Transcriptomics and AI are improving drug discovery

Artificial intelligence (AI) is accelerating the development of new prescription drugs¹. It uses large datasets, complex mathematical models, and advanced computational algorithms to help humans design new drugs, identify therapeutic targets, and conduct clinical trials.

Computer-aided drug design has improved the efficiency of new drug development. It uses molecular modeling to analyze the structure of potential drug candidates and predict their interactions with therapeutic targets, as well as their activity, toxicity, and bioavailability. This enables better planning and guides drug discovery. For example, captopril, saquinavir, ritonavir, indinavir, tirofiban, doxorubicin, zanamivir, aliskiren, and poprevir were discovered using virtual screening.

Researchers use natural language models such as BioGPT and Med PaLM to analyze microscopy and pathology data as well as chemistry models to improve predictions. They can also use digital twins and other models to follow each patient’s health and reactions to treatment. The goal is to predict, prevent, and personalize every patient’s treatment while enabling them and their caregivers to participate in the therapy.

The first drug discovered and designed entirely by generative artificial intelligence began human clinical trials. It is a novel anti-fibrotic small molecule called Rentosertib. It was developed by Insilico Medicine. The drug targets idiopathic pulmonary fibrosis and has successfully advanced into multi-regional Phase 2 clinical trials.

However, AI is only as good as the data that are used to train it. Much existing data on the efficacy and toxicity of currently approved drugs is based largely on animal testing. This has led to many costly failures, since human biology is unique and quite different from that of rats, mice, and other animals previously used in evaluating the safety and efficacy of new drug candidates. So, health care professionals and government agencies are encouraging the use of new analytical methods to test the efficacy and toxicity of new drug candidates².

The FDA aims to accelerate the validation and adoption of these innovative methods by working with the National Institutes of Health, the National Toxicology Program, the Department of Veterans Affairs, and the Interagency Coordinating Committee on the Validation of Alternative Methods. Instead of testing new drugs on animals, they can be tested using AI-based computational models of toxicity.

Machine learning methods use enormous amounts of data from patient records and previous clinical trials to generate predictive models of patient response to an intervention. The creation of digital twins from many patients can help evaluate the effects of treatments on each individual and identify adverse side effects in their earliest stages—before they become serious. This could decrease the number of patients required for a clinical trial and enhance postmarket surveillance once a drug is approved and begins being used by many people. Theoretically, the model’s accuracy would progressively increase with each subsequent trial based on the newly generated data.

They can also be tested using cells, organs on a chip and organoids made from cells taken from each individual patient. Organoids are cultures of stem cells capable of differentiating and spontaneously self-organizing into small 3D structures that mimic organs. Heart, lung, and other organoids offer screening platforms for drugs, as well as mechanistic insights. Organs on a chip contain tiny channels lined with living cells and are designed to reflect the architecture and physiology of an organ. This involves capturing the basic elements required for biological activity, including various cell types, structures, and microenvironments, and recreating them in a matrix.

By stringing organs-on-a-chips together in a biologically relevant fashion, researchers can create multi-organ systems or even a human-on-a-chip. Such efforts recently resulted in the first FDA approval of human trials for a drug candidate without preclinical animal efficacy data. The toxicity testing is in a laboratory setting using New Approach Methodologies or NAMs. Implementation of the regimen will begin immediately for investigational new drug applications, where the inclusion of NAMs data is encouraged. This new approach aims to improve drug safety and accelerate the evaluation process while reducing animal experimentation, the cost of research and development, and, subsequently, drug prices.

The responses of patients to therapy are affected by intricate networks of genomic variants, epigenetic modifications, and metabolic pathways³. The analysis of multi-omics addresses this complexity by capturing genomic, transcriptomic, proteomic, and metabolomic data. This provides a systemic view of each patient. AI tools such as deep neural networks, graph neural networks, and representation learning techniques help to detect hidden patterns, fill gaps in incomplete data sets, and enable simulations of treatment responses. This improves predictive accuracy while deepening mechanistic insights. This can reveal how genes interact with each other and the environment through epigenetics to affect therapeutic outcomes.

