In this week’s AI ∩ Bio paper, we look at a multi-dataset benchmark where deep learning models trained to predict single-cell perturbation outcomes did not outperform two deliberately simple baselines: “assume nothing changes” and “predict the additive combination of single perturbations.” Performance improved when models were reduced to just their internal embeddings and a linear decoder—no deep layers, just a direct projection from embedding space to gene expression. In other words, the models had learned biologically meaningful representations during training, but using those representations to make predictions through a complex decoder often made things worse.

This result may not surprise those familiar with traditional machine learning approaches. As the bias–variance tradeoff reminds us: when signal is weak and noise is high, simpler models often generalize better and overfit less, leading them to outperform more complex alternatives.

This study revealed both conditions:

  • The signal from genetic perturbations is often modest or additive.
  • The noise in single-cell gene expression is high, due to dropout, amplification bias, and biological variability.

If your model must detect subtle shifts in expression, but your measurement system is saturated with noise, then even an optimal model will struggle to generate meaningful predictions. And a large model may do worse, because it fits spurious correlations that simpler models ignore.

AI ∩ Bio

Illustration of scientific paper with the title "AI ∩ Bio: Reading the Revolution"

More in this series

  1. ChatNT: The future of biological assistants—or a mirage in a lab coat?
  2. GET: A Foundation Model for Transcription, Still Between Promise and Proof
  3. X-Atlas/Orion: Your Model is Only as Good as Your Training Data
  4. From Better Models to Better Questions: A Pathology AI Rethink
  5. The Model Isn’t the Magic: How Cytoland shows that domain expertise—not just deep learning—is what makes AI in biology work.
  6. Boltz-2: How much can 3D structure really tell us about molecular binding energetics?
  7. Investigating the volume and diversity of data needed for generalizable antibody–antigen ΔΔG prediction
  8. Beyond Binding: Rethinking Drug Design in the Age of AI and Structural Biology
  9. Testing the Physics Beneath the Predictions Beyond RMSD: What AlphaFold3 Really Understands
  10. Hype, Hurdles, and Hepatotoxicity: A Bold Step for AI-Designed Drugs, But Still Miles to Go
  11. When Complexity Misleads
  12. What Are Genomic Transformers Actually Learning?
  13. mRNABench and the Future of AI in Biology: Why Domain Knowledge Wins
  14. OncoGAN Creates Synthetic Cancer Genomes, Opening New Paths for Privacy-Preserving Precision Oncology
  15. Readable Rules, Testable Models: A New Grammar for Virtual Cells
  16. Multimodal CustOmics: Fusing Pathology Images and Tumor Genomics for Next-Gen Cancer Diagnostics
  17. Beyond Perturbation Simulations: PDGrapher Shows a Faster Way to Identify Actionable Targets
  18. Can AI design epigenetic anti-aging strategies?
  19. What happens when physicians use GPT-4 for diagnosis
  20. Can generative AI predict emergent phenomena?
  21. DeepSomatic and the question of how AI learns from itself
  22. Toward Mechanism-Centric Interpretability in Genomic Machine Learning
  23. From AlphaEvolve to DeepEvolve: What We’re Learning About Machine-Led Scientific Discovery
  24. Kosmos and the Culture of Discovery
  25. From Embeddings to Insight
  26. From Prediction to Explanation: How BIOREASON Reframes Genomic AI as a Reasoning Problem
  27. AI Models Need Better Truth—Platinum Pedigree Shows How
  28. From Evolutionary Intolerance to Clinical Insight: What popEVE Teaches Us About Missense Variants
  29. Interpretable Latent Spaces, Messy Biology: What AUTOENCODIX Teaches Us About Autoencoders in the Wild
  30. The Gene Ontology Knowledgebase in 2026

Before adding more architecture, it’s worth asking:

  • Have we clearly defined our outcome?
  • Do we understand the system we are working with?
  • Can our data realistically support the question we’re asking?

Otherwise, we risk building a computational Rube Goldberg machine: impressive, fragile, and ultimately unclear.