Designing new Tuberculosis drugs using AI

 Tuberculosis or TB is the world’s deadliest infectious disease, sickening over 10 million people each year. A cocktail of four antibiotics, taken for six months, is the clinical standard of care for TB.

 The spread of drug resistant TB has created a pressing need for new treatments. A major impediment in finding these treatments is the large number of drug combinations, which makes experimental testing infeasible.

Antibiotics are used in combinations to control the rise of drug resistant strains and increase potency. The 28 drugs currently used to treat TB could be assembled into nearly 24,000 different 4-drug combinations. TB therapy design would be revolutionized by approaches that can rapidly evaluate the efficacy of antibiotic combinations.

 To address this problem, we have developed a computational model (INDIGO-MTB) that significantly outperforms traditional methods in terms of both speed and throughput. The largest studies have so far analyzed up to 200 TB drug combinations. In this study, we have predicted outcomes for a million combinations with high accuracy, which enabled us to identify new synergistic drug combinations. INDIGO also suggested ways to improve the synergy of existing drug treatments, which we confirmed through experimental testing, in collaboration with the Sherman lab at UW.

 We then went back and analyzed historic clinical trials of TB drug combinations. We asked if knowing the synergy of drugs can shed light on the ultimate clinical efficacy of treatments. Note that there are many other factors that may impact clinical outcome, including side effects, drug dosing, or presence of other co-infections. Nevertheless, strikingly, we found a strong correlation between the extent of synergy of a drug combination predicted by INDIGO and its corresponding clinical efficacy.

This suggests that INDIGO is a promising approach to rapidly identify new drug combinations for TB and other antibiotic resistant infections. We are now planning to test promising TB drug combinations in mouse models.

 Ref: Ma S, Jaipalli J, Larkins-Ford J, Lohmiller J, Aldridge B, Sherman D, and Chandrasekaran S, Transcriptomic signatures predict regulators of drug synergy and clinical regimen efficacy against Tuberculosis, mBio, 2019

Scott Campit, Sriram Chandrasekaran

Each cell in our body makes tiny machines called proteins that carry out essential tasks. The instructions to make these essential molecules are stored in DNA. While the genetic code itself is vital for life, proteins work together with DNA to maintain health. One example is histones, the proteins that wrap around DNA. By tightly wrapping around a region of DNA, histone proteins control the activity of genes encoded in the DNA segment. Attaching or removing specific metabolites onto histones acts like a switch, turning genes ON or OFF. The collection of different chemical modifications on histones is called the ‘histone code’. Several diseases including Alzheimer's disease, hypertension, and cancer are predicted to occur due to alterations in the composition of the histone code.

To discover new therapies for these diseases, our lab is developing computational models of how genes, histones, and metabolites interact together in a cell. In this study, we focused on how nutrition and metabolism of a cell affects the levels of a key histone modification called acetylation. This modification results in turning ON of genes in the DNA associated with the modified histone. Using our computer model, we simulated how starving or feeding cells with glucose and other nutrients impacts histone acetylation. Our model demonstrated that the excess levels of a key metabolite -  acetyl-coA is predictive of increase in histone acetylation levels. This observation can explain how in metabolic disorders or in tumors the change in metabolism can affect gene activity.

Next, we wanted to see if we could use these computer models to identify new cancer therapies. Several drugs that are used to treat cancer kill tumors by changing histone acetylation levels. Yet, it is difficult to determine how effective these drugs for different cancer types. Our computational model accurately predicted cancer cell types that are sensitive to these drugs based on their metabolic activity. This approach will ultimately allow researchers and clinicians to identify new cancer drugs that inhibit specific types of tumors using computer models.

Reference: Shen, F., Boccuto, L., Pauly, R., Srikanth, S. and Chandrasekaran, S., 2019. Genome-scale network model of metabolism and histone acetylation reveals metabolic dependencies of histone deacetylase inhibitors. Genome biology, 20(1), p.49.