Events

The Gerstein lab conducts computational biology & bioinformatics research in the biomedical and genomic fields. We use various computational analytics methods including artificial-intelligence / machine-learning techniques to analyze large biomedical datasets and develop bioinformatics tools. The lab has particular focuses on the following areas of research: neurogenomics, personal genomes, genomic privacy and genome annotation.

https://medicine.yale.edu/profile/joel-rozowsky

Lab: https://www.gersteinlab.org/

Microbial communities in the gut, soil, and ocean hold thousands of species, yet we still lack a simple theory for which species thrive where and what the community does. In the macroecology of plants and animals, measurable traits like leaf thickness or beak shape are found to predict species abundances. But in microbiomes, the physiological traits that govern growth cannot be measured organism by organism inside a living community. We instead plan to infer them directly from data. Using the minimax-entropy principle,  a community’s composition is the least-biased distribution consistent with a few community-averaged traits, and the optimal traits are the ones that minimize the residual entropy,  we reverse-engineer the traits of a microbiome directly from data. This compresses a community of thousands of taxa into a handful of interpretable low dimensional axes that predict composition, read out the environment, and forecast function, and we test the inferred traits against real genomes and metabolism to confirm they are genuine biology.

https://sites.google.com/view/dixitlab/home

This project will evaluate the Tenstorrent TT-QuietBox 2 as a platform for running and adapting local AI models for privacy-sensitive tasks. We will benchmark open-weight language and multimodal models on document analysis, information extraction, question answering, summarization, and code generation, measuring accuracy, latency, throughput, memory use, energy consumption, and multi-device scalability. We will also develop smaller, task-specialized models using approaches that may interact with larger frontier models. The project will produce reproducible benchmarks, deployment tools, and practical guidance for determining when local AI can complement or replace cloud-based systems while reducing external exposure of sensitive data.

https://quanquancliu.com

Human economies run on contracts. As LLM agents increasingly transact on our behalf — and with each other — the same need arises at machine speed. The technology is essentially ready: a markdown contract, interpreted by an LLM judge, settled on-chain. The economics is not: under what conditions do agents actually adopt contracting, rather than merely being capable of it?

We probe this with a simulated economy of LLM agents facing problems too large to solve alone, able to talk publicly, huddle privately, and sign contracts. Across 45+ runs, three obstacles block a working contract economy. (1) Information: with incomplete records of others’ actions (history) or one’s own (memory), contracts fail to emerge. (2) Protocols: when interaction sequences deviate from human economies, contracts that work among humans become ineffective — and agents stay blind to this even when instructed. (3) Objectives: agents contract for fairness over the profit maximization they were instructed to pursue — likely priors absorbed from human training data.

The next phase has two pillars: moving from an export economy of problem-solvers to agent value chains modeled on real markets now taking off (agentic commerce, pay-per-crawl content markets), and developing solutions to the three obstacles through information design and the training pipeline.

Requirements: Strong interest in the agentic economy and multi-agent LLM systems; comfort building and running LLM-agent simulations; being a Ph.D. student would be preferred but not necessary.

https://sites.google.com/view/soheil-ghili/home

https://kekexinz.github.io

This project tackles two related forecasting problems in the electricity sector. Both share a common core challenge: making reliable predictions from limited, sparse, or anomalous data. The first thread involves day-ahead market clearing price forecasting for European electricity markets (e.g. Greek Market). These markets often have a short historical record (~6 years). Part of that record is also distorted by structural anomalies tied to major regional political or energy crises. We have tested existing ML and deep learning models, including LASSO, PCA, RF, LGBM, MLP, and transformer-based models like TOTO and Chronos. These models perform on par with published benchmarks. However, they noticeably underperform relative to results reported for larger, more stable markets (e.g. NordPool). This gap is likely a direct consequence of the short and anomalous training window. The second thread involves EV charging demand prediction, using proprietary station-level data from an industry partner. Aggregate hourly demand can be forecasted accurately. However, station-level forecasts currently perform little better than random guessing. This is because many stations see intermittent usage, concentrated in just a few scattered hours across sparse active days.

We’re looking for students with a strong statistics/math background to bring rigorous statistical thinking to both problems. For price forecasting, this means characterizing the structural break in the data. It also means building interpretability and uncertainty analysis, to help explain the performance gap versus peer markets. For EV demand, this means applying statistical methods suited to sparse, intermittent series. It also means developing a smarter station clustering approach. Our naive geographic clustering did not work, so the new approach should incorporate road type, urban/rural context, and altitude. Both threads offer a strong opportunity to apply statistical modeling to real, messy, industry data, with direct impact on power-sector decision-making.

https://smartnets.yale.edu

Every day, patients in the ICU are kept sedated and on the mechanical ventilator longer than necessary – not because it’s medically required, but because clinicians lack a fast, reliable way to synthesize the dozens of data points (vital signs, medications, ventilator settings, labs) needed to determine readiness for “wake-up” trials. The result: unnecessary sedation, prolonged ventilation, and higher rates of delirium and long-term functional and cognitive decline leading to loss of independence. This project needs a computer science collaborator to build and validate an algorithm that processes structured, multi-domain EHR data – hourly, in near real-time – to flag when a patient meets clinical criteria for sedation and ventilator weaning trials. The core challenge is a time-series classification problem: integrating heterogeneous streams (hemodynamics, respiratory parameters, neuro status, medication records) across, then validating algorithm output against clinician-documented gold-standard labels.  This is a rare chance to design an algorithm with a clear path to clinical deployment – feeding directly into a decision-support tool that will be piloted across multiple health systems once validated.

https://medicine.yale.edu/profile/snigdha-jain/

One of the most natural learning abilities of humans and other animals is the ability to accumulate skills from past experience and compose them to accomplish more complex tasks. While reinforcement learning in artificial intelligence has made progress in decomposing complex behavior into reusable skills in a top-down fashion, it remains challenging to learn such skills, bottom-up, incrementally from a stream of experience, retain them, and flexibly compose them to solve increasingly complex tasks. Given the difficulty of hand-designing algorithms for this kind of continual compositional learning, this project aims to design a framework to metalearn them in a data-driven fashion.

https://www.cs.yale.edu/homes/irie/