NeuroCell-LM AD
A donor-aware benchmark of frozen Cell2Sentence representations, with public code and aggregate evidence for assay transfer, brain-region transfer, and cross-sectional Alzheimer’s disease neuropathology-severity prediction.
Research use only. This repository is not a diagnostic model, clinical decision-support system, medical device, or validated patient-level predictor.
What this repository is
This is a code-and-evidence release for a frozen development benchmark. It does not contain a standalone language model, C2S weights, raw SEA-AD matrices, donor-level predictions, embedding vectors, or donor-lockbox data. It packages analysis and evaluation code, sanitized aggregate results, frozen protocol summaries, figures, and cryptographic manifests. The evaluated foundation-model backbone is vandijklab/C2S-Scale-Gemma-2-2B; this is not a comparison of multiple foundation-model backbones. Public result inspection is possible from this release, but full scientific regeneration requires upstream data and donor-level intermediates that are not included here.
The central question is whether frozen single-cell foundation-model representations provide robust or incrementally useful biological signal when evaluation respects donors as the independent unit.
Why donor-level evaluation matters
Single-cell datasets contain many cells from each donor. Treating those cells as independent replicates creates pseudoreplication and can greatly overstate certainty. NeuroCell-LM AD keeps all records from one donor together during splitting and cross-validation. The frozen design contains 56 training donors, 14 one-time development-validation donors, and 14 lockbox donors reserved for evaluation. Donor metadata and pathology summaries, including CPS_Global, were used before partition freeze to balance candidate splits. After freeze, partition assignments were retained to exclude lockbox donors from development and evaluation. No lockbox predictions or performance metrics were generated.
Data, assays, regions, and cell populations
The benchmark uses official SEA-AD/Allen Institute resources acquired from upstream, not redistributed here. The broad cell-identity task covers five inhibitory subclasses—Sst, Pvalb, Vip, Lamp5, and Sncg—using MTG 10xV3.1 as the source setting, paired Multiome GEX as an assay-transfer setting, and MEC 10xV3.1 as a held-out-region setting.
The donor-level endpoint is CPS_Global, interpreted only as a cross-sectional AD-severity target. It is not longitudinal progression.
Representations and baselines
- frozen
vandijklab/C2S-Scale-Gemma-2-2Bembeddings; - sparse top-500 ranked-gene linear features;
- 2,000-HVG raw-expression linear features;
- donor-level raw-count pseudobulk PCA/ridge and ElasticNet;
- donor covariates and full-census cell composition;
- pooled and subclass-resolved frozen C2S donor aggregates.
No C2S or Gemma weights are redistributed. The release records the pinned upstream model revision and provides scripts that recreate embeddings after users accept the upstream model and data terms.
Broad cell-subclass benchmark
| Benchmark | Frozen C2S | Ranked-gene linear | Raw-HVG linear |
|---|---|---|---|
| MTG in-region | 0.9921 | 0.9911 | 0.9925 |
| MTG → Multiome | 0.9911 | 0.9912 | 0.9923 |
| MTG → MEC | 0.9870 | 0.9843 | 0.9872 |
The broad five-subclass problem remained near saturation across donor, assay, and region shifts. Frozen C2S and the simpler expression-based models had similarly high point estimates. This is a descriptive comparison, not a formal equivalence or superiority result, and it does not establish clinical validity. The displayed F1 values are rounded to four decimals; full-precision values are retained in reports/public/.
Donor-level AD-severity evaluation
For the donor-level severity models, preprocessing and hyperparameter selection were fitted within the relevant training folds. Seven frozen model families were then applied once to the 14 development-validation donors. This should not be generalized to every cell-classification preprocessing step: the ranked-gene vocabulary and raw-expression HVG list were selected from the full 56-donor training partition before its internal grouped cross-validation. See the implementation clarification.
| Frozen model | MAE | RMSE | Spearman ρ | R² |
|---|---|---|---|---|
| Frozen C2S pooled PCA/ridge | 0.1689 | 0.2338 | 0.477 | 0.209 |
| Covariates-only ridge | 0.1785 | 0.2473 | 0.235 | 0.115 |
| Raw pseudobulk PCA/ridge | 0.1896 | 0.2555 | 0.376 | 0.055 |
| Frozen C2S subclass-resolved PCA/ridge | 0.1934 | 0.2579 | 0.512 | 0.038 |
| Raw pseudobulk ElasticNet | 0.2026 | 0.2604 | 0.363 | 0.019 |
| Cell composition ElasticNet | 0.2082 | 0.2657 | — | -0.021 |
| Training-mean null | 0.2082 | 0.2657 | — | -0.021 |
The descriptive lowest-MAE model was Frozen C2S pooled PCA/ridge (MAE 0.1689). Its point MAE improvement over covariates-only was 0.0096, with paired 95% interval [-0.0622, 0.0834]. This contrast did not demonstrate clear positive incremental value. The prespecified advancement rule considered three primary transcriptomic models versus covariates-only: raw pseudobulk PCA/ridge, raw pseudobulk ElasticNet, and pooled frozen C2S PCA/ridge. None had a paired interval entirely above zero, so the advancement gate was not met.
