BioML Seminar 3.1 - Abhinav Adduri on Modeling cell perturbations with STATE

BioML Seminar 3.1 - Abhinav Adduri on Modeling cell perturbations with STATE

by Abhinav Adduri

Formal & Physical Sciences Biology PSALife sciences: general issuesPSAXComputational biology
🎙 Abhinav Adduri 👥 14K 📅 October 30, 2025 ⏱ 66 min 👁 207 📄 original study 🧭 2026-08-16
Available in: English (current) Français

Keywords

STATEperturbation responsesingle-cell RNA-seqtransformercell state

Summary

Abhinav Adduri presents STATE, a transformer-based model for predicting cellular responses to perturbations. The talk begins by motivating the need for in silico cell models, citing the complexity of eukaryotic cells and the limitations of traditional systems biology. STATE addresses the challenge of predicting perturbation effects across diverse cellular contexts by operating on sets of cells, leveraging self-attention to generalize over averaging and single-cell mapping approaches. The model consists of two components: a transition model that learns perturbation effects on covariate-matched cell sets, and an embedding model that provides a shared representation across datasets. Training on over 100 million perturbed cells, STATE improves discrimination of effects by more than 30% and identifies differentially expressed genes with higher accuracy. The talk highlights the importance of scaling data and improving model architectures, drawing parallels to AlphaFold’s success. STATE also enables zero-shot inference on novel cell types and is open-sourced, with a virtual cell challenge offering a $100,000 prize. The presentation includes technical details on tokenization, attention patterns, and comparisons with optimal transport approaches.

171 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides valuable insights into a cutting-edge application of machine learning to biology. The argumentation is solid, grounded in the speaker’s research and supported by quantitative results. The speaker clearly explains the challenges of single-cell data (e.g., noise, batch effects, lack of instance-level correspondence) and justifies the design choices of STATE, such as using sets of cells and covariate matching. The comparison with existing methods (pseudobulk, scVI, optimal transport) is informative, and the attention map analysis offers intuitive understanding. The speaker also acknowledges limitations and areas for improvement, such as the simple one-hot encoding of perturbations, which adds credibility.

Scientific Rigor, Source Quality, Title Accuracy

The talk demonstrates scientific rigor through a clear methodology, quantitative evaluations, and references to open-source resources. The speaker cites relevant prior work (e.g., AlphaFold, scVI) and mentions the availability of the model and competition. The title accurately reflects the content. The presentation is a seminar talk, not a peer-reviewed publication, but the speaker’s affiliation with Arc Institute and the inclusion of detailed results enhance credibility. The talk also mentions a sponsor (Amplify) but this does not affect the scientific content.

195 words

Title / Content Match

The title accurately reflects the content: a seminar on modeling cell perturbations with the STATE model.

Quality & Reliability

8/10

The talk presents a novel machine learning model (STATE) with clear methodology, quantitative results, and references to open-source resources. The speaker is a research scientist at Arc Institute, and the work is part of a larger initiative (virtual cell challenge). However, the presentation is a seminar talk, not a peer-reviewed publication, and some details are simplified.

Key Moments

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

Cited Sources

  • STATE GitHub repository — Mentioned as open-source repository for the model
  • Virtual Cell Challenge — Mentioned as a Kaggle-style competition with $100,000 prize
  • Tahoe Therapeutics — Mentioned as a company that released 100 million cells of single-cell data

Concurring Sources

  • AlphaFold — Referenced as an example of deep learning success in biology
  • scVI — Mentioned as a baseline method for single-cell analysis

Contribution & Novelties

STATE introduces a novel approach to perturbation prediction by using a set transformer that operates on groups of cells, enabling the model to account for cellular heterogeneity and generalize to unseen contexts. The model’s ability to learn from over 100 million cells and improve discrimination by 30% is a significant advancement. The open-source release and virtual cell challenge further contribute to the field.

Pour aller plus loin :

  • AlphaFold — Landmark deep learning model for protein structure prediction, referenced as inspiration.
  • scVI — A variational autoencoder for single-cell data, mentioned as a baseline.
  • Optimal transport — Mathematical framework used for comparing distributions, compared to STATE.

105 words

Radar Profile

The radar profile shows high scores in information quantity, quality, and reliability, with a slightly lower technical level, indicating a talk that is informative and credible but accessible to a broad audience.

Reliability 8/10

💬 No comments were provided for analysis.