AET5

A disease signature and a drug's transcriptional response are vectors over the same 978 landmark genes, so a compound that pushes expression opposite to the disease is, in principle, a therapeutic candidate. The obstacle is the measurement: L1000 profiles carry batch effects, platform drift and non-specific probe hybridization, and the usual ways of handing them to a generator make this worse. A linear projection has too little capacity to separate signal from noise, and a variational encoder is optimized for a smooth latent space — exactly the smoothing that blurs the perturbation-specific structure distinguishing one profile from another.

AET5, short for Anti-disease Expression T5, decouples the two halves of the problem and trains them separately. A gene-level transformer treats each landmark gene as a token and learns transcriptomic representations by contrastive self-supervision, with no molecules involved at all; that encoder is then frozen and used only to emit a conditioning vector. Molecular generation is handled by MolT5, a pretrained SMILES sequence-to-sequence model, adapted with lightweight adapters and selective layer unfreezing rather than retrained, so the chemical syntax it already knows survives intact. Because the disease enters only through the conditioning profile, retargeting means swapping the expression vector at inference, not refitting the generator.

The work comes from a group spanning The University of Hong Kong, Suzhou University of Science and Technology, Harbin Institute of Technology and Halmstad University. It sits alongside MolGene-E, which aligns single-cell expression with chemistry through paired variational autoencoders, and DLEPS, which ranks known compounds against a disease signature rather than generating new ones.

#Key Features

• Decoupled two-stage training: The transcriptome encoder and the molecular generator are trained independently, so representation learning is never entangled with the generation objective and either half can be replaced on its own.
• Augmentation designed for expression noise: Four operators — feature dropout, random scaling, adaptive Gaussian noise, and structured Gaussian noise drawn from the data's own regularized covariance — build the positive pairs, teaching invariances that match assay artifacts.
• Conditioning by tokens, not concatenation: The latent condition becomes a fixed-length sequence of tokens that enter the generator's encoder and reach decoding through cross-attention, a richer channel than a vector glued onto the input.
• Parameter-efficient adaptation: Most pretrained generator weights stay frozen while adapters, the condition-mapping network and selected upper layers are updated, keeping training viable on the small paired datasets this task supplies.
• Interpretable conditioning: Generator attention over the input loci can be read back out; 25 of the 978 genes matched prostate-cancer targets in Open Targets at confidence above 0.5.

#Technical Details

The encoder projects each of the 978 L1000 landmark genes into an embedding through a feed-forward network, processes the resulting gene tokens with a two-layer feature-wise attention stack, and compresses them through a projection head to a 128-dimensional latent optimized with InfoNCE at temperature 0.2; structured noise is sampled via Cholesky decomposition of the scaled empirical covariance. The generator starts from the MolT5 caption-to-SMILES checkpoint, the condition vector becomes 32 condition tokens, and training uses teacher forcing plus variant-SMILES augmentation. Paired L1000 drug-perturbation profiles and compound SMILES are split 9:1 — the released data covers the MCF7 cell line — while the SARS-CoV-2 (NGDC PRJCA002273, GEO GSE147507) and TCGA-PRAD prostate-cancer profiles are held out entirely as case-study conditions. Against eight baselines including GxVAEs, Gx2Mol, TransGEM and SmilesGEN, AET5 reached 100% validity, 100% uniqueness, 97.3% novelty, internal diversity 0.87 and FCD 6.09, produced 960 unique scaffolds against 953 in the reference set, and matched the reference QED of 0.60 at a synthetic accessibility of 3.32 versus 3.40. Dropping the pretrained initialization sends novelty to 48.1%.

#Applications

The intended use is hypothesis generation at the front of a hit-finding campaign, for a team that has a disease expression signature but no validated target. Inverting that signature and decoding under it yields chemically valid, largely unseen structures that can be filtered on synthetic accessibility and predicted ADME before docking. The two case studies span mechanistically unlike settings — SARS-CoV-2, where transcriptomic data are scarce and heterogeneous, and prostate cancer, where they are abundant but noisy — with candidates carried through AutoDock Vina docking and molecular dynamics against Nsp3, RdRp, CHEK2 and TUBB6.

#Impact

AET5's contribution is architectural rather than chemical: the encoder that reads the biology and the decoder that writes the molecule do not need to be trained together, and keeping them apart buys robustness where the conditioning data are sparse and noisy. Removing either the learned transcriptomic embedding or the generator's pretraining degrades every generation metric. The limits are equally clear. All validation is computational — docking, dynamics and predicted pharmacokinetics, with no assay behind any candidate — robustness to input noise varied by disease and decoding strategy, and no target structures, pathway annotations or interaction networks constrain generation, so mechanistic control remains indirect.