Reinforcement Learning Launches a New Paradigm for Spatial Omics Experiment Design

Spatial omics can measure thousands of RNAs and proteins at cellular or sub‑cellular resolution, but selecting the optimal field‑of‑view (FOV) is expensive and time‑consuming.

SOFisher reinforcement‑learning framework

SOFisher equips spatial‑omics instruments with a “look‑and‑move” capability, replacing blind scanning with adaptive sampling guided by a reinforcement‑learning (RL) policy.

Biological assumptions

Non‑randomness assumption – tissue exhibits highly ordered spatial topology of cell phenotypes (cell types, gene‑expression profiles).

Association assumption – target tissue landmarks (TTLs) are strongly spatially correlated or causally linked with surrounding cellular phenotypes.

Under these assumptions each sampled FOV acts as a “compass”; the AI infers the current ecological niche from the observed cell‑type distribution and predicts the direction of likely landmarks.

Training‑testing information isolation

During training the model receives full cell‑type and landmark annotations, learning the association between phenotype patterns and landmark locations. During testing the model observes only the cell‑type composition of the current FOV, without any landmark labels, and navigates solely on the learned association.

Simulation experiments

Validation used real mouse primary motor‑cortex spatial transcriptomics data (64 slices, ~300 k cells). Simulated landmarks were placed with a 20 % probability around L4/5 IT cells. In 50 sampling steps SOFisher captured 2–5× more landmarks than random sampling and reached 10 landmarks with ≈60 % fewer steps.

Generalization across ages and FOV sizes

SOFisher maintained performance on mouse brain slices of different ages and with FOV diameters of 150 µm and 300 µm, demonstrating robust generalization.

Application to Alzheimer’s disease (AD) data

Using a pre‑trained SOFisher policy, inexpensive spatial single‑omics (cell‑type only) was performed on a few small FOVs. The AI guided sampling to ecological niches where Aβ plaques and p‑tau neurofibrillary tangles intersect. This limited‑view data reproduced the pathogenic cell subtypes and gene programs previously discovered only with large‑scale multi‑omics scans.

SOFisherWR variant with restart mechanism

SOFisherWR extends the reward function to multi‑modal spatial gene‑expression gradients. In highly heterogeneous tissues such as colorectal cancer, the model automatically triggers a “restart exploration” to ensure isolated tumor cores are not missed.

Implications

The study positions AI‑driven frameworks like SOFisher as a new embodied‑intelligence paradigm for biological instruments, enabling dynamic, data‑driven sampling from the first moment of contact.

Paper: https://www.nature.com/articles/s41467-026-73404-6

Code example

来源：ScienceAI
本文
约2000字
，建议阅读
5
分钟
强化学习驱动的FOV采样，仅凭细胞类型就能精准定位。