The 2026 QBI Hackathon marked the largest edition of the event to date, bringing together 19 teams for 48 hours of collaboration at the intersection of biology, computation, and medicine. Participants from academia and industry worked across disciplines to tackle active research questions, develop new computational tools, and generate fresh biological insights.

From artificial intelligence and protein dynamics to more accessible scientific software, teams applied diverse expertise to real-world challenges while building new connections and exchanging ideas. Rooted in open science and cross-disciplinary collaboration, the QBI Hackathon Series continues to foster innovation and accelerate discovery. Explore the 2026 team projects and browse links to their code below.

EvoEmu is an in silico directed evolution platform that engineers protein sequences toward user-defined conformational landscapes. Current AI tools predict structure but ignore this equilibrium. EvoEmu uses BioEmu, a generative deep learning biomolecular emulator, to sample conformational ensembles for any sequence, then scores mutations using the log-likelihood ratio as a thermodynamic fingerprint of the landscape. A genetic algorithm iteratively proposes and ranks mutations, converging toward a target LLR without exhaustive search. Applications span oncology, crop engineering, and therapeutic antibody optimization.

Members: Nikhila Manikandan, Ibraham Aasim, Diego Velazquez, Gilbert Shen, Imane Jannanen, Antranig Baghdasssarian

HiveBlot extracts structured experimental evidence from western blot figures in published papers, something no existing database does. Western blots appear in roughly 8–9% of protein publications, but the data is locked inside figure images and unstructured captions, forcing researchers into hours of manual literature review to answer basic questions like "has anyone shown protein X in cell line Y?" HiveBlot takes uploaded PDFs, parses figures and captions, classifies western blots using Qwen 3-VL vision, and extracts per-lane evidence: protein, cell line, organism, experimental condition, result, and antibody. Extracted records are stored in Supabase PostgreSQL and queried through a natural language interface. FastAPI converts researcher questions into SQL via LLM, returning structured, citation-linked results instantly. What took a week of manual searching now takes seconds.

Members: Nikhilesh Suravarjjala, Yashvi Jaladi, Shrushti Madhure, Saitheeraj Thatigotla, Vedant Shinde, Suhas Kunte, Ananya Jha, Shravan Arumugam, Tanvi Bharadwaj, Thim Paruvangada

AGeneTic is an agentic AI system that addresses a critical challenge in genomics: interpreting the biological meaning of large-scale gene perturbation data interactively and at scale. Motivated by the complexity of gene regulation — exemplified by a 2024 Nature study testing 2,000 potential causal genes to observe cellular responses to perturbation — AGeneTic aims to accelerate and streamline this interpretive process. The system features two core agents: a Literature Evidence Retrieval Agent, which uses a Retrieval-Augmented Generation (RAG) architecture (adapted from PaperQA2) to search public paper databases, chunk and embed text into semantic vectors, rank relevant evidence, and generate grounded answers while mitigating LLM hallucination; and an Experimental Evidence Retrieval Agent, which infers first-degree gene perturbation effects on downstream genes. Together, these agents produce traceable, multi-sourced annotated outputs from experimental data and literature, enabling fast, scalable, and interpretable analysis of large-scale genetic experiments. 

Members: Alexandra Mo, Lilly Lew

NeuroZip is a task-aware neural codec for single-trial EEG. The codec is trained to preserve the decodable semantic content of each epoch, defined as the CLIP embedding of the viewed stimulus, rather than to minimize waveform reconstruction error. A frozen EEG-to-CLIP projector scores whether stimulus identity survives compression, and the codec is trained against that score; gradient propagates through the frozen projector into the decoder. The fidelity baseline and the task-aware codec share architecture and training budget and differ only in the weight of the task term (lambda_task = 0 versus 3.0), so observed differences are attributable to the objective rather than to model capacity.

