BioML Seminar 4.4 - Shreshth Gandhi on Scaling Perturbation-Trained Single-Cell Foundation Models

Value of the Information & Strength of the Argument

The talk provides substantial value by presenting a novel large-scale dataset and model, with detailed technical insights into architecture and training. The argumentation is solid, backed by benchmark comparisons against existing models. The speaker transparently discusses limitations, such as the performance gap in zero-shot perturbation prediction. The engineering details, like custom attention kernels and data streaming, add practical value. However, some claims lack external validation, and the talk is a seminar rather than a peer-reviewed presentation.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is high, with clear methodology and evaluation. The speaker references relevant prior work (e.g., scGPT, GeneFormer, scBase, Cell by Gene) and datasets (Cancer Dependency Map, MSigDB). The title accurately reflects the content. No public comments were provided for analysis.

The title accurately reflects the content: a seminar on scaling perturbation-trained single-cell foundation models, presented by the lead author.

8/10

The talk presents original research with technical depth, including model architecture, training details, and benchmark results. The speaker is a domain expert with relevant industry experience. However, the presentation is a seminar talk without peer-reviewed publication details, and some claims (e.g., cost, performance) are not independently verified.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 02:00Introduction to Tahoe-x1 and the Tahoe-100M dataset.
• 05:00Explanation of pooled experiments and genetic demultiplexing.
• 10:00Overview of the model architecture and training objective.
• 15:00Discussion on engineering optimizations and cost reduction.
• 20:00Application to gene dependency prediction from CRISPR screens.
• 25:00Application to gene set membership classification.
• 30:00Perturbation prediction results and performance gap.
• 35:00Lessons learned and future directions towards a virtual cell.

• scBase — Mentioned as a compilation of publicly available datasets from the Arc Institute.
• Cell by Gene — Mentioned as a corpus compiled by CZIS.
• Cancer Dependency Map — Used for CRISPR knockout screens and gene dependency scores.
• MSigDB — Used for gene set membership benchmarks.

• scGPT — The model architecture is derived from scGPT, and the talk reports improved performance.
• Geneformer — The model is inspired by Geneformer, and the talk compares performance.