Diptanshu Sikdar

About

AI Reasoning for Science

I am a Computer Science Ph.D. student at UC San Diego, advised by Prof. Lianhui Qin. My research focuses on constrained reasoning at both the model and agentic levels: how models and AI agents can reason reliably when their outputs have to respect physical, biological, and experimental constraints.

I am currently involved in three AI for Science projects spanning geophysics and cell biology. Earlier, my work involved histopathology images, single-cell and spatial transcriptomics, and multimodal generative models.

Before UCSD, I completed my B.S. in Computer Science with a specialization in Intelligent Systems at the University of California, Irvine, where I worked in Zhang Laboratory on vision-omics learning, tissue morphology generation, and cell-cell communication analysis.

2026 - present Ph.D. in Computer Science, UC San Diego

Advised by Prof. Lianhui Qin.

2022 - 2026 B.S. in Computer Science, UC Irvine

Specialization in Intelligent Systems.

Research Directions

Constrained reasoning within and across models, applied to scientific discovery.

Level 01

Intra-model reasoning

Complex reasoning inside a single model, where generation follows hard constraints such as physical laws and biological mechanisms, and soft constraints such as priors and the spatial and temporal organization of the data, built directly into generative dynamics like flow matching.

Level 02

Inter-model reasoning

Reasoning at the agentic level, where multiple models interact as agents to ask questions, form and critique hypotheses, and explore complex scientific problems systematically within explicit constraints.

AI for Science Projects

? Question Hypothesize Explore Critique
  • Inter-model reasoning
  • Geophysics
  • MAESTRO initiative

Multi-Agent ScienTific Exploration and Research System

An AI-driven framework designed to accelerate scientific discovery by fostering deeper reasoning, asking meaningful questions, and systematically exploring complex scientific problems. Developed for the Geophysics MAESTRO initiative.

0 5 10 15 20 min
  • Benchmark
  • Single-cell dynamics
  • Virtual cells

ChronoBench

A planned resource and reproducible benchmark built around ChronoSeq perturbation time courses.

What does minute-scale single-cell data reveal about cellular responses, and how can these measurements support the development and evaluation of dynamic virtual-cell models?
  • Intra-model reasoning
  • Spatial transcriptomics
  • Cell-cell communication

Spatial Perturbation-Induced Communication Evolution

A world model of tissue: given a genetic perturbation, SPICE simulates how cells and their communication evolve in space and time by bridging spatial transcriptomics (ST) and cell-cell communication (CCC) data. Its novelty lies in jointly learning the highly dynamic nature of cells alongside genetic perturbations, and in modeling how their effects propagate across neighboring cells.