Well today’s post has been a few years in gestation and yes I have had to keep it quiet for a long, long time which was extended as our recent paper was embargoed for what seemed like an eternity. It starts back in 2009 when I was using some of the NIAID/SRI whole cell screening data for mycobacterium tuberculosis (here and here). This lead to building machine learning models to predict whole cell activity. One of the side effects of this were tables of fragments that were important or not in compounds used in the model. Little did I know that years later they would be used to design some new molecules and help find new antibiotics. I guess you could say this is an example of the unintended consequences of research. Over the years, predominantly with my collaborator Joel Freundlich (Rutgers) we have generated more machine learning models for TB (here and here). Then a few years back along with Gyanu Lamichhane at Johns Hopkins University we embarked on an R21 aimed at developing oral carbapenems for drug resistant TB. Joel wanted to make use of the data from the earlier machine learning models to design evolved carbapenems that would have properties that would bias them towards having activity against Mtb. Our new paper published today in Nature Chemical Biology is an extensive study on the mechanism of carbapenems which inhibit L,D-transpeptidases of Mtb as well as ESKAPE pathogens. The new evolved carbapenems made by Joel’s lab and tested by Gyanu’s lab, along with crystal structures, show how these compounds can bind covalently in different orientations in the enzyme binding site . This study also showed that biapenem was more potent against Mtb and was active in the mouse model (especially when combined with Rifampicin). This work represents an important step towards showing carbapenems are feasible for treating Mtb. It also highlights how some machine learning efforts can ultimately impact design of drugs, albeit not in the manner originally intended, but nonetheless creatively! It represents another collaboration that has been incredibly productive and goes some considerable way to showing what is possible when combining computational and experimental research for Mtb. I am eager to see where this work goes as Gyanu and Joel push this work onwards!