Although pooled frozen C2S had the lowest validation MAE, its mean prediction error (prediction minus observed) was 0.0628, compared with 0.0046 for the covariates-only model. Positive values indicate average overprediction. Validation calibration was prohibited and was not performed.
Frozen conclusion
Supported
- strong broad inhibitory-subclass transfer for frozen C2S and simpler linear representations;
- donor-aware, paired benchmarking across assay and brain-region shift;
- the lowest observed validation MAE for pooled frozen C2S among the seven model families, without established superiority over donor covariates.
Not supported
- clinical or diagnostic use;
- longitudinal progression claims;
- clear incremental AD-severity value beyond conventional donor covariates;
- confirmatory lockbox performance;
- a rationale for post-hoc QLoRA, validation calibration, or ensemble construction.
The lockbox advancement gate was not met. The lockbox remained unevaluated after partition freeze: no predictions or performance metrics were generated. This does not mean that donor metadata or pathology summaries were never used during the earlier split-balancing stage.
Repository contents
scripts/research_pipeline/: archived source stages without individual donor identifiers;scripts/release/: preparation, publication, remote verification, figure, and smoke-test tooling;reports/public/: aggregate-only public results generated from frozen reports;protocols/: sanitized frozen-method summaries;figures/: figures generated from aggregate outputs;manifests/: file hashes, public-release inventory, schemas, and audit results.
Reproduction
Public integrity checks and aggregate-figure regeneration are not a complete independent rerun of the scientific study. Donor-level outcomes and predictions are needed to recompute the paired resampling results; they are not in this public release. See REPRODUCIBILITY.md for the staged workflow and DATA_AVAILABILITY.md for official source acquisition. Reproducing the full analyses requires upstream SEA-AD data, substantial local disk space, and—in the C2S stages—a CUDA-capable GPU. Reviewing the public results and running the release smoke test requires only a normal Python environment.
Documentation status and publication status
This documentation clarification describes the implementation in the original v1.0.0 scientific release at commit ef6b31994a96173eb160c78e02ec490620be1a4d. The research-pipeline scripts, frozen scientific configuration, and public numerical reports have not been changed. The original tag remains the scientific-release reference; this documentation update must not move or replace it. Details of the wording corrections are recorded in DOCUMENTATION_CLARIFICATIONS.md.
A manuscript describing the study was submitted to PLOS ONE on August 4, 2026. Submission is not journal acceptance, publication, or independent confirmation of the findings. This repository is a benchmark code-and-evidence release, not a newly trained language model.
The study was conducted independently using public data and author-owned computing resources. Bristol Myers Squibb did not sponsor this study or provide its data or computing resources; the author's paid contract relationship is disclosed in the manuscript.
AI assistance
OpenAI ChatGPT assisted with code troubleshooting, discussion and explanation of results, literature and journal-policy research, manuscript and supporting-document preparation, figure formatting, and reproducibility/documentation audits. The author is responsible for checking and approving AI-assisted contributions. The submitted manuscript contains a more detailed disclosure. Historical model-version labels are author-reported rather than independently verified by the repository, and this release does not use repository timestamps as evidence of AI-service access dates.
A private project snapshot provided during audit included coded donor-to-partition assignments and aggregate lockbox summaries. This is distinct from model evaluation: no lockbox predictions or performance metrics were generated or assessed. Those private identifiers and summaries are not redistributed here. No new biological data or numerical scientific results are created by this documentation patch, and no Cell2Sentence weights were trained or updated.
Licensing and attribution
This is a mixed-license research release. Original release tooling and original project code are offered under Apache-2.0; original prose and figures are offered under CC BY 4.0. SEA-AD/Allen-derived aggregate evidence remains subject to the Allen Institute Terms of Use and Citation Policy, including noncommercial-use and attribution requirements. C2S and Gemma are governed by their upstream licenses. See LICENSE and THIRD_PARTY_NOTICES.md.
Limitations and nonclinical warning
See LIMITATIONS.md. The donor sample is small, the lockbox was not evaluated, technical and study confounding remain possible, and the benchmark is not a clinical validation.