GNN

This page describes the GNN row of the Papers.md matrix: graph neural networks, meaning learned message-passing models over graph-structured data, including graph convolutional networks, graph attention networks, graph isomorphism networks and graph autoencoders. The row’s authoritative scope is its Taxonomy.md definition; this page synthesizes what currently sits in it.

Scope boundary

The taxonomy draws a line that is easy to cross by accident: the model must be a trained, message-passing network. Classical graph algorithms, random-walk network propagation, and non-learned graph embeddings such as metapath2vec or node2vec are not GNNs and do not belong here, even though all of them operate on graphs and all of them appear in cell-ag work. A paper using node2vec embeddings as features for a downstream classifier sits in Deep Learning or wherever that classifier belongs.

The row splits cleanly into two halves that share a mathematical object and nothing else. In Cellular Engineering the graph is inferred: cells or genes are nodes and the edges are a modelling choice the paper has to defend. In Sensory Prediction the graph is given: a molecule or a peptide has bonds, and the graph is the chemistry. The second half is easier and the results are correspondingly stronger.

Cellular Engineering

#13 scGNN (Wang et al. 2021, Nature Communications): a hypothesis-free framework that formulates and aggregates cell-cell relationships with graph neural networks while modelling gene expression with a left-truncated mixture Gaussian. That second component is the distinctive part and is aimed squarely at scRNA-seq dropout: expression of a gene across cells is treated as a mixture of Gaussians corresponding to discrete regulatory signals, with the left-truncation assumption splitting truly measured values from left-censored ones, and the number of components chosen by Bayesian information criterion. Three iterative multi-modal autoencoders sit on top. Evaluated on gene imputation and cell clustering across four benchmark datasets and applied to 13,214 single nuclei in an Alzheimer’s study. Code at juexinwang/scGNN.
#8 graph-sc (Ciortan & Defrance 2022, Bioinformatics): models the expression matrix as a gene-to-cell graph and runs a graph autoencoder to produce cell embeddings, which are then clustered by K-means or Leiden. The preprocessing is reported as a tuned choice rather than a convention, with genes expressed in fewer than two cells dropped, per-cell count normalization, and the top 3,000 most variable genes retained, all justified by an ablation showing it maximizes clustering performance. The paper is candid that there is still no consensus best clustering method for scRNA-seq, which is the honest framing for this whole cluster. Code at ciortanmadalina/graph-sc.
#12 GLAE (Shan et al. 2023, Information Sciences): attacks the assumption the two papers above rely on. Most GNN-based scRNA-seq methods build a pre-fixed cell-relation graph from all genes at once and feed it to the network, which means sequencing error in the graph propagates into every downstream result, and which ignores that different gene subsets imply different cell relations. GLAE instead learns the cell-relation graph end to end from several sub-modules, each looking at a different feature subset, and aggregates their graphs and embeddings iteratively. For anyone deciding whether to use a GNN on their own culture’s scRNA-seq, this is the paper that names the failure mode.
#121 GEARS (Roohani et al. 2024, Nature Biotechnology, Leskovec lab): predicts the transcriptional outcome of single and multi-gene perturbations, including combinations of genes never experimentally perturbed, which is what makes it useful given the combinatorial explosion in possible multigene screens. Its structure is worth stating precisely because it is what keeps GEARS in this row rather than a foundation-model one: a GNN encoder maps each perturbation to an embedding over a gene-gene knowledge graph, a second GNN encoder maps each gene to an embedding over a coexpression graph, a compositional module combines them, and a cross-gene decoder emits the post-perturbation expression vector, trained end to end with a direction-aware loss. Genes are learnable embeddings rather than scalars, so relative heterogeneity of perturbational response is representable. Reported roughly 40% higher precision than prior approaches on genetic-interaction subtypes. It is trained on one perturbation dataset rather than pretrained and transferred, which is exactly why the taxonomy places it here and not in Cell-State & Perturbation Prediction. Code at snap-stanford/GEARS.

