Single-Cell Foundation Models in Biotechnology: A Scoping Review and Bibliometric Analysis of Multimodal AI for Cell-State, Perturbation, and Biomarker Prediction

Technical Summary: Single-Cell Foundation Models in Biotechnology

Problem Statement
The field of single-cell biology has rapidly adopted "foundation models"—large neural networks pretrained on broad datasets and fine-tuned for specific tasks—mirroring trends in natural language processing (NLP) and protein modeling. While models like Geneformer, scGPT, and scBERT promise unified embeddings for cell-state analysis, perturbation prediction, and biomarker discovery, the systematic advantages of these architectures over simpler alternatives remain incompletely evaluated. Critical concerns persist regarding benchmark rigor, data leakage, reproducibility, and whether reported performance gains are genuine or artifacts of poor experimental design. There is a lack of empirical, code-driven analysis characterizing the landscape of these models, the quality of their benchmarks, and the state of reproducibility across the literature.

Methodology
This study employs a systematic scoping review combined with an original computational bibliometric analysis. The authors constructed a fully automated pipeline to analyze literature published between 2020 and April 2026.

• Data Retrieval: Four databases (PubMed, bioRxiv, Semantic Scholar, Crossref) were queried using eight structured search strings, yielding 1,295 raw records.
• Screening & Deduplication: A programmatic relevance filter retained 1,249 records, followed by deduplication (exact DOI and fuzzy title matching), resulting in a final corpus of 1,042 unique records.
• Automated Extraction: Scripts extracted 45 fields per record, including model family, task type, input modality, and reproducibility markers (code, weights, data availability).
• Bibliometric Analysis: The study utilized TF-IDF vectorization and K-means clustering (k=8) for topic modeling, keyword co-occurrence network analysis, and publication timeline tracking.
• Quality Scoring: A custom benchmark quality score (0–4) was assigned programmatically based on the presence of: (1) named benchmark datasets, (2) stated baseline comparisons, (3) explicit evaluation metrics, and (4) independent validation.
• Reproducibility Audit: Availability of code, model weights, and training data was flagged via keyword matching in abstracts to compute a composite reproducibility index.

Key Contributions

1. Systematic Landscape Mapping: The paper provides the first large-scale, code-driven characterization of the single-cell foundation model field, categorizing 1,042 records by model architecture, task, and modality.
2. Benchmark Rigor Audit: It introduces a quantitative scoring system to assess the methodological quality of benchmarking in this domain, revealing a systemic lack of baseline comparisons and independent validation.
3. Reproducibility Assessment: The study quantifies the "reproducibility gap," documenting the low rates of shared code, weights, and data, and correlating these with benchmark quality.
4. Critical Evaluation of Claims: The analysis empirically tests the narrative of "foundation model superiority" against the reality of reported baselines and task performance.

Key Results

• Publication Growth: The field has experienced approximately 24-fold growth in annual publication volume between 2020 and 2025, with a sharp inflection point post-2022 coinciding with the rise of large language models.
• Model Landscape: Multimodal fusion architectures are the dominant category (29.4%), followed by standard attention/transformer models (22.0%) and VAE-based generative models (17.6%). True "foundation-scale" pretraining models (claiming massive pretraining objectives) constitute only 13.9% of the literature.
• Task Distribution: Cell-type annotation (29.9%), spatial reconstruction (28.6%), and perturbation prediction (27.9%) are the most common downstream tasks.
• Benchmark Quality: The median benchmark quality score is 1/4. Only 12.4% of papers achieve a score of 3 or 4. Crucially, explicit baseline comparisons appear in only 25.7% of abstracts, and independent validation is present in just 9.3%.
• Reproducibility: Code availability is detectable in 40.4% of records, while model weights (15.0%) and training data (19.5%) are rarely shared. The composite reproducibility index averages 0.25.
• Baseline Performance: The analysis supports the finding that simple baselines (e.g., logistic regression on PCA embeddings, scVI) are frequently competitive with large transformer models, yet are rarely included in published comparisons.

Significance and Claims
The paper argues that while the field is expanding rapidly and producing sophisticated tools, the cumulative scientific value is currently limited by inconsistent reproducibility and benchmark rigor.

• Modest Claims on Superiority: The authors assert that evidence for the superiority of foundation models over simple baselines is often weak or context-dependent. For tasks like cell-type annotation, pre-trained representations show modest but consistent advantages in few-shot scenarios, but claims of massive gains in perturbation prediction or drug response are often unsupported by rigorous out-of-distribution testing.
• The "Scale" Question: The paper notes that many models labeled as "foundation models" are trained on data scales (tens to hundreds of thousands of cells) orders of magnitude smaller than NLP counterparts, and the biological necessity of such scaling has not been rigorously demonstrated.
• Call for Standardization: The primary significance of this work is the identification of specific gaps: the need for standardized benchmark suites (analogous to GLUE in NLP), mandatory inclusion of simple baselines in evaluations, and infrastructure for sharing model weights. The authors conclude that broader adoption of open code, data, and standardized protocols is required to transition the field from a period of rapid proliferation to one of cumulative, reproducible scientific advancement.