Check Claims Before You Publish

 Long Explanation Copied

Author Review: Maria Carilli (ORCID 0000-0002-8977-7224)

Maria Carilli is an emerging computational biologist trained in the Pachter lab (Caltech) with prior telomerase biochemistry work (University of Colorado Boulder / HHMI). OpenAlex records 22 works with 252 citations and an h-index of 8, with a clear productivity surge: 6 works in 2024 (149 citations that year) and 5 works in 2026 OpenAlex citation metrics for Maria Carilli. A Semantic Scholar profile showing h-index 1 and 2 papers appears to be an incomplete, name-collision-prone merge and should not be taken at face value as her true impact.

First-author methodological depth. Carilli is first author of biVI (Nature Methods 2024), which embeds biophysical transcription/splicing models inside the scVI variational autoencoder for bimodal single-cell RNA-seq data biVI in Nature Methods. Her 2025 preprint ρPCA derives an exact, parameter-free Rayleigh-quotient contrastive PCA, solving the generalized eigenproblem Σ_Tv = λΣ_Bv; it outperforms cPCA in accuracy and speed (10,000 samples × 1,000 features: cPCA ≈ 2 min vs ρPCA in seconds) and is validated on 89,940 mouse kidney snRNA-seq nuclei (47,798 female, 42,142 male; 17 cell types), with open code (rhopca), reproduction repos, and a Zenodo deposit — all scoring 9/9 on BGPT quality/novelty/reproducibility extraction ρPCA contrastive PCA preprint. A 2026 follow-up extends this to spatial (k-ρPCA, colorectal cancer Visium HD, 471,258 bins) and functional (f-ρPCA, 23-patient longitudinal vaccine-response RNA-seq, 439 samples) settings ρPCA II spatial/functional extension.

Collaborative breadth. She contributes to viral sequence detection at single-cell resolution (Nature Biotechnology 2025) Single-cell viral detection in Nature Biotechnology and is a middle author on a 2026 foundation-models perturbation-prediction benchmark across Norman, Essential, Sciplex3, and Tahoe-100M datasets. That preprint is strong on scale (e.g., 964,451 single cells; 15,932 Tahoe triplets) and honest about null results for some modalities, but all authors are current/former GenBio AI employees with financial interest, and GenBio AI funded the work — a material conflict-of-interest caveat GenBio AI perturbation-prediction preprint.

Earlier wet-lab pedigree. Her NAR 2021 telomerase paper (38 citations) with Cech and Wuttke labs shows mechanistic biochemistry competence, ruling out a purely computational profile CST-telomerase mechanism paper.

Critical caveats. (1) h-index 8 with 252 citations is modest but consistent with an early-career author whose signature works are recent; her 2026 output is mostly preprints not yet peer-reviewed. (2) ρPCA papers rely on covariance estimates with Gaussian-ish assumptions and limited non-mouse-kidney validation; the authors' own falsification criteria (identical covariances, cPCA-with-tuned-α superiority) remain untested externally. (3) The Foundation Models paper carries industry funding and authorship conflicts; independent replication is pending. Confidence: moderate-high — trajectory is strongly upward, but several headline results await independent validation.

 BGPT Author Review

Scientific Quality

Strong early-career trajectory: first-author Nature Methods tool (biVI), a theoretically rigorous exact Rayleigh-quotient contrastive PCA framework with two extensions, contributions to a Nature Biotechnology viral-detection method, and a top-cited Nature consortium paper (93 citations). H-index 8/252 citations is modest but appropriate for career stage. Red flags: heavy reliance on preprints in 2026, one industry-funded benchmark with GenBio AI employment conflicts, and limited independent external validation of ρPCA claims.

Communication Quality

Papers include worked biological examples (sex-specific kidney variation, tumor boundaries, interferon trajectories), open code, reproduction repos, and Zenodo data, indicating clear, reproducible communication. BiVI and ρPCA papers connect linear algebra to biology accessibly. Deduction: heavy mathematical framing (generalized eigenproblems, functional bases) may be opaque to wet-lab readers, and some claims (interpretable loadings) lack downstream biological validation.

Author Novelty

ρPCA reframes contrastive PCA as an exact Rayleigh-quotient problem, eliminating cPCA's α-tuning — a genuinely new, parameter-free formulation with novel spatial (k-ρPCA) and functional (f-ρPCA) extensions. biVI uniquely fuses biophysical kinetics with VAEs. The foundation-model benchmark is more derivative (a large comparative evaluation) though it usefully resolves contradictory prior claims.

Scientific Rigor

Method papers derive exact solutions, state limitations explicitly (positive definiteness, regularization, covariance-estimation sensitivity), predefine falsification criteria, and release full code/data — hallmarks of rigor. The GenBio AI paper is internally honest about negative results for some modalities, but sponsorship and authorship financial interests, plus unreviewed preprint status, temper confidence. Sample sizes are large where reported.