Deciphering how drugs alter cellular gene expression is essential for new drug screening and personalized medical treatment. Scientists mainly rely on high-throughput sequencing to capture drug perturbation transcriptomic profiles, yet this experimental method suffers from high costs and long testing cycles, making it impossible to cover countless combinations of chemical compounds, cell lines and administration dosages. Traditional computational prediction tools show obvious performance drops when handling unseen drugs or cell types, and lack a universal coding system to fit diverse biological perturbations. While generative diffusion frameworks have achieved outstanding performance in text and image generation, few studies adapt this technology to predict post-drug transcriptional responses from baseline cell expression data. Against this gap, the research team develops a novel generative tool to cut experimental costs and boost prediction accuracy for drug development.

Recently, a research team from Prof. Jin Gu at Tsinghua University published an article titled “Predicting drug‐perturbed transcriptional responses using multi‐conditional diffusion transformer” in Quantitative Biology. The team built a multi-condition diffusion computing framework named PertDiT to predict drug-triggered gene expression shifts only using baseline cellular transcriptome and chemical molecule information. This work designs two tailored network branches for distinct research scenarios and creates a universal text-based pipeline to convert drug structures into computable features. Systematic tests prove the tool outperforms mainstream existing algorithms on untrained drugs, untrained cell lines, and organ toxicity prediction tasks, and single-gene analysis verifies its reliability in capturing subtle expression changes.

Figure 1 illustrates the complete structural workflow of the multi-condition diffusion prediction system, which estimates post-drug gene expression profiles by taking untreated cellular baseline transcriptome data and drug molecular information as inputs. The whole workflow consists of three sequential functional modules. The first module is the perturbation representation unit, which converts chemical structure codes alongside textual descriptions of dosage and compound types into standardized computational drug features via two text-based pre-trained models. The second module serves as data preprocessing adapters, containing three independent units to standardize drug features, raw cellular gene expression data and timestep parameters for diffusion computation; it also supports concatenating baseline and noisy transcriptional data for unified processing. The third module includes two independent computational branches for feature fusion. The first branch processes baseline cell data and drug information separately and integrates the two data types through cross matching operations, which fits the scenario of screening new candidate drugs. The second branch first extracts inherent expression patterns from concatenated cellular datasets before merging drug information via cross matching. Both branches stack multiple normalization layers and feature extraction units iteratively to generate final predicted transcriptional profiles.
DOI
10.1002/qub2.70016