We all have those annoying little quirks that left unchecked can drive us mad. I always look at objectives. The big data projects have them, clinical trials have them (endpoints), medical education, research articles (null hypothesis), they are every where. My stone in the shoe is how poorly formed and ill-defined they are. Regardless of what we call them, often they are not actionable, measurable, temporal or in the case of clinical trials--relevant. Let me explain.
It seems that precise and well defined objectives carefully placed on a path to robust discoveries may be self-limiting. What if the most innovative discoveries aren't waiting to be predicted in advance? I would argue that maybe we need to identify the steps along the way--not knowing where they might lead. For example, perhaps there are populations of cell types or immuno-signatures potentially missed if we don't redefine what success looks like along the way to a singular objective--improving overall survival for example.
If Deep Learning focuses on programming an artificial neural network (ANN) to learn, Neuroevolution focuses on the origin of the overall architecture. We learn about artificial intelligence by comparing it to how the brain works. Our network connections simulate the neurons of the brain. ANN simulates these connections with stronger connections having more nodes or "weight".
What does it mean to make progress in neuroevolution? In general, it involves recognizing a limitation on the complexity of the ANNs that can evolve and then introducing an approach to overcoming that limitation. For example, the fixed-topology algorithms of the '80s and '90s exhibit one glaring limitation that clearly diverges from nature: the ANNs they evolve can never become larger.
If this area interests you or you are working in machine learning and large datasets read about a new class of algorithms called "Illumination algorithms" a modern approach to neuroevolution. The shift is to focus on a broad cross-section of workable variations of what might be possible instead of looking for a single "optimal" solution.
The main reason for integrating neuroevolution into the discussion has to do with the latest advancements in immuno-oncology. If we think of recent phase III failures--perhaps we are too focused on objectives or clinical study endpoints such as overall survival and progression free survival. Defining a broad cross-section of workable variations might be more successful and informative.
Chimeric antigen receptor (CAR)-T cell infusion has demonstrated an increase of many genes in predefined immune signatures, including t-cell related genes, chemokines, checkpoint inhibitors, and lymphocyte-activation protein 3 (LAG3). Maybe there are too many clinical steps along the path for a singularly-defined objective to serve research goals or more importantly--patients. For example, wouldn't it be informative to address the complexity of cancer as a system? We have multiple measures rolled into a clinical trial endpoint that decides whether a investigational product progresses to the next phase, what patient may benefit, and who funds the research. Meanwhile we must consider pre-existing immunity, type and density of immune cells, spatiotemporal dynamics of intratumoral immune cells, cancer vaccines, the role of cytokines, CD122, T cell bispecifics, and a host of other immune factors. Bringing to Life the Science around Innovative New Drugs, Gene and Cell Therapies--noveltargets.com
A straight path never leads anywhere except to the objective.--Andre Gide Comments are closed.
|
