I wanted to follow up on something that came up much earlier in the coronavirus pandemic. Back in February, a group at BenevolentAI proposed the kinase inhibitor baricitinib as a possible therapeutic for the coronavirus. They identified this through their company’s machine-learning approach to the medical literature and disease mechanisms, and identified the compound due to its proposed effects on endocytosis. The drug is used in arthritis therapy as a Janus kinase inhibitor, but it’s also known to inhibit adapter-associated kinase 1 (AAK1), which is where the endocytosis comes in (and that’s certainly a candidate for being able to affect viral entry).
Targeting this enzyme and related ones in the pathway the numb-associated kinase (NAK) family, had been suggested several times over the years as a possible antiviral therapy (the earliest paper I know of is from 2007). So it seems that this idea was going to come up one way or another, AI or not. Some of the coverage at the time was a bit breathless, as is often the case with AI stories that hit the popular press. I had some comments on this at the time, mostly to the effect that this was a good example of literature searching and curation of a useful database, and that there’s nothing wrong with that. If software can help us do that, so much the better – the literature is a gigantic shaggy mound, and we need whatever help we can get in extracting actionable things from it.
Since then, baricitinib and other JAK inhibitors have been tried out in the clinic. In August, a paper from Lilly, BenevolentAI, and other collaborators provided more details. Baricitinib did indeed seem to be effective in cellular models, and a case series of patients treated with it showed some promise. The FDA issued an Emergency Use Authorization for the combination of baricitinib and remdesivir, and now we have the data that led to that decision, published in the NEJM.
At the end of this process, what you see is that the patients getting both drugs recovered a median of one day faster than the ones getting remdesivir alone. Differences in mortality between the two groups showed a trend towards improvement in the dual-treatment group. There was evidence that the combination led to less use of oxygen and mechanical ventilation, and all of these differences seemed to be more pronounced in patients who were in more serious condition at the start of the study. The combination actually produced fewer adverse events than remdesivir alone. These numbers sound pretty similar to dexamethasone, but the paper notes that this trial and the RECOVERY one had different designs and can’t be directly compared. You’d have to run a head-to-head between a remdesivir/baricitinib group and a remdesivir/dexamethasone group to sort that one out, and the good news is that that trial is going to happen.
What we don’t know is whether baricitinib works via the mechanism proposed by the original BenevolentAI paper. Remember, it was first identified for the AAK1 activity; this was before we knew so much about damping down cytokine activity as a needed therapy in the later phases of coronavirus therapy in some patients. JAK inhibitors already do that, which is why they’re used in arthritis treatment, and were independently proposed for coronavirus therapy for that reason. But it’s worth noting that another JAK inhibitor (ruxolitinib) just failed to improve the recovery of coronavirus patients in a separate trial. It does seem that of the JAK inhibitors (which all should affect cytokine levels) that baricitinib has the best additional activity on the endocytosis-related kinase targets, so the original proposal is definitely still alive.
We’ll see from the dexamethasone comparison trial, though, how useful it is under real world patient care conditions (especially when compared to such an inexpensive drug as dexamethasone). That’s something you can’t predict with any AI in the world – not yet, anyway, and it’s going to be a long time before that’s feasible. We’ll take what we can get.