This abstract has been withheld from publication.

Mapping context-dependent signaling networks is essential for decoding cellular responses and identifying therapeutic targets. However, reconstructing these networks from data remains challenging; manual curation lacks scalability, while mathematical models struggle to integrate vast, unstructured biological context. We address this by employing Large Language Models (LLMs) as "inverse problem solvers" that formalize latent biological knowledge into logical constraints. We applied this approach to CRISPR knockout screening data for reversing cellular senescence, tasking an LLM to integrate functional hits into a predictive signaling network. Preliminary results reveal a "logic-truth gap": the LLM generates logically consistent conclusions that frequently diverge from empirical ground truth. This suggests the primary bottleneck is structural network fidelity rather than the LLM’s internal reasoning. We discuss strategies to bridge this gap, providing a roadmap for integrating generative AI into automated pathway discovery.

Python is the dominant language in data science, with a recent Jetbrains survey finding that 51% of Python developers are involved in data analysis. Unlike other forms of software engineering, data scientists often develop in an exploratory manner, using insights from previous runs to guide their analysis. However, traditional scripting runtimes do not persist state across executions, requiring repeated recomputation of unchanged results. Computational notebooks are a widely adopted solution, enabling users to only re-execute individual cells in a persistent session. While convenient, if the user ​​modifies cells without executing their dependents, the retained state diverges from the code, undermining correctness. Reactive notebook alternatives automatically re-execute the minimal code necessary to refresh the state, but sacrifice expressiveness by disallowing common idioms like variable mutation. To address these limitations, we introduce Pync, a system that automatically memoizes function calls while correctly capturing and replaying their side-effects in general Python scripts. The system traces function calls by instrumenting program operations, recording their dependencies and side-effects upon the program state. On subsequent executions, Pync replays the side-effects and restores the return value of long-running functions. Benchmarks on a diverse set of data science notebooks indicate the potential for up to 32% performance boosts on subsequent executions of a script with minimal modifications.

This abstract has been withheld from publication.

Frequency combs are widely used in spectroscopy and frequency metrology, but their effectiveness relies heavily on the control of laser repetition rates and frequency offsets. Traditional systems often utilize pump-current feedback, which responds too slowly to suppress fast-changing phase noise. This project examines whether an analog feed-forward locking scheme could yield a self-referenced alternative with heightened short-term stability. We built a self-referenced frequency comb system around a 325 MHz Er:Yb glass oscillator. An f-2f interferometer was used to measure the carrier-envelope offset frequency, and that was combined with a repetition-rate-derived signal to drive an acousto-optic modulator to cancel the original offset. The comb repetition rate was then stabilized by phase-locking an optical beat note against a narrow-linewidth continuous-wave laser. Performance was evaluated using linewidth measurements, phase-noise analysis, and modified Allan deviation. Integrated carrier-envelope offset phase noise was measured to be 53.1 mrad from 10 Hz to 1 MHz, corresponding to 43.7 attoseconds of timing jitter, greatly outperforming conventional combs in short-term stability. However, it demonstrated worse long-term stability than traditional feedback, revealing a tradeoff between fast noise suppression and internal error accumulation. These results outline both the advantages and limitations of self-referenced feed-forward stabilization for future laser system designs in metrology and spectroscopy.