, panel, cell references, original identifier, Gq efficacy, arrestin efficacy, HTR magnitude and units. Include a reconciliation check that paired response values and compound identities agree. No unreported deduplication or hand-edited numeric corrections.
• analysis.py or analysis.R, predictions.csv, metrics.csv, influence.csv and transfer-range.csv: reproduce all five models and five evaluation settings, training-only preprocessing, every held-out prediction, baseline-relative comparisons, deletion sensitivity and transfer-range flags. The deletion analysis is required for the three leave-one-out settings; transfer deletion fits are not required.
• forecast-errors.svg or forecast-errors.png: actual-versus-held-out-predicted views and a clear model/baseline comparison that retains the two panel labels and distinguishes within-panel, pooled and transfer results.
• decision.md: recommend which, if any, tested signaling model merits a next independent validation step and what evidence would change that decision. Give a concrete prospective validation design: intended compound domain, locked predictor/outcome definitions, selection independent of observed HTR, comparator, prespecified error metric and a defensible precision target to be chosen before collecting data. Do not invent a validated error tolerance or successful future experiment. Explain how affinity, exposure, off-targets, assay context and dose selection remain untested here.
• README.md: environment, execution command, provenance and all exclusions or unresolved issues. All required bytes must be included; external URLs do not substitute for the data or computed results.

1. Wallach et al. (2023), *Identification of 5-HT2A receptor signaling pathways associated with psychedelic potential*, DOI 10.1038/s41467-023-44016-1, PMC10724237: https://pmc.ncbi.nlm.nih.gov/articles/PMC10724237/ and https://www.nature.com/articles/s41467-023-44016-1 . Figure 5 and its methods define the two panels, efficacy assays and maximum HTR outcome.
2. Required Source Data file 41467_2023_44016_MOESM12_ESM.xlsx, 297,248 bytes, SHA-256 c32b4d301d6acac4ccb18ca9102f166104395cb501e037cc4643d0d5b46cbabd, verified 17 September 2026. Linked by the article at https://pmc.ncbi.nlm.nih.gov/articles/instance/10724237/bin/41467_2023_44016_MOESM12_ESM.xlsx . An unauthenticated public retrieval route verified for this campaign is https://www.ebi.ac.uk/europepmc/webservices/rest/PMC10724237/supplementaryFiles ; extract the named workbook from that archive. Other supplementary files are context only; submit the named workbook, not the entire supplementary archive. The required worksheet ranges are B3:D16 (A), B18:D31 (B), B33:D56 (C), B58:D81 (D). Preserve published identifiers such as CPM as compound labels; do not confuse them with the response-unit abbreviation.

Project context: https://openlabs-git-codex-openlabs-elgora-adapter-bio-xyz.vercel.app/projects/98397606-1b31-48f8-818a-ff21621cf5e7 . Independent contribution without project-owner endorsement. Estimated work is four to six hours on an ordinary laptop; no paid API, raw sequencing, GPU or new animal experiment is needed.

Only submissions satisfying all criteria are eligible. Prefer fewer material numerical or scientific errors, then the most defensible next-validation decision supported by held-out and transfer evidence, then stronger reproducibility and provenance. Higher claimed accuracy, a positive mechanism story or endorsement of a particular model earns no preference. Remaining ties go to earlier on-chain submission timestamp, then lower numeric submission ID. No winner is required if none qualifies.