Sensory Prediction

#14 Principal Odor Map (Lee et al. 2023, Science, Osmo and collaborators): the row’s landmark result. A graph neural network over molecular structure produces a 256-dimensional embedding, the principal odor map, that preserves perceptual relationships and predicts odor quality for uncharacterized odorants. On a prospective validation set of 400 out-of-sample molecules the model’s odor profile matched the trained panel mean more closely than the median human panelist did. Training data came from the Good Scents and Leffingwell collections, with descriptor variants and misspellings merged and any descriptor occurring 30 or fewer times discarded, leaving a vocabulary that is itself informative: it includes beefy, meaty, fatty, roasted and savory, which are precisely the descriptors a cultivated-meat sensory panel works in. Code at osmoai/publications.
#36 (Qian et al. 2023, eLife): dissects that model rather than building a new one, freezing the pretrained network and using the 256-dimensional embedding as a fixed representation. It finds that molecules close together in the map are more likely to co-occur in a natural substance and to be metabolically closely related, and that the representation predicts receptor, neural and behavioural responses across nearly all terrestrial organisms studied in olfactory neuroscience. The claim, that odour space is organized by metabolism, matters for cell-ag because the volatiles a cultured tissue emits are metabolic products, so the map’s axes are not arbitrary with respect to what a bioprocess can change. Code at osmoai/publications.
#286 BitterPep-GCN (Srivastava et al. 2024, Journal of Cheminformatics): a feature-agnostic graph convolutional network for bitter-peptide prediction, representing a peptide as a simple path graph with amino acids as nodes, bonds as edges and one-hot node encoding, learning amino-acid embeddings rather than using hand-engineered descriptors, and pooling for classification. Benchmarked on BTP640. The dataset caveat is stated in the paper and should travel with any use of it: the 320 bitter peptides were experimentally validated, but the 320 non-bitter ones were sampled from BIOPEP and were not experimentally shown to be non-bitter, so the negative class is an assumption. Code at srivastavaprashant/BitterGNN.
#285 (Steuer et al. 2026, npj Science of Food): uses BitterPep-GCN as the classifier that filters language-model-generated peptides, and is the loop closed to a human taste panel. Described in Foundation Models: Next-Token Prediction.
#320 BitterGNNs (He et al. 2024, Bioinformatics Advances): a modular architecture search rather than a single model. Four GNN modules (an edge-attention module the authors introduce, plus GATv2, MPNN and Weave) are each given a fixed position and a present-or-absent flag, generating 16 combinations that are compared against five popular GNNs on public datasets, with the best four retained as the predictor. Reported AUC 0.87 on both bitter versus non-bitter and bitter versus sweet, beating fingerprint-based baselines. The bitter-versus-sweet task is the harder and more useful one for formulation work.
#324 (Song et al. 2023, Foods): a comparative study over 2,601 molecules curated from ChemTastesDB, testing molecular feature representations against learning algorithms for taste prediction. GNN-based models outperform the alternatives, but the paper’s actual finding is that a consensus of molecular fingerprints plus a GNN beats either alone, which is evidence that the learned graph representation and the hand-engineered one capture different things. It also reports the class distribution honestly, 977 sweet and 1,183 bitter against 38 sour and 12 salty, which is the same data cliff #342 runs into.

Metabolic Modeling

#68 D2Cell (Li et al. 2024, bioRxiv): mines metabolic-engineering strategies from the literature at scale using LLMs, over abstracts from 10,000 Web of Science, PubMed and Scopus articles plus 1,340 open-access full texts, through a three-stage pipeline of named-entity recognition, relation extraction, and entity resolution. One detail explains why it works: strain IDs are extracted as entities even though they are unstandardized author inventions like QW101, because they are what lets successive genetic modifications within one study be traced to the same strain. The result is a database of over 29,006 metabolic-engineering entries across 1,210 products and 751 organisms, feeding a hybrid deep-learning and mechanistic model that predicts engineering targets. Also in Domain-Specific Biomedical Agents. Code at LiLabTsinghua/D2Cell.

AI Tooling / Methodology

#52 BioMedReasoner (Mulyadi et al. 2025, NeurIPS AI for Science workshop): multi-hop reasoning over biomedical knowledge graphs via path-based relational learning, framing link prediction as estimating the probability of a triple under a query relation, following NBFNet. Node features are initialized from the query relation itself at the source node and updated by message passing, with a permutation-invariant aggregation and a feed-forward scoring head. It is in the tooling column because it is a general knowledge-graph reasoning method rather than a cell-ag application, and in this row rather than Agent Infrastructure because the contribution is a trained message-passing model, not a graph to query. Note this ref’s evidence is thinner than the rest of the row: it is a workshop poster and the corpus carries no abstract for it.

Adjacent methods

Foundation Models: Cell-State & Perturbation Prediction: where a pretrained-and-transferred perturbation model belongs, as opposed to GEARS.
Deep Learning: the catch-all for networks not better described by a graph architecture, including anything using non-learned graph embeddings.
GAN / VAE: the other route to a single-cell embedding, and the generative counterpart to the graph autoencoders here.
Foundation Models (other modalities): NaFM’s graph pretraining, and the protein-language-model peptide predictors these GCNs are benchmarked against.
Agent Infrastructure (Frameworks, KGs, Protocols): curated and queried biomedical knowledge graphs, as opposed to learned message passing over them.

GNN