NITMB Industry Exchange

with Risi Kondor

RNA function depends on dynamic folding, which standard models often miss. Graph neural networks encode these folding rules directly, predicting riboswitch function both accurately and interpretably with far fewer parameters. Building on this framework, P-tensors and diffusion models are combined to design new functional RNA sequences, which are then validated experimentally.

Genes burst, taxa bloom and vanish, and tumors alternate between dormancy and growth—sparse, bursty dynamics that Gaussian tools handle poorly. This project develops practical probabilistic models for these dynamics, with applications in predicting tumor response to unseen treatments.

with Jake Soloff

Model selection is often unstable: dropping a single data point can flip which model gets chosen, threatening reproducibility. This project’s method returns a small set of near-best models rather than a single fragile pick, with a formal stability guarantee, demonstrated on Lotka–Volterra dynamics.

with Seppe Kuehn

When does a microbial community stay stable, and when does it adapt? Combining consumer-resource modeling with dynamical systems, this project shows how traits like auxotrophy and sequential nutrient use shape resilience and coexistence in fluctuating, “boom-and-bust” environments, with testable predictions.

with Vincenzo Vitelli & Madhav Mani

Sequencing tells us which species are present in a microbial community, but not which functional groups drive its behavior—information critical for engineering probiotics, fermentation cultures, and soil products. A neural network addresses this gap by grouping microbes by function rather than genetic relatedness, while quantifying uncertainty. At a larger scale, work on nitrogen cycling in farmland soil links shifts between growth, stationary, and death phases to environmental conditions such as pH and, ultimately, soil carbon release. More broadly, these approaches show how modeling collective microbial function can enable more predictable and controllable bioprocesses across probiotics, biomanufacturing, and agtech.

with Yogesh Goyal & Ben Kuznets-Speck

Predicting how high-dimensional cell states respond to perturbations is central to biology. Using statistical physics, CIPHER predicts a cell’s genome-wide responses to genetic perturbations from baseline gene covariance alone (validated across 1M+ cells). Its companion, Keeping SCORE, turns diffusion models into interpretable, calibrated classifiers. Together they enable causal inference and mechanism discovery.

with Ilya Nemenman

Glucose regulation is a control problem spread across physically separated organs. A minimal multi-organ model shows why purely local sensing runs into unavoidable tradeoffs, and how hormones like insulin overcome them. This model is relevant to diabetes and closed-loop insulin-delivery devices.

with Siwei Wang & Alexander Pearson 

Immunotherapy response depends on how immune cells are spatially arranged around tumors—something visible in routine histopathology, but difficult to quantify. This project builds spatial graphs that capture both the relationship between cells and their shape, generating interpretable biomarkers of treatment response that could help spare non-responders from unnecessary toxicity.