TL;DR
In this talk, Abhinav Adduri presents a framework for predicting perturbation effects conditioned on cell state at the Foundation Models for Biology Seminar Series by GenBio AI. The approach separates a learned representation of cell state from a perturbation-effect component.
The idea is to first capture what state the cell is in, then model what the perturbation does given that state. This separation of state and effect makes the approach easier to interpret, more reliable across messy datasets, and better at spotting when results are driven by batch effects, rare cell types, or weak signals.
Introduction
Imagine you are running a CRISPR screen or testing a new compound. You expect to see changes in gene expression, but those changes will look very different depending on whether the target is a stem cell, an immune cell, or a tumor cell. Standard models blur these differences, treating the perturbation effect and the cell’s identity as one, which limits both generalization and interpretability.
Abhi and colleagues introduce state-conditioned modeling to address this. The idea is simple but powerful: first capture the baseline state of the cell, then model the effect of the perturbation given that state. By separating state and effect, the framework makes it easier to explain results, identify failures on rare cell types, and integrate data across labs and experiments.
Full Talk
Watch the full talk below:
The Problem
- Entanglement of state and effect: Standard models conflate baseline cell identity with the perturbation’s causal impact.
- Heterogeneity & batch effects: Datasets span labs, donors, and protocols; naive pooling invites spurious signals.
- Sparse labels & rare exemplars: Many perturbations and cell states have few observations, limiting learnability.
- Metrics vs utility: Validation loss can diverge from biological usefulness without task-aware evaluation.
- Baselines matter: Simple centroid/mean baselines are often surprisingly strong; new models must beat them reliably.
This work addresses these gaps by decoupling state and effect, handling covariates explicitly, and evaluating with task-aware metrics alongside strong baselines.
What is State-Conditioned Perturbation Modeling?
A modeling paradigm that decouples the representation of cell state (what the cell is) from the perturbation effect (what the intervention does), then recombines them to predict outcomes. The learned mapping is not strictly additive; it allows interactions, so effects can depend nonlinearly on the underlying state.
Architecture Overview
The architecture includes a state encoder that captures baseline cell identity with optional covariates, a perturbation encoder that maps target, compound, and dose into an effect embedding, and a transformer that models their interaction beyond simple additivity for state-dependent responses. The cell embedding is trained similarly to an autoencoder, using a bottleneck layer to learn meaningful cell embeddings that capture transcriptome-wide content. Benchmarking and Evaluation
The state-conditioned model is evaluated against mean/centroid baselines and standard deep models, with the largest gains in heterogeneous and data-sparse regimes. Generalization is explored via hold-out cell types and states to test extrapolation beyond the training manifold. Beyond loss, task-aware metrics include effect-size fidelity, rank correlation for differential expression, and hit enrichment for key targets. Ablations show that removing covariates or collapsing state/effect weakens robustness.
Use Cases in Modern Genomic Research
- CRISPR perturbation screens: Predict state-specific transcriptional responses to edits or gene knockdowns.
- Drug response modeling: Estimate effects across diverse cell states.
- Design of experiments: Prioritize informative perturbations for rare or clinically relevant subpopulations.
- Cross-study integration: More reliable transfer across labs and protocols via explicit covariate handling.
Future Directions
Looking ahead, Abhinav highlights several priorities for advancing state-conditioned perturbation modeling. He emphasizes the need for fair evaluation frameworks that separate seen and unseen perturbations to measure true generalization. Dataset-specific ablations will help tune choices such as cell-set and batch sizes for different experimental contexts. Stronger baselines, including simple statistical measures like L2 distance, are essential to ensure that performance gains are meaningful rather than incremental. Task-aware metrics such as effect-size fidelity, rank agreement, and PR-AUC can provide more biologically relevant assessments than loss values alone. Together, these steps aim to produce models that are more rigorous, interpretable, and useful for real-world perturbation studies.
Final Thoughts
State-conditioned perturbation modeling brings us closer to answering the fundamental biological question: what happens in this specific cell when we intervene? By separating state from effectand grounding evaluation in strong baselines, it provides a practical path to more accurate and interpretable predictions in single-cell biology. This approach helps align computational models with experimental reality, opening the door to more robust and actionable perturbation science.
The full paper on state-conditioned perturbation modeling is available on bioRxiv.
Join us in our mission to push the frontiers of AI-driven biology and strive to make a lasting impact on medicine, biotechnology, and human health. We are hiring across teams. Visit our Careers page to learn more and apply. Follow us on X, YouTube, and LinkedIn.
