Gautam Ahuja
Senior Scientific Computing Engineer at Nygen Analytics, Sweden. I build tools for single-cell genomics, from analysing tens of millions of cells to AI agents that annotate them and cite the evidence.
Research interests agentic AI for science | single-cell & multi-omics | graph & signal | language models | biomedical multimodal learning
WUB WBSLQJ "ELOO" DQBZKHUH.
Fig. 1. A synthetic single-cell atlas of an imaginary tissue, grown in your browser from seed —— (…). Labels come from the Cell Ontology. A new atlas re-embeds every 10 seconds; hover over a cluster to annotate it.
Fig. 1, interrupted. Meet Bill: a one-eyed triangle in a top hat and bow tie, a dream demon who lives in the mindscape, makes deals with anyone who will shake his hand, and hides his secrets in ciphers. The ten symbols around him mark the ten people foretold to stop him; hover over one to read it. Drawn with the atlas's own cells. Reality returns in a few seconds.
About
At Nygen Analytics, I build software and AI systems for single-cell and omics data. The goal is biological analysis that is more useful, more reproducible and able to scale. It starts with the data itself: I work on CyteArc, which analyses RNA, ATAC and multi-omic experiments without loading them into memory, and CyteBase, which opens public datasets straight from the cloud. On top of those sit agentic systems that reason over the results: CyteType annotates cell types and cites its evidence, and CyteOnto checks those annotations against the Cell Ontology.
I studied computer science and mathematics at Ashoka. Before Nygen, I worked on biomedical AI at Ashoka University's Koita Centre for Digital Health and Mphasis AI & Applied Tech Lab, on clinical NLP, medical imaging, knowledge graphs, health-data interoperability and LLM-based decision support.
I'm most interested in how AI systems can work with structured scientific knowledge, such as ontologies and knowledge graphs, and with high-dimensional biological data.
I also like to teach, mentor and collaborate with students. I have taught programming, discrete mathematics, machine learning, data science (and some electronics) at Ashoka, IIT Jodhpur, Plaksha University. I take the Ashoka Horizons course for high school students.
Selected work
-
A multi-agent AI system that annotates cell types in single-cell RNA-seq data and shows the evidence behind each call, mapped to the Cell Ontology. Available for Python and for R/Seurat (CyteTypeR).
-
Compares two sets of cell-type annotations by mapping each label onto the Cell Ontology with LLM-written descriptions and embeddings, so synonyms and near-misses are scored fairly rather than by exact string match.
-
Our open-source engine for single-cell analysis: RNA, ATAC, protein and multi-omic data analysed out of core, scaling past a hundred million cells within a set memory budget, with provenance tracked through every result.
-
A searchable catalogue of public single-cell datasets that open directly as cloud-hosted CyteArc stores, so an analysis can start without downloading anything first. I version the dataset buckets by release.
-
Clinical decision support for abdominal pain in which physician-written diagnostic rules do the reasoning and small open-source language models handle the conversation and symptom extraction. Built with gastroenterologists at AIIMS New Delhi; accepted at GenAI4Health @ NeurIPS 2025.
Publications
-
Multi-agent AI enables evidence-based cell annotation in single-cell transcriptomics
bioRxiv, 2025
CyteType, in which LLM agents propose competing cell-type hypotheses, check them against biological databases and score their own confidence. It is benchmarked with CyteOnto against the Cell Ontology.
-
PRISM: Physician Rules Integrated with Small large language Models for probable diagnoses associated with Abdominal Pain
GenAI for Health Workshop (GenAI4Health) at NeurIPS 2025, 2025
Physician-written rules do the diagnostic reasoning while small open language models handle conversation and symptom extraction, to suggest probable diagnoses for abdominal pain in outpatient settings.
workshop -
Artificial intelligence for automated classification of antinuclear-antibody indirect immunofluorescence patterns
EULAR Rheumatology Open 2(1), 2026
An AI model that classifies antinuclear-antibody staining patterns in immunofluorescence images, automating a manual and subjective laboratory read.
Order from local rules
Fig. 3. Order from local rules. Nothing here follows a blueprint: each element responds only to its immediate neighbours, and the structure emerges anyway.
a Turing patterns. In 1952 Alan Turing showed that two chemicals which diffuse and react can organise themselves into spots and stripes, a possible mechanism for how tissues get their shape. This is the Gray–Scott model. Every few seconds a patch is cleared, like a small wound, and the pattern grows back into it. Click or drag to add a drop of v.
∂u/∂t = Du∇²u − uv² + F(1 − u) ∂v/∂t = Dv∇²v + uv² − (F + k)v
b Flocking. In Craig Reynolds' boids (1987), each bird steers away from crowding neighbours, matches their heading and moves toward their centre. No bird leads, yet the flock turns as one. Move your cursor through it to play the predator.
vi ← vi + ws Σj (xi − xj) / ‖xi − xj‖² + wa(⟨v⟩i − vi) + wc(⟨x⟩i − xi)
The sum and the averages ⟨·⟩i run over the neighbours j within a small radius of bird i. Here each bird also looks only ahead of itself and has a slowly wandering urge to turn; without that, the flocks soon freeze into one.
News
CyteBase is out: a searchable catalogue of public single-cell datasets that open as cloud-hosted CyteArc stores, no download needed. (PyPI | docs)
CyteArc is out: fast, out-of-core single-cell analysis for RNA, ATAC, protein and multi-omic data, scaling past a hundred million cells, with provenance tracked through every result. (PyPI | docs)
Submitted our write-up to the AMP Challenge (NeurIPS 2026 Competition Track), a three-stage generative pipeline for designing antimicrobial peptides. Slides, posters and other write-ups are now on the journal page.
Moved to Bangalore (temporarily)!
The DREAM Olfactory Mixtures challenge paper, A semantic-based community model for high-fidelity tuning of olfactory mixture distances, is published in PNAS. I’m part of the challenge consortium.
Promoted to Senior Scientific Computing Engineer at Nygen Analytics, where I look after AI at Nygen, including CyteType and CyteOnto.
Contact
Email is the best way to reach me. I'm always happy to talk and collaborate.
gautam.ahuja [at] nygen [dot] io