AROMA

<p align="center"> <img src="figures/logo.jpg" alt="AROMA Logo" width="120"> </p> <h2 align="center"> 🧬 AROMA: Augmented Reasoning Over a Multimodal Architecture for Virtual Cell Genetic Perturbation Modeling<br>(ACL 2026 Findings)</h2> <p align="center"> 📃 <a href="https://arxiv.org/pdf/2604.20263" target="_blank">Paper</a> • 🐙 <a href="https://github.com/blazerye/AROMA" target="_blank">Code</a> • 🗂️ <a href="https://huggingface.co/datasets/blazerye/PerturbReason" target="_blank">Datasets</a><br> </p> </p>

Please refer to our repository and paper for more details.

🌐 Overview

AROMA is a novel multimodal architecture for virtual cell modeling that integrates textual evidence, graph topology, and protein sequences to predict the effects of genetic perturbations.

<p align="center"> <img src="figures/overview.jpg" alt="Overview"> </p>

The overall AROMA pipeline is illustrated in the figure above and is divided into three stages:

• Data stage. AROMA constructs two complementary knowledge graphs and a large-scale virtual cell reasoning dataset for evidence grounding.

• Modeling stage. AROMA adopts a retrieval-augmented strategy to incorporate query-relevant information, thereby providing explicit evidence cues for prediction. In addition, it jointly leverages topological representations learned from graph neural networks (GNN) and protein sequence representations encoded by ESM-2, and applies a cross-attention module to explicitly model perturbation-target gene dependencies across modalities.

• Training stage. AROMA first performs multimodal supervised fine-tuning (SFT), and is then further optimized with Group Relative Policy Optimization (GRPO) reinforcement learning to enhance predictive performance while generating biologically meaningful explanations.