How do you go from what you want a cell to do, to the DNA that makes it happen? This is a central challenge in synthetic biology - and the inverse problem at the heart of my current work. I lead the computational efforts on this project in the Weiss Lab at MIT, where we're building what we call a "biocompiler."

The biocompiler is a machine learning framework that learns to predict cell behavior from genetic circuits, then runs in reverse to design new circuits for desired functions. At its core are Biomorphic Neural Networks (BNNs) - compositions of neural functions that map directly to cellular processes like transcription, translation, and RNA cleavage. We train on experimental data, learning embeddings for DNA parts that capture their effect on cell behavior.

Once trained, we freeze the biological "constants" and optimize in the other direction: given a target function - a bandpass filter, an analog sensor, and eventually, programmed multicellular behaviors - we use gradient descent to find the DNA that produces it. Unlike traditional approaches built on boolean logic, we embrace the fact that cells are fundamentally analog: continuous, noisy, and richly expressive. The output is a circuit design ready for the lab.

This is what I mean by building translators for the language of life.

As a Postdoctoral Associate in the Marks Lab at Harvard Medical School, I focused on intrinsically disordered proteins (IDPs) - a family whose resistance to traditional investigation has earned it the name "Dark Proteome." IDPs are involved in nearly every major cell process, and their association with prion-like mechanisms and amyloidosis makes them key players in neurodegenerative diseases.

I led the computational side of the Biostasis project, which aims to identify protein sequences responsible for cryptobiosis - the ability of organisms like Tardigrades to "hibernate" and later be resuscitated. I developed new ways to classify IDP sequences, including a spectral representation that highlights periodicity patterns relevant to their interactions.

• Scalable nonparametric Bayesian models that predict and generate genome sequences.

  Neurips Workshop, 2021
• + journal article in preparation

This was my first and main research project as a Postdoctoral Associate at MIT. I was working under the direct supervision of Prof. Neri Oxman and in close collaboration with Profs. Christopher Voigt, Cullen Buie, and Timothy Lu. The ambitious goal of this project was to program and simulate the growth of a multicellular aggregate, compile the virtual genotype of the simulated organisms into real DNA, and subsequently implant this into real cells. The model organism of interest was Escherichia coli.

We want to be able to computationally design a phenotype - a desired pattern - and have the software output the corresponding DNA code (genotype). This DNA will then reproduce the same growth process and the corresponding patterns when inserted into a living cell.

I developed a specialized agent-based simulation engine, a logic-gate based virtual gene regulatory network and a machine learning framework that heavily relies on evolutionary algorithms. The machine learning part is particularly fascinating, as it applies to both the calibration of the simulation (trying to reduce the reality gap) and the inverse problem (finding the correct genotype for a desired phenotype).

• Viva in Silico: A position-based dynamics model for microcolony morphology simulation.

  Artificial Life Conference, 2018
• + journal article in preparation

Working with emergence - rather than against it - is possible at many levels and scales: from molecular to architectural, from gene networks to insect swarms. With Silk Pavilion II, the Mediated Matter group explored this at the architectural scale, working with silkworms as biological co-designers.

The pavilion - a 6-meter tall structure exhibited at MoMA's Material Ecology exhibition in 2020 - was spun by over 17,000 silkworms onto a carefully designed dissolvable scaffold. The scaffold's geometry was crafted to let the worms spin freely without triggering their instinct to enclose themselves in a cocoon - allowing them to simply deposit silk and live out their full life cycle.

• Silk Pavilion II on oxman.com

Since a considerable amount of my research required state of the art cell physics simulation, I developed a specialized cell physics engine. MecaCell is an open-source platform for the agent-based simulation of cells, with a strong focus on modularity and performance. Depending on the core physics model used, it is able to handle several millions of cells on a conventional CPU. It is a lightweight platform designed with artificial life and morphogenetic engineering in mind. It has been in use throughout most of my PhD and postdocs.

• MecaCell: an Open-source Efficient Cellular Physics Engine

  Artificial Life Conference, 2015

• Viva in Silico: A position-based dynamics model for microcolony morphology simulation.

  Artificial Life Conference, 2018

• Learning Aquatic Locomotion with Animats

  ECAL, 2017
• MecaCell on github

I have used and researched better representation of aGRNs both as cell controllers (for cell cycle regulation) and as controllers for more abstract problems (for example, in this video, a GRN was evolved to control a ship).

This research gave insights on the dynamics of these complex systems and allowed for interesting improvements and variations in their design, encoding, and evolvability.

• A comparison of Genetic Regulatory Network Dynamics and Encoding

  GECCO, 2017

• Dangerousness Metric for Gene Regulated Car Driving

  Evostar, 2016

• Artificial Gene Regulatory Networks for Agent Control

  Evolutionary Computation in Gene Regulatory Network Research

• Investigating Artificial Cells’ Spatial Proliferation with a Gene Regulatory Network

  Procedia computer science, 2017

My research led me to use a wide variety of evolutionary computations & genetic algorithms that, in some cases, required a degree of customization and versatility that I didn't find in other libraries. I put everything together in a modern C++ project.

• GAGA on github

This is a soft robotics research project I worked on while a visiting PhD student at EPFL in Pr. Dario Floreano's team. By letting each deformable cell unit define its adhesion pattern, unconnected cells can, through random motion, assemble into desired geometric structures, such as the formation of snowflakes patterns.

Starting from a single cell with identical local rules governing communication, growth, and specialization, a varied panel of morphologies can develop. These environmental rules and cell controllers can be discovered and explored through various evolutionary computation tools. In much of this research, virtual creatures are composed of cells housing an artificial Gene Regulatory Network (GRN). The GRN is evolved in order to favor the organisms that would best fit their environment and its energy constraints.

One of my main interests in this area is studying how very simple fitness functions make for the emergence of complex phenotypes. For example, in a series of experiments that I did during my PhD, the organisms only had to "survive"; the complex phenotypes and life-like behaviors emerge from the simulation constraints (mostly related to energy). Here, an example of growth and auto-organisation into shells (nutritive cells feeding a shield of very resistant cells), and another example showing the emergence of roots and leaves (to produce energy from simulated sunlight and nutrients buried in the ground).

• Self-organization of symbiotic multicellular structures

  Artificial Life Conference, 2014

• Evolved developmental strategies of artificial multicellular organisms

  Artificial Life Conference, 2016

This is more of a hobby than a "true" research project, but, to me, being a part of such a great robotics team was an invaluable opportunity to apply and improve my electronics skills, as well as some computer vision and planning algorithms.

I have designed, from scratch, the electronics of the two robots of our team that had to compete in the French Robotics Cup. I also wrote some low-level code and helped with computer vision algorithms. Overall, this was an incredible experience, especially since we have had excellent results (winners of the 2021 edition, Innovation Price in 2016, First Runner-Up in 2017 and 2019, 3rd at the European Robotics Cup of 2017).

My cell simulation software has been used to model tumor spheroids and study the formation of a necrotic core due to various constraints (oxygen & nutrient depletion and mechanical constraints mostly). Combining this software with evolutionary computations, I have also helped with the simulation of the evolutionary dynamics of the "arms race" between the immune system and melanoma.