Two results follow on THINGS-EEG (subject sub-01). First, the compressed corpus remains searchable by natural language: a free-text query returns the epochs recorded while the subject viewed the corresponding object, although the model is never trained on text. At matched bitrate the task-aware codec improves top-5 image-prompt retrieval by 4 to 11 percentage points over the fidelity baseline, at a cost of at most 5 percent in reconstruction MSE. Second, because the objective retains only stimulus-decodable content, aggressive compression acts as an instrument for localizing task-relevant signal: at 144x compression the preserved information concentrates at occipitotemporal electrodes around the N170 (150 to 200 ms), and an independently trained classifier identifies the viewed object with 100 percent accuracy, versus 88 percent for the fidelity baseline. Object identity in single-subject EEG is therefore low-dimensional, spatially and temporally localized, and robust to compression. 

Members: Avinash Senthil, Rian Butala

The adoption of AI agents has reimagined how biomedical knowledge is structured, typed, and sourced. We present the Biomedical Open Knowledge Format (BioOKF), a knowledge harness defining a closed, molecularly grounded vocabulary of entities (genes, variants, molecules) and relationships, with provenance required on every claim. Within it, an agent reading unstructured sources (manuscripts, slide decks, posts) ingests concepts and relationships in a knowledge base that follows one standard across curators, models, and topics, and these knowledge bases can be merged with others to give agents persistent, well-sourced context. BioOKF extends Google's Open Knowledge Format and ships as an open-source stack, an MCP server, a command-line tool, and a desktop interface, linking curation, querying, linting, and visualization to the agent. In benchmarks holding the model and source text fixed, curating under BioOKF captured more biologically meaningful entities and relationships, represented them more accurately, and yielded more explainable bases than a bare model or generic OKF. With BioOKF, we aim to build a periodic table of biomedical knowledge that grows collaboratively in the generative-AI era.

Members: Wanjun Gu, Gianmarco Bellucci, Ilan Ladabaum, James Xue, Jonathon Xue, Xi Chen

To design a new binder (an antibody or mini-protein that sticks to a disease target), you first have to prepare that target structurally. Today a structural biologist does this by hand, and it can take hours per target.

Bonsai automates it. Give Bonsai a protein (a gene name, protein name, or UniProt accession) and it prepares a design-ready target: it resolves the UniProt entry, finds and ranks structures, trims to the relevant domain, identifies candidate hotspot residues and an epitope patch, builds an avoid set, and renders it all in an interactive 3D viewer. When no suitable experimental structure exists, it falls back to the AlphaFold Database automatically. A built-in chat assistant lets you ask why a structure was chosen or re-run with new settings in plain English. The output drops straight into de novo binder-design pipelines like BindCraft, RFdiffusion, or BoltzGen.

Members: Krishna Raja, Adrian Lobdill, Zachary Tokuno, Rucha Deo, Nabiel Wahedi

CellScope bridges the gap between raw microscopy data and actionable insights in stem cell research. Built on top of DeepSea — an image segmentation AI developed by the UCSC Ali Shariati's Lab — CellScope transforms per-cell shape statistics into meaningful differentiation classifications without requiring any programming expertise. DeepSea processes microscopy images and extracts morphological data at the individual cell level. CellScope then aggregates those measurements into per-image averages and maps them to a four-stage differentiation model, automatically classifying where cells fall along the developmental spectrum. Results are presented in plain, accessible language, making the pipeline usable by researchers at any technical level. By connecting a powerful segmentation backend to a researcher-friendly interface, CellScope democratizes stem cell analysis — turning complex morphological outputs into clear, interpretable science.

