AROMA: Gene Perturbation Prediction Model

• AROMA is a virtual cell modeling system for predicting gene expression changes due to gene perturbations, providing evidence-grounded explanations.
• It uses a retrieval-augmented multimodal architecture that integrates textual evidence, graph-based pathway data, and protein sequence information.
• Empirical evaluations across multiple cell lines demonstrate that AROMA improves prediction accuracy and interpretability over previous models.

Searching arXiv for the primary AROMA paper and closely related virtual-cell perturbation modeling work. AROMA is a virtual cell modeling system for genetic perturbation prediction that estimates how the expression of a target gene changes—upregulated, downregulated, or non-differentially expressed—when a perturbation gene is genetically perturbed in a specified cell line, while also generating an evidence-grounded explanation for that prediction (Wang et al., 22 Apr 2026). It is formulated around queries of the form: in cell line $c$, if gene $g_p$ is perturbed, what happens to gene $g_t$, and why? Within this framework, AROMA addresses three limitations identified in prior approaches: unconstrained reasoning, uninterpretable predictions, and topology-agnostic retrieval. Its central design is a retrieval-augmented multimodal architecture that integrates textual evidence, graph-topology information, and protein sequence features, and that is trained to output both a ternary label and a reasoning trace (Wang et al., 22 Apr 2026).

1. Task formulation and virtual cell setting

AROMA operates on triplets $x = (g_p, g_t, c)$, where $g_p$ is the perturbation gene, $g_t$ is the target gene, and $c$ is the cell line context. The prediction label is $y \in \{\text{Up}, \text{Down}, \text{Non-diff}\}$, indicating how the expression of $g_t$ changes under perturbation of $g_p$ in cell line $g_p$0 (Wang et al., 22 Apr 2026). The model is trained as

$g_p$1

Unlike standard classifiers, however, the output is not limited to $g_p$2; the system must also emit a reasoning sequence that explains the label in terms of genes, pathways, and cell context (Wang et al., 22 Apr 2026).

The underlying labels are derived from single-cell perturbation data. Raw UMI counts $g_p$3 are normalized using log-transformed counts-per-10k,

$g_p$4

and perturbed cells are compared with non-targeting control cells using a Wilcoxon test with BH-adjusted $g_p$5-values $g_p$6. The mean expression difference for gene $g_p$7 is

$g_p$8

Labels are assigned as Up when $g_p$9 and $g_t$0, Down when $g_t$1 and $g_t$2, and ND otherwise (Wang et al., 22 Apr 2026).

This formulation makes AROMA a gene–gene perturbation-effect model conditioned on cellular context. A plausible implication is that it occupies an intermediate position between transcriptome-level perturbation simulators and purely language-based reasoning systems: it is more structured than free-form explanation models, but narrower than models that jointly predict the full post-perturbation expression state.

2. Data resources and knowledge representation

A central component of AROMA is the construction of reusable resources for the virtual cell domain. The first is PerturbReason, a perturbation reasoning dataset containing more than 498k samples (Wang et al., 22 Apr 2026). Each sample augments perturbation labels with retrieved evidence and synthetic reasoning traces. For each perturbation instance, the pipeline retrieves perturbation and target gene functional descriptions from the SUMMER knowledge base, up to three shortest paths between $g_t$3 and $g_t$4 in the pathway graph, and a cell-line description from sources such as Wikipedia. These are assembled into an instruction-style prompt, and DeepSeek-R1 is used to generate a > ...</think> reasoning trace and a final <answer>Up|Down|not significantly changed</answer> output, filtered for label correctness and structural validity (Wang et al., 22 Apr 2026).

The second major resource is the pair of knowledge graphs used both for retrieval and structural encoding. The Gene Knowledge Graph (Gene-KG) is constructed from STRING and CORUM and contains 18,479 genes and 752,612 undirected association edges, with average degree 81.46 (Wang et al., 22 Apr 2026). CORUM complexes are expanded into pairwise edges, and all edge types are treated as statistical associations rather than explicit causal relations. The Pathway Knowledge Graph (Path-KG) is built from Gene Ontology and Reactome, with generic GO terms removed and hub nodes of degree at least 1000 filtered out. It contains 80,020 nodes and 441,176 edges, with average degree 11.03 (Wang et al., 22 Apr 2026).

