My industrial research work at Adagene is, for now, not published and under non-disclosure agreement. I work as computer scientist, designing algorithm, models and implementing them. I never get tired of coding.
The bulk of my effort currently goes into machine learning for protein structures. Large public databases of protein structures are available, the RCSB Protein Data Base for instance. We can learn very, very useful things from those databases. At Adagene, we want to learn how to build synthetic antigenes that will be able to target very precily a set of targets in a human organism. Indeed, we are not interested in one or two targets, but any target.
To accomplish this, I am developping predictive models that link the protein shape, its shape and its physical properties. It involves a lot of clustering, interesting geometry and statistics problems and some engineering too. Those dataset do not quite fit within the RAM of a computer, and you have to distribute the computations over several computing units.
I also do some combinatorial optimization, where the problem is to find a set of DNA maximizing an objective while satisfying a set of local and global constraints. The difficulty lies on finding answers within a short time and without using too much ressources.
- Clustering for protein structures
- Predictive models for protein structures
- Geometrical alignement
- Large scale clustering
- Out-of-core clustering
- Directional statistics
- Combinatorial optimization
My academic research work focuses on exploring ways to delegate conception and design of objects or mechanisms to computers, by using optimization algorithms. The objects to conceive can be robots, planes, turbines, neural networks...
An other angle of my research work is swarm : large communities of agents without centralized control. Those agents, by interacting with their close neighboring environment, will eventually perform a task all together.
So my research combines those two aspects: optimization of objects, where what is optimized is not an object, but the capacity of a swarm to build objects fit to a task. The interesting thing in such an approach, is that the optimization task does not depends on the scale of the objects we want to design, nor the size of the swarm we are going to use to generate an object. Potentially, this would allow the design of very large designs with reasonable computational resources. Also, decentralized construction fits nicely with emerging technologies like 3d printing and meta-materials.
I tinker with neural networks, evolution strategies, physicals simulations and computational geometry to do all that. My work have strong connections spatial computing and self organization. I tend to be very empiric, mostly in-silico experiments.
- Morphogenesis of buildable objects
- Artificial embryogeny
- Evolutionary algorithms
- Swarm intelligence