Major breakthrough! AI identifies drug candidates for more than 17,000 diseases. Are orphan drugs for rare diseases no longer a distant dream?

“For patients with rare diseases, having access to medicines is always the top priority.”

Bi Jingquan, chairman of the China Center for International Economic Exchanges, said this in 2022.

Rare diseases, as the name implies, are a type of disease with a relatively low incidence rate . Due to the small number of patients, small market demand, and the difficulty, high cost, and long cycle of drug development, the difficulty of treating rare diseases remains high, and drugs for rare diseases are even called "orphan drugs." The vast majority of patients with rare diseases face the problems of difficult treatment and drug shortages.

According to the World Health Organization (WHO), among the more than 7,000 rare diseases found worldwide, less than 10% have approved treatment plans or drugs, and most of them require lifelong drug treatment. In this severe context, how to find effective treatment drugs for rare diseases is a key step in alleviating the treatment difficulties of patients with rare diseases.

This step is expected to be a "small step" forward with the help of artificial intelligence (AI).

Recently, a research team from Harvard Medical School and its collaborators developed an AI model called TxGNN, which is the first new method developed specifically for identifying candidate drugs for rare diseases and untreatable conditions. It has identified candidate drugs for more than 17,000 diseases from existing drugs, many of which have no existing treatments. Compared with similar AI models used for drug repurposing, TxGNN has an average improvement of nearly 50% in identifying candidate drugs and a 35% higher accuracy in predicting which drugs will have contraindications.

The related research paper, titled “A foundation model for clinician-centered drug repurposing”, has been published in Nature Medicine, a subsidiary of Nature.

How can AI promote drug repurposing?

Although the traditional drug repurposing strategy can accelerate the clinical application of new drugs by utilizing the safety and efficacy data of existing drugs, this method is often accidental and opportunistic, and it is difficult to systematically solve the drug development problems for rare diseases.

In this context, TxGNN came into being, bringing revolutionary breakthroughs in drug repurposing, especially showing great potential in drug discovery for rare diseases.

TxGNN is a basic model based on graph neural network (GNN) specifically for zero-shot drug repurposing. Unlike traditional methods, TxGNN is not limited to the known relationship between existing drugs and diseases, but embeds the complex relationship between diseases and drugs into a potential representation space by training a medical knowledge graph (KG), so that potential therapeutic drugs can be predicted for any given disease. The knowledge graph contains medical concepts such as 17,080 diseases, 7,957 drugs, and 27,671 proteins, providing a rich data foundation for the training of TxGNN.

Figure | TxGNN is a graph-based model for drug repurposing that can identify candidate drugs for diseases with limited treatment options and limited molecular data. (Source: The paper)

In multiple experiments, TxGNN demonstrated its powerful predictive ability. Compared with 8 existing methods, TxGNN showed significant advantages in zero-sample environment. According to the experimental data in the paper, TxGNN improved the accuracy of drug indication prediction by 19% and the accuracy of side effect prediction by 23.9%.

These results show that TxGNN can not only find potential new uses for existing treatment options, but also accurately predict possible drugs when there are no known treatment options. This is of great significance for the treatment of rare diseases, because more than 95% of rare diseases have no existing treatment drugs, and TxGNN creates the possibility of rapid discovery of new drugs for these diseases.

Figure | TxGNN can accurately predict drug indications and contraindications. (Source: This paper)

In addition to its predictive capabilities, TxGNN also has a specially designed explanation module to help doctors and researchers understand the model’s predictive logic. This module shows the potential connection between drugs and diseases through multi-hop paths.

The explanation module not only points out why a drug may be effective for a specific disease, but also provides a detailed medical knowledge path, allowing users to trace the scientific basis behind the prediction. With this feature, TxGNN overcomes the "black box problem" of many AI models in medical applications and greatly improves the interpretability and trustworthiness of the model.

Figure | Development, visualization, and evaluation of multi-hop explainable paths in TxGNN Explainer. (Source: the paper)

The research team validated the practical application of TxGNN in rare disease drug discovery and achieved encouraging results. In the experiment, many predictions of TxGNN were highly consistent with actual clinical over-the-counter drug use.

For example, in an analysis of electronic medical record data from 1,272,085 patients, researchers found that the top drugs predicted by TxGNN were used significantly more frequently in these patient groups than random predictions. Analysis of log(OR) (co-occurrence ratio) showed that the association between the use of drugs predicted by TxGNN and rare diseases was 107% higher than the underlying drug predictions, further demonstrating the effectiveness of the model in actual clinical settings.

Figure｜Evaluating the prediction results of TxGNN in a large medical system. (Source: This paper)

A specific example is that TxGNN recommended deferasirox as the most promising candidate drug when predicting potential therapeutic drugs for Wilson disease, a rare disease that causes abnormal copper metabolism. This drug has been used in the clinic to treat iron overload diseases, and TxGNN showed through its explanation module that deferasirox may have a positive effect on the treatment of Wilson disease through metabolic pathways. This prediction is also consistent with relevant research results in the medical literature, showing the scientific rationality of the model.

In the current context of long and costly drug development cycles, TxGNN provides a systematic solution for the repurposing of existing drugs.

In the future, as this technology continues to improve, it is expected to become a key driving force for accelerating drug development, especially making drugs available to patients with rare diseases.

AI has great potential in medicine

In addition to the field of rare diseases, AI is also widely used in other medical fields.

In September 2019, AI drug development company Insilico Medicine developed a generative AI platform called GENTRL, which has been used to design new drugs for the treatment of diseases such as fibrosis and cancer. By using GENTRL, researchers discovered a potent inhibitor of the discoidin domain receptor 1 (DDR1) in just 21 days, greatly shortening the drug discovery time.

In 2022, researchers at the Institute of Cancer Research in London created a prototype test that can predict within 24-48 hours which drug combinations might be effective for cancer patients. They used AI to analyze large-scale data from tumor samples to predict how patients would respond to drugs more accurately than current methods.

In 2023, US company Nuance Communications launched a new AI-driven voice-activated software application, Dragon Ambient Experience (DAX) Express. The program aims to effectively help reduce the administrative burden on clinicians by leveraging natural language processing technology to facilitate real-time transcription of medical records during patient consultations.

In 2024, a research team from Harvard Medical School and its collaborators proposed the Clinical Histopathology Imaging Evaluation Foundation (CHIEF) model for extracting pathological imaging features for systematic cancer assessment. On 15 datasets containing 11 cancer types, CHIEF achieved an accuracy rate of nearly 94% in cancer detection, significantly outperforming previous AI methods.

There is no doubt that AI has the potential to improve rare disease detection, drug repurposing, and clinical process optimization, which will better help us fight diseases.

Author: Ruan Wenyun