These graphs play complementary roles. Gene-KG is a dense gene–gene association graph, whereas Path-KG is a multi-typed graph linking genes to processes and pathways and enabling path extraction between perturbation and target genes (Wang et al., 22 Apr 2026). This distinction is crucial to AROMA’s claim of topology-aware retrieval: the textual evidence is not gathered by semantic similarity alone, but partly by shortest-path structure in Path-KG.

Resource	Contents	Size
PerturbReason	Reasoning dataset for SFT	more than 498k samples
Gene-KG	Gene–gene association graph from STRING + CORUM	18,479 nodes, 752,612 edges
Path-KG	Gene/process/pathway graph from GO + Reactome	80,020 nodes, 441,176 edges

The paper emphasizes that these resources are reusable. This suggests that AROMA is not only a single model instance but also an attempt to formalize a perturbation-reasoning substrate for subsequent virtual cell research.

3. Multimodal architecture

AROMA is described as a retrieval-augmented domain LLM with a Qwen3-8B backbone, enriched with graph and protein modalities (Wang et al., 22 Apr 2026). Its textual input for a query $g_t$5 is

$g_t$6

where $g_t$7 comprises gene functional descriptions, $g_t$8 comprises up to three shortest paths between $g_t$9 and $x = (g_p, g_t, c)$0 in Path-KG extracted by BFS on hop count, and $x = (g_p, g_t, c)$1 is a curated cell-line description (Wang et al., 22 Apr 2026). These components are concatenated into a prompt that includes placeholders for multimodal embeddings such as <ESM-2_EMB>, <Path-KG_EMB>, and <Gene-KG_EMB>.

For graph topology, AROMA uses two separate Graph Attention Network (GAT) encoders, one for Gene-KG and one for Path-KG, each with 3 layers, 4 attention heads, hidden dimension 64, output dimension 1024, ELU activation, and independent parameters (Wang et al., 22 Apr 2026). The layer update is

$x = (g_p, g_t, c)$2

These encoders are pretrained by link prediction using

$x = (g_p, g_t, c)$3

and binary cross-entropy loss

$x = (g_p, g_t, c)$4

The reported pretraining AUROC is 0.95 for Gene-KG and 0.92 for Path-KG (Wang et al., 22 Apr 2026). At inference time, AROMA extracts default 1-hop subgraphs around both $x = (g_p, g_t, c)$5 and $x = (g_p, g_t, c)$6 and mean-pools the resulting embeddings into perturbation- and target-specific graph representations.

Protein sequence information is supplied by ESM-2 (esm2_t48_15B_UR50D), kept frozen (Wang et al., 22 Apr 2026). For perturbation and target proteins with sequences $x = (g_p, g_t, c)$7 and $x = (g_p, g_t, c)$8,

$x = (g_p, g_t, c)$9

where $g_p$0 denotes mean pooling over amino-acid token embeddings (Wang et al., 22 Apr 2026).

A distinctive element of the architecture is that it models perturbation–target interactions explicitly rather than embedding genes independently. For each modality $g_p$1, AROMA computes a cross-attention interaction feature: $g_p$2 projects it into the LLM embedding space,

$g_p$3

and injects the projected vectors at predefined placeholder tokens in the textual prompt (Wang et al., 22 Apr 2026). The resulting LLM input combines retrieved text with modality-specific interaction embeddings, so the LLM conditions on relation-level evidence rather than isolated gene descriptors.

4. Augmented reasoning and interpretability

The paper defines “augmented reasoning” as reasoning that is not performed in a vacuum but is forced to condition on retrieved evidence, graph structure, and protein-derived interaction signals (Wang et al., 22 Apr 2026). This is presented as a direct response to unconstrained reasoning in general or synthetic-trace LLMs, which can hallucinate biologically implausible causal narratives. AROMA instead grounds its generation in explicit gene summaries, pathway paths, cell-line descriptions, local KG neighborhoods, and protein features (Wang et al., 22 Apr 2026).

The output format is explicitly structured: $c$0 During supervised fine-tuning and reinforcement learning, the model is trained to produce both the reasoning trace and the answer, and the <think> section is encouraged to reference pathways, cell-type-specific biology, molecular functions of $g_p$4 and $g_p$5, and possible buffering or compensatory mechanisms (Wang et al., 22 Apr 2026).

