Canopy

Metabolic engineering — reprogramming microbes to manufacture chemicals and ingredients through fermentation — depends on connecting deeply heterogeneous data: gene sequences, enzyme functions, metabolite structures, reaction and pathway networks, host strains, and the outcomes of past fermentation experiments. These entities live in different modalities and separate databases, and no single learned representation has captured them jointly, which forces strain-design decisions to rely on fragmented evidence.

Canopy, developed by Twig Bio, is a heterograph foundation model that unifies these data types in one representation for metabolic and strain engineering. It assembles a large biological knowledge graph and applies a Heterogeneous Graph Transformer over it, using frozen embeddings from established modality-specific models as node features: ESM-2 for protein sequences, MoLFormer for small-molecule chemistry, and PubMedBERT for text. The work was presented at the ICML GenBio Workshop.

Rather than train new modality encoders from scratch, Canopy fuses pretrained embeddings as node features and learns to reason over the graph's relational structure. A single pretrained backbone then serves downstream tasks — such as predicting fermentation titer — through lightweight probes on frozen embeddings, without fine-tuning the underlying network.

#Key Features

• Heterogeneous knowledge graph: A graph of roughly 6.9 million nodes spanning 13 node types — genes, enzymes, metabolites, pathways, strains, and fermentation experiments — connected by 34 edge types and assembled from 10 data sources.
• Multimodal frozen node encoders: Protein, chemistry, and text nodes are initialized with frozen ESM-2, MoLFormer, and PubMedBERT embeddings, letting the graph model integrate sequence, structure, and literature signals without retraining those encoders.
• Heterogeneous Graph Transformer backbone: Type-aware attention respects the distinct semantics of each node and edge type, so relationships between, for example, an enzyme and the metabolite it acts on are modeled explicitly.
• Four self-supervised objectives: Pretraining combines link prediction, masked node modeling, distance prediction, and contrastive clustering to learn transferable node representations from the graph alone.
• Frozen-embedding downstream use: The learned embeddings support lightweight probes for new tasks, demonstrated on fermentation titer prediction with an R² of 0.41, without retraining the backbone.

#Technical Details

Canopy is built on a Heterogeneous Graph Transformer applied to a biological knowledge graph of approximately 6.9 million nodes across 13 node types and 34 edge types, curated from 10 upstream sources covering genes, enzymes, metabolites, pathways, strains, and fermentation records. Node features come from frozen, pretrained embedding models — ESM-2 (protein), MoLFormer (chemistry), and PubMedBERT (text) — rather than encoders trained within the graph. The backbone is pretrained with four self-supervised objectives (link prediction, masked node modeling, distance prediction, and contrastive clustering). Its foundation-model behavior is shown by freezing the learned embeddings and attaching a lightweight probe to predict fermentation titer, reaching an R² of 0.41 without adapting the backbone.

#Applications

Canopy targets metabolic and strain engineers who must prioritize pathway, enzyme, and host choices before committing scarce wet-lab and fermentation resources. By embedding genes, enzymes, metabolites, and experimental outcomes in a shared space, it can score candidate designs, surface related enzymes or metabolites, and predict fermentation performance from prior data — capabilities that fit into an industrial biotechnology workflow for producing sustainable chemicals and ingredients, aligned with Twig Bio's engineering platform.

#Impact

Canopy applies the foundation-model and knowledge-graph paradigm to industrial metabolic engineering, a domain where the relevant data are unusually fragmented across modalities and databases. Reusing frozen multimodal embeddings as node features and demonstrating a frozen-embedding probe is a concrete signal that the learned representations transfer across tasks without retraining. As a workshop preprint, its results are in-silico and await peer review; no public code or weights have been released and the license is undisclosed, so external replication is not yet possible.