MolGene-E

Conventional drug discovery starts from a target: pick a protein, find a molecule that binds it. Systems pharmacology proposes an inversion — start from the cellular state you want. If single-cell transcriptomics can describe precisely how a diseased cell differs from a healthy one, the design question becomes "what molecule would push this expression profile back?" That framing is attractive for diseases with no clean druggable target, but it has been blocked by the data: single-cell expression measurements are noisy, sparse, high-dimensional, and scarce relative to the bulk perturbation catalogs that most cheminformatics models were built around.

MolGene-E, from Lei Xie's group at Hunter College and the CUNY Graduate Center with collaborators at Weill Cornell Medicine, is a deep generative framework built to make that inversion tractable. It pairs two models: a cross-modal component that harmonizes and denoises chemically perturbed bulk and single-cell transcriptomic data into a shared representation, and a contrastive generative component that produces new molecular structures conditioned on that representation. The preprint was posted in February 2025 by Rahul Ohlan, Raswanth Murugan, Li Xie and colleagues, with a revision in 2026.

What distinguishes it from earlier expression-conditioned generators is the evaluation. Rather than reporting held-out performance within the same perturbation catalog it trained on, MolGene-E is assessed in a zero-shot setting on two independent datasets it never saw — CRISPR knockout signatures and single-cell drug-response profiles — which tests whether the learned expression-to-chemistry mapping transfers rather than memorizes.

#Key Features

• Bulk and single-cell harmonization: A denoising variational autoencoder maps noisy single-cell profiles and dense bulk profiles into one latent space, so the abundant bulk perturbation catalogs can supervise a task defined on scarce single-cell data.
• Contrastive expression-to-chemistry alignment: A CLIP-style supervised contrastive objective aligns gene-expression embeddings with molecular embeddings, giving the generator a conditioning signal grounded in a shared space rather than a bolted-on regression.
• SELFIES-based molecular decoder: Generation runs over SELFIES strings through a recurrent variational autoencoder pretrained on ZINC, a representation that makes every decoded string a syntactically valid molecule.
• Zero-shot transfer across perturbation types: The same trained model is applied to CRISPR knockout signatures and to single-cell drug-response profiles without refitting, testing generalization across both perturbation modality and assay platform.

#Technical Details

The gene-expression arm is a denoising VAE with hidden layers of 1024, 512 and 256 units compressing to a 128-dimensional latent. The chemical arm is a SELFIES variational autoencoder pretrained on ZINC, with a bidirectional GRU encoder and a three-layer GRU decoder of 512 hidden units. A projection network and a multilayer-perceptron prior map gene embeddings into the chemical latent space, trained with a supervised contrastive loss adapted from CLIP.

Training drew on L1000toRNAseq, an inferred whole-transcriptome expansion of the LINCS L1000 catalog covering 23,614 genes across 221 cell lines and more than 30,000 chemical perturbagens; for the MCF7 experiments this was reduced to 3,116 high-variance genes. Data were split 70/15/15 with no chemical overlap between splits. Zero-shot evaluation used 1,218 L1000 CRISPR knockout signatures across 44 transcription factors and the Sciplex-3 single-cell dataset of 188 compounds across three cancer cell lines. On the CRISPR benchmark, MolGene-E generated molecules with validity 1.00, uniqueness 1.00, novelty 1.00 and diversity 0.89, against 0.31 validity for MolGAN and 0.93 for GxVAE, with a mean synthetic accessibility score of 3.20. Docking-score distributions for generated molecules differed significantly from baselines by Kolmogorov-Smirnov test.

#Applications

The intended user is a systems-pharmacology or chemical-biology group that has a transcriptomic description of a disease state — a knockout signature, a patient-derived single-cell profile, a differentiation trajectory — and wants chemical starting points without first nominating a target. Practical uses include generating hit-like scaffolds for phenotypic screening campaigns, proposing chemistry for targets long considered undruggable, and prioritizing which compounds to synthesize when the desired outcome is a cell-state shift rather than inhibition of one protein.

#Impact

MolGene-E's contribution is the harmonization step: showing that bulk perturbation catalogs can be made to supervise molecule generation conditioned on single-cell measurements, rather than treating the two data types as separate problems. The zero-shot design of its evaluation is the stronger claim, since expression-conditioned generators frequently look strong within one catalog and collapse outside it. Two caveats belong on the record. The metrics reported — validity, uniqueness, novelty, diversity, docking scores — are in-silico proxies, and no generated molecule has been synthesized or tested experimentally in the published work. And no public code or weights repository accompanies the preprint, which was still awaiting peer review, so the results have not been independently reproduced.