Who sees the patient, and why

If the patient is ‘at the centre of everything we do’, the lens that most Discovery applies is ‘yes, and their cellular receptors are at the centre of the patient…’

The question of who ‘sees’ the patient, and when, remains pertinent. Humans who become patients remain complex, not just in their biology, but their psychology, behaviour and more. Developing drugs in silico, therefore, requires a model which is ‘if we take care of the disease, the patient will take care of itself’, or it requires the patient to be modelled into their disease. That, today, would blow the circuits of any machine - quantum computing, computational biology, etc., all are better than last year’s tools, but still not enough for that task.

Theoretically, the pattern recognition of AI/ Machine Learning should find pockets of homogeneity in this complexity, in which case the training set becomes critical. However, like self-driving car development, there will come a point when the complexity of the real world is not just a mathematically bigger version of a simulation, but an intractable problem. The?‘chicken chased into the road’ problem ?is not an exception, but emblematic of wholly expected problems. To think that it is something we can ignore is to see the simulation as ideal, and the real world as annoyingly problematic - analogous perhaps to the sterile clinical trial environment versus the real world. If most people with depression can’t be put into most of our studies of depression because they’re too complicated, that is a model problem.

Disease models improve, but none are ‘done’. This problem is important - there is no view of something as ‘simple’ as Type II diabetes, for example, which is complete (at time of writing - if you know different, just let me know…). There are hundreds of pathways, even with our current understanding, with interdependencies and hierarchies, and contingencies and redundancies. Tomorrow, it will be different. As Brian Christian’s remarkable?The Alignment Problem ?notes, on the assumption that the real world looks like the training set:

One of the simplest violations of this assumption, however, is the world’s stubborn and persistent tendency to change. I recall hearing one computational linguistics researcher complaining that no matter how hard they tried, they could not get their model to replicate the accuracy of another researcher’s published result from a year or two earlier. Over and over they checked their work. What were they doing differently? Nothing, as it turned out. The training data were from 2016. The English being written and spoken in 2017 was slightly, but measurably, different. The English of 2018 was more different still. This is one example of what researchers know as “distributional shift.” No one trying to reproduce the paper’s results ever would reach the same level of accuracy as the original researchers, at least not with that training data. The model trained on 2016 data was slowly bleeding out its accuracy as the world moved on.

We have to guard against this problem in learning. If the English language is complicated, but finite, and it changes year on year, the complexity that’s driving it is English language?speakers, and their environment, fashions and preferences, etc. Analogous for us is Type II diabetes and the patients that have it. If we don’t have a ‘map’ of Type II diabetes, we certainly don’t have a map of Type II diabetes patients, however many ‘patient journeys’ we choose to collect.

Brian Christian : Taken together, we have a litany of reminders that “the map is not the territory.” As Bruno Latour writes, “We have taken science for realist painting, imagining that it made an exact copy of the world.”

We are in danger of losing control of the world not to AI or to machines as such but to models. To formal, often numerical specifications for what exists and for what we want.

Rather than dwell on that irreducible complexity here, let’s focus then on who wants to understand the patient. We may proclaim that patients drive what we do, but it’s a paternalistic view that says we know what will be good for them. We might drag some patients in after we’ve already decided what they are going to have. Patients’ data might well have been borrowed for the purposes of Discovery, but they’ll be called ‘subjects’ in clinical development - measured and tracked, but probably not listened to as?people. They may well be called in as advocates once confidence is higher in getting towards phase III or approval - in some more enlightened companies, perhaps they will be asked to input to the phase III plan. But the patient at that point will be rather fixed, and look a lot like the homogeneous one that the Clinical Development Plan picked for us.

In many companies now, we have?thought about who gets to talk to patients , but we haven’t really laid out?who gets to listen to them . One of the challenges with our faith that better machines will help our process is that it further distances ‘the science’ from ‘the patient’. The molecule is not the drug - the ‘golden age’ of psychiatric drug development was typically observational: understanding the real world situation led to great insights. We might pretend that we’re better than that now, but surely it’s an ‘and’ rather than an ‘or’… As Safi Bahcall’s Loonshots notes, insights that led to Avastin were observational. We know that anthropology led to PCSK9s, that Opdivo’s early win in the IO battle was human, and that the expertise of investigators in SLE makes a difference to trial success in phase II and phase III…?

Seeing real people is something babies can do, but we still have to teach computers. Wanting to listen to them means deciding where in our process we start, and who does the listening. As asymmetries go, better listening is one of the most powerful.