Also, real-world data from diverse patient populations is broadening the data base. It is important to include many data sets and population-specific algorithms to reduce disparities in healthcare. International teams are improving data harmonization, interpretability, and regulatory oversight. They also improve the synergy between multi-omics integration and AI-driven analytics. This is helping to revolutionize modern precision medicine.

Multi-omics merges diverse data types. This includes DNA variants, RNA expression matrices, methylation profiles, proteomic readouts, metabolite concentrations, and sometimes 16S ribosomal RNA or metagenomic data. Each one has unique noise features and measurement biases. They require specialized bioinformatic pipelines. Data scientists, statisticians, molecular biologists, and clinicians are collaborating to ensure these heterogeneous layers are processed and interpreted properly.

One of the key new analytical methods is the analysis of all the RNA that is produced when DNA is being transcribed in cells. This is called the transcriptome. Transcriptomics measures the expression of RNA and provides information about what human cells are doing at a given moment^4-5. It is very useful when studying disease states, treatment responses, and how different cell populations behave under changing biological conditions. Moreover, advances such as single-cell sequencing, spatial transcriptomics, and multi-omics have expanded the amount of information that one can obtain.

The analysis of RNA sequences has helped us understand patterns of gene expression and understand the transcriptomes of different cells in an organism⁵. The emergence of computational approaches for analyzing RNA sequence data has produced much more accurate data. Several bioinformatics tools and statistical approaches enable transcriptome shotgun sequencing. This helps to identify genes that are associated with various pathologies.

Moreover, AI helps physicians diagnose and treat people as individuals who require personalized and precision treatments. AI models are used to study cancer transcriptomes. They make analyses more efficient while showing how RNA sequences affect cancer subtypes, biomarkers, and tumor heterogeneity. Also, the analysis of dynamic changes in gene expression and pathway enrichment analysis can identify key genes that could be effective drug targets. Moreover, identifying biomarkers can deliver personalized medicine and enable proactive epidemiology.

For example, DUCT-BRCA-CSP is a supervised ML-based web server that predicts the different stages of progression of invasive ductal carcinoma (IDC)⁵. The model was trained on RNA sequence data from The Cancer Genome Atlas, which contains the expression profiles of 610 patients in different stages of IDC. The web server uses two ML classifier models to distinguish IDC's early and late stages. The first classifier uses several methods to train and test the data sets of the model. The other classifier is trained with the same model with data sets that include only IDC biomarker genes. This optimized the prediction of different stages of IDC.

For people trying to build AI systems around biology, generating better data means moving beyond static datasets and toward experiments that capture how human cells and tissues respond in real time ⁵. This helped to inspire the emergence of perturbation atlases, where genes are modified systematically to measure the biological effects. This led to a project focused on generating one of the world’s largest human liver functional genomics datasets. It is built around large-scale gene perturbation experiments in primary human hepatocytes across multiple donors and disease states.

Researchers and companies industrialized a complex transcriptomics workflow without losing biological resolution. However large-scale perturbation datasets are only useful for AI analysis if every sample is processed with consistent quality. Also, technical variation should be minimized, and the data structure must remain compatible with downstream modeling. That combination of RNA expertise, scalable operations, and quality control is making this project successful.

Another important aspect of the project was patient (tissue donor) variability. Rather than treating disease biology as uniform, the dataset incorporated multiple patients and disease conditions to capture how biological responses differ across human populations. The importance of this type of variability is now clearer for AI-driven biology because models trained on homogeneous datasets may fail to generalize across real patient populations.

We are not machines. Modern medicine is predictive, preventive, personalized, and participatory. There is a synergy between AI and laboratory methods to analyze the RNA. Human researchers and health care professionals are learning more every day by studying the active transcription of genes and the effects that it has on each patient’s health.