Interpretability arises from several sources. First, the input evidence is itself human-readable: gene summaries, cell-line descriptions, and explicit pathway strings such as TULP1 → photoreceptor inner [segment](https://www.emergentmind.com/topics/segment) → DRAM2 → apoptotic process → ITCH (Wang et al., 22 Apr 2026). Second, the reasoning trace is trained to align with that evidence. Third, the paper reports sentence-level source-tracing case studies in which reasoning statements are manually aligned to retrieved evidence fields such as "Perturbation_Gene_summary_1", "Pathway_text", and "Cell_line_description" (Wang et al., 22 Apr 2026). No separate interpretability metric is introduced, but the case-study methodology is used to argue that the generated explanations are evidence-supported rather than arbitrary.

The retrieval design is especially important here. Path evidence is generated from shortest paths in Path-KG rather than generic text retrieval, so the explanation is topologically aligned with the graph of plausible biological relations (Wang et al., 22 Apr 2026). This does not make Path-KG causal—the paper explicitly states that the edges are not strictly causal—but it constrains retrieval to biologically organized routes rather than unconstrained similarity matches.

A plausible implication is that AROMA’s interpretability is best understood as evidence-traceability rather than mechanistic proof: the model can indicate which retrieved entities and relations supported its answer, even if the underlying KG edges are association-level.

5. Training procedure and optimization strategy

AROMA is trained in two stages. Stage 1 is multimodal supervised fine-tuning on PerturbReason, where each example consists of retrieval-augmented text with injected multimodal embeddings as input $g_p$6, and the concatenated reasoning trace plus final answer as output $g_p$7 (Wang et al., 22 Apr 2026). The loss is standard autoregressive language modeling: $g_p$8 During this phase, the pretrained GNN encoders and ESM-2 are frozen, while the cross-attention modules, projectors, and the Qwen3-8B backbone via LoRA adapters are trainable (Wang et al., 22 Apr 2026).

Stage 2 uses Group Relative Policy Optimization (GRPO), described as a PPO-like reinforcement learning algorithm (Wang et al., 22 Apr 2026). In this phase all encoders and projection modules are frozen, and only the LLM LoRA parameters are updated. For each instance, multiple reasoning trajectories are sampled—16 completions in the main experiments—and rewarded according to label correctness and output format. The reward is +5.0 for a correct label and -1.0 otherwise, plus +0.5 if the reasoning trace follows the <think>... format and +0.5 if the answer contains exactly one valid category (Wang et al., 22 Apr 2026). The objective is written abstractly as

$g_p$9

with $g_t$0 the reference policy and $g_t$1 the KL regularization coefficient (Wang et al., 22 Apr 2026).

The rationale for the two-stage scheme is explicit. Supervised fine-tuning teaches the model domain knowledge, evidence formatting, and the joint reasoning-plus-answer output distribution, whereas GRPO supplies task-level feedback that penalizes plausible-but-wrong conclusions and rewards correct, well-formed trajectories (Wang et al., 22 Apr 2026).

The paper’s ablation study summarizes the resulting gains. A vanilla Qwen3-8B backbone without domain training achieves Macro-F1 around 0.26; SFT alone raises it to around 0.65; SFT plus GRPO improves it further to around 0.68; and adding retrieval plus multimodal modules yields around 0.71–0.73 average (Wang et al., 22 Apr 2026). This indicates that the largest single improvement comes from reasoning data plus SFT, while RL and multimodal evidence provide additional complementary gains.

6. Empirical performance and robustness

Evaluation is conducted on four human cell lines: K562, HepG2, Jurkat, and RPE1, using PerturbQA-based splits and Macro F1 over the three classes ND, Up, and Down (Wang et al., 22 Apr 2026). Baselines include general LLMs prompted with textual context—DeepSeek-R1, OpenAI o4-mini, GPT-5, Gemini-2.5-pro, Qwen3-235B—and domain-specific models including SynthPert, GAT, STATE, GEARS, scGPT, GenePT-Gene, GenePT-Prot, and SUMMER (Wang et al., 22 Apr 2026).

