Learning Molecular Representations from Cellular Phenotypes with Structure Preservation

When scientists test thousands of compounds in cellular assays, they generate a wealth of phenotypic data that links chemical structure to biological outcome. Turning that noisy signal into useful molecular representations remains a challenge, especially when models prioritize aligning chemistry with phenotype at the expense of preserving the underlying chemistry.

What You Need to Know

The paper introduces PhenMol, a multimodal framework designed to learn molecular embeddings that are both phenotype‑aware and structurally faithful. Instead of forcing a single shared space for molecules and cells, PhenMol splits each modality into two parts: a shared component that captures phenotype‑relevant information and a private component that retains modality‑specific details. For molecules, the private branch is a dedicated encoder that is trained to reconstruct the input graph, ensuring that important substructures, stereochemistry, and physicochemical properties are not lost during alignment.

The cellular side uses a similar split, allowing the model to learn which aspects of the phenotype are truly driven by chemical structure versus experimental noise or cell‑line‑specific effects. The shared space is then optimized with a contrastive loss that pulls together matched molecule‑phenotype pairs while pushing apart mismatched ones. Because the molecular private branch is decoupled from this alignment pressure, the resulting embeddings retain a high fidelity to the original chemical graph, as measured by reconstruction accuracy and preservation of scaffold similarity.

Experiments on public phenotypic screens (e.g., the JUMP‑CP dataset) show that PhenMol improves downstream tasks such as activity prediction and compound‑pair similarity ranking compared with baseline methods that omit the private molecular branch. Ablation studies confirm that removing the structure‑preserving branch leads to a drop in both reconstruction quality and predictive performance, confirming the trade‑off that the design alleviates.

Why It Matters

In drug discovery, the ability to generalize from screened compounds to novel chemotypes hinges on representations that capture both the biology of interest and the essential chemistry. Distorted embeddings can cause virtual screening to miss active scaffolds or to over‑emphasize irrelevant features, wasting resources on false leads. By explicitly guarding chemical information, PhenMol offers a more reliable foundation for tasks like lead optimization, scaffold hopping, and interpreting structure‑activity relationships.

Beyond early‑stage screening, the approach can be adapted to other multimodal settings where one modality has a well‑defined structural grammar (e.g., proteins, materials). The disentanglement strategy provides a principled way to balance cross‑modal signal with modality‑specific integrity, a balance that is increasingly important as phenotypic profiling scales to millions of measurements.

Key Details

• Molecular encoder consists of a graph‑isomorphism network (GIN) followed by a private reconstruction head that predicts node and atom features.
• Cellular encoder uses a multilayer perceptron on raw phenotypic profiles, also split into shared and private heads.
• The shared space is trained with a normalized temperature‑scaled cross‑entropy loss (contrastive) over 10,000 matched molecule‑phenotype pairs.
• Reconstruction loss for the molecular private branch is weighted at 0.5 relative to the contrastive loss, tuned via validation on scaffold split.
• Ablation shows that removing the molecular private branch drops top‑1 retrieval accuracy from 0.42 to 0.31 on held‑out phenotypes.
• Embedding similarity computed from PhenMol correlates more strongly with Tanimoto similarity (r = 0.63) than baseline methods (r = 0.48).

What’s Next

Future work will explore extending PhenMol to multi‑task phenotypic panels, incorporating temporal data from live‑cell imaging, and evaluating the embeddings in prospective synthesis campaigns. The authors also plan to release pretrained models and a benchmark suite to facilitate comparison with other structure‑preserving multimodal approaches.

📌 Source: Arxiv Ml