DLEPS

Conventional drug discovery begins with a target: a protein whose inhibition or activation is believed to move the disease. Obesity, hyperuricemia and nonalcoholic steatohepatitis have no such consensus target, which leaves a target-first pipeline with nothing to screen against. DLEPS — the Deep Learning based Efficacy Prediction System — sidesteps the problem by describing the disease instead of the target: supply the genes that go up and the genes that go down in the diseased state, and the system ranks compounds by how strongly they are predicted to push the transcriptome the other way.

The mechanism has two halves. A neural network reads a compound as a SMILES string and predicts the transcriptional change that compound induces, expressed over the 978 landmark genes measured by the L1000 assay and then expanded linearly to 12,328 genes. That predicted signature is scored against the user's up and down gene sets with a Connectivity Map-style enrichment statistic, and the sign is inverted so that signature-reversing compounds rank highest. Neither step is target-aware and neither is indication-specific: one fixed set of weights serves every disease, because the disease enters only as two lists of gene identifiers.

DLEPS was built by Zhengwei Xie's group at Peking University Health Science Center together with Beijing Gigaceuticals, and published in Nature Biotechnology in 2021. It predates the transformer-based perturbation models that followed it — DrugPT among them, which uses DLEPS as a baseline — and differs from most of them in taking no cell-line context as input at all.

#Key Features

• Disease specified as a gene set, not as training data: A new indication needs only an upregulated and a downregulated gene list; the shipped repository drives four independent signatures — browning, fibrosis, hyperuricemia and NASH — through the identical checkpoint.
• Frozen chemical encoder: The SMILES side reuses a grammar variational autoencoder pretrained on ZINC, held non-trainable while the expression head is fit, so chemical structure is represented independently of the L1000 supervision.
• Landmark-to-transcriptome expansion: A dense linear map lifts the 978 predicted landmark values to 12,328 genes, so enrichment is scored against the wider transcriptome rather than the assay's measured subset.
• Screening libraries included: The repository ships an FDA-approved drug set of 961 compounds and a natural product library of 2,719 compounds, formatted for direct scoring.
• Predictions taken to animals: Top-ranked candidates were tested in mouse disease models rather than left as in-silico rankings.

#Technical Details

The chemical encoder is the ZINC grammar VAE, which parses SMILES through a context-free grammar and produces a 56-dimensional latent code. A sampled latent vector feeds a four-layer fully connected head — three 1,024-unit ReLU layers and a 1,024-unit tanh layer, each with 0.25 dropout — terminating in a linear 978-output layer that regresses the chemically induced expression change. Training minimizes mean squared error on chemically induced transcriptional profiles from the L1000 project, with the encoder's weights frozen. On molecules absent from training, predicted and measured profiles correlate at a Pearson r of 0.74. Scoring converts predicted landmark values to absolute expression, applies the 978-to-12,328 matrix, subtracts a benchmark baseline, and computes a Kolmogorov-Smirnov enrichment score against the query gene sets.

Applied to obesity, hyperuricemia and nonalcoholic steatohepatitis, DLEPS nominated perillen, chikusetsusaponin IV and trametinib respectively, each of which produced disease-relevant effects in mouse models. The trametinib result implicated MEK-ERK signaling as a tractable axis for anti-NASH agents, an inference drawn from the predicted mechanism rather than from a prior target hypothesis.

#Applications

The system suits any program where a disease state can be written as a differential expression signature but no protein target is agreed: metabolic disease, fibrosis, aging phenotypes, and repurposing campaigns over approved drugs or natural product collections. Because the input is a SMILES string and two gene lists, a group with a published patient or model-organism signature can triage a several-thousand-compound library before committing any assay capacity, then take the top of the ranking directly into cell or animal work.

#Impact

DLEPS made the case that a chemical-structure-to-transcriptome predictor, coupled to Connectivity Map scoring, can carry candidates all the way from a gene signature to validated animal efficacy across several unrelated indications — and it has become a standard baseline for later perturbation-response models. Its practical caveats are real: the reference implementation targets TensorFlow 1.15 and Keras 2.3, the published GitHub tree contains the driver code and gene signatures but not the trained weight files, and the source carries an "all rights reserved" header with no repository license. The L1000 training measurements come from cultured cell lines, so a predicted signature is a cell-line-average response rather than a tissue-specific one, and efficacy in a mouse model of a metabolic disorder remains several steps short of clinical benefit.