Overall results place AROMA at approximately 0.73 average Macro F1, whereas general LLMs lie around 0.33–0.41 and the strongest domain baselines around 0.64 (Wang et al., 22 Apr 2026). On K562, for example, AROMA achieves per-class F1 of 0.96 for ND, 0.66 for Up, and 0.36 for Down, compared with SUMMER’s 0.94, 0.56, and 0.25 and GAT’s 0.94, 0.62, and 0.21 (Wang et al., 22 Apr 2026). The gains are concentrated especially in the Up and Down classes, which are harder and more imbalanced.

The zero-shot evaluation is notable. Training on K562, HepG2, and Jurkat and testing on unseen RPE1 yields zero-shot AROMA performance of ND 0.96, Up 0.57, Down 0.66, average around 0.73, compared with in-domain RPE1 training performance of ND 0.97, Up 0.60, Down 0.73, average around 0.77 (Wang et al., 22 Apr 2026). The drop is modest, suggesting transfer of cross-cell-line structure through external knowledge and multimodal reasoning.

Robustness is further examined in long-tail and knowledge-sparse settings by stratifying test genes by PubMed mention frequency and Path-KG degree into High, Mid, and Low groups (Wang et al., 22 Apr 2026). On RPE1, the ablated model without RAG and multimodal components drops from 0.69 to 0.56 across popularity High to Low, whereas full AROMA drops only from 0.77 to 0.71; by degree, the ablated model falls from 0.72 to 0.57, versus AROMA from 0.79 to 0.69 (Wang et al., 22 Apr 2026). The paper interprets this as evidence that the system is not merely memorizing well-studied genes but can maintain performance in long-tail regions via multimodal evidence and topology-aware retrieval.

Sensitivity analysis on GRPO trajectory count shows that sampling more reasoning trajectories during RL improves performance: around 0.70 average at 4 trajectories, around 0.72 at 8, and around 0.73 at 16, the setting used in main experiments (Wang et al., 22 Apr 2026).

7. Uses, limitations, and implementation details

The paper presents AROMA as a tool for target prioritization, mechanistic exploration, design of experiments, and regulatory network analysis (Wang et al., 22 Apr 2026). In target prioritization, a perturbation gene $g_t$2 can be queried against a panel of targets $g_t$3 in a given cell line to identify likely up- or down-regulated genes. In mechanistic exploration, the reasoning trace and explicit paths can be used to inspect how effects are hypothesized to propagate through pathways or cellular processes. In experimental design, the model can suggest which downstream genes or pathways to monitor and whether a non-differential prediction may reflect buffering or compensatory mechanisms (Wang et al., 22 Apr 2026).

The authors highlight three limitations. First, AROMA currently models single-gene perturbations only and does not directly address combinatorial perturbations or multi-target drug effects (Wang et al., 22 Apr 2026). Second, it predicts one target gene at a time rather than the full post-perturbation transcriptome jointly (Wang et al., 22 Apr 2026). Third, its performance depends on the quality and completeness of external KGs and text; genes with poor annotation or missing graph connectivity remain challenging, and performance may degrade on extreme unknowns beyond the long-tail regime evaluated (Wang et al., 22 Apr 2026).

Implementation details underscore the system’s engineering profile. The base LLM is Qwen3-8B fine-tuned with LoRA; GAT encoders use input dimension 1024, hidden 64, output 1024, 3 layers, 4 heads, ELU activation, and dropout 0.5, pretrained with AdamW at learning rate $g_t$4 for 200 epochs (Wang et al., 22 Apr 2026). SFT uses LoRA rank 16, $g_t$5, dropout 0.1, learning rate $g_t$6, batch size per device 4 with gradient accumulation 4, 3 epochs, and bfloat16 precision. GRPO uses learning rate $g_t$7, Adam $g_t$8, $g_t$9, weight decay 0.1, warmup ratio 0.1, cosine schedule, batch size per device 2 with gradient accumulation 8, 16 trajectories per query, max completion length 2048, and 2 epochs (Wang et al., 22 Apr 2026).

Model weights are released on Hugging Face and code on GitHub (Wang et al., 22 Apr 2026). The release framing suggests that AROMA is intended both as a callable perturbation-reasoning model and as a basis for more complex virtual cell agents, including multi-step experimental planners. This suggests a broader research trajectory in which perturbation prediction becomes one component of agentic biological reasoning systems, provided that the current limitations in perturbation scope, target granularity, and knowledge dependence are addressed.