Members: Kalpita Balu, Devika G, Avlyn Singer, Dhanvanth R, Varun S, Rishima A

CryoClear is an open, real-time copilot for cryo-EM particle picking and junk removal. To build a 3D protein structure, scientists must first locate thousands of particles in noisy micrographs but automated pickers like Topaz over-pick junk (ice, carbon edges, aggregates), and cleaning it is slow, manual, and done after collection. CryoClear picks particles, flags junk by type as data streams in, and learns from the scientist's corrections live, so it improves with every interaction. Built on industry-standard pickers (Topaz, CryoSegNet) with a lightweight per-particle junk classifier and active learning, it reports real precision/recall against expert ground-truth (CryoPPP) and exports clean coordinates plus 2D class averages as proof the picks are real. Unlike closed platforms such as cryoSPARC Live, CryoClear is open-source (MIT), lightweight, and junk-aware running interactively on a single GPU. 

Members: Eva Lin, Tony Li, Bindu Bhargava Chintam

CryoSight is a deep learning-based platform for particle detection within Cryo Electron Tomographic Volumes. Its core backend utilizes a custom Unet-based neural network to classify and segment 6 different particle classes within tomograms. It also has a complementary graph-based neural network for additional insight on top of Unet Predictions for clustering, density, size, etc. The system is equipped with a UI stack of 3d renderings for tomograms and particle graphs, as well as an LLM-generated report for coherence and readability.

Members: Akshay Lakkur, Anton Kubo, Paul Gedrimas, Krishi Jaladi

A pipeline that places a resting-state EEG on a one-dimensional state-stability axis derived from microstate dynamics. Recordings are harmonized to a shared 16-channel 10-20 montage, resampled to 100 Hz, average-referenced, and band-pass filtered (1–40 Hz). Topographies are extracted at global-field-power peaks and clustered into four canonical maps via polarity-invariant modified k-means, then every recording is backfit onto a shared group template. From the state sequence we compute dwell time, coverage, occurrence, switch rate, and transition entropy (the conditional Shannon entropy of the transition matrix). A logistic-regression axis trained on 121 ADHD/control children separates the groups at leave-one-subject-out cross-validated AUC ≈ 0.70; spectral band power — including theta/beta — performs at chance. A Zen monk's empty-mind recording anchors the stable extreme. Placement is age-matched, seeded, and fully reproducible.

Members: Elisa Ma, YiRen Lin, Robert Khunis, Arjun Kumar

Our project aims to aid in the rehabilitation progress of patients with lateral ankle strains, which make up 85% of cases of the 2 million ankle sprains that occur annually in the US. Our project is a cluster of wearable biosensing stickers designed to monitor movement and localized mechanical strain around injured ligaments during rehabilitation. The flexible sticker, similar in size and material to a hydrocolloid patch, is placed at three sites of each major ankle ligament (ATFL, CFL, and PTFL). Our deliverable consists of a mockup of a mobile application that is paired with our product, and provides real time feedback throughout the recovery process rehabilitation. Key features include a color coded strain scale that classifies movement, a physical therapy/exercise log, a history of recent falls, and an option to contact patient physicians.

Members: Tulsi Arunkumar, Breena Bamaca, Aditi Bhat, Rishi Bhat, Sharvari Bulbule, Leilah-Louise Cuenca, Andy Tang

MelaTect is a multimodal melanoma risk assessment assistant that helps clinicians interpret both patient history and lesion images in a more transparent way. The system uses MedGemma in two specialized roles: one agent analyzes EHR-style clinical data such as age, family history, prior melanoma, and risk factors, while another analyzes the lesion image for suspicious visual features like asymmetry, border irregularity, and color variation. Each agent outputs a risk score, confidence level, and structured evidence. A reasoning layer then compares the two assessments, checks whether the confidence claims are supported, highlights agreement or disagreement, and produces a calibrated synthesis. Finally, an evidence-grounded report is generated for the clinician. MelanaTect is not meant to replace dermatologists, but to make AI-assisted melanoma screening more interpretable, auditable, and clinically useful.

Members: William Vo, Mark Amirkhan, Avi Kapur Srinivasan, Yash Shetty

Onco Pulse is a patient-specific cancer digital twin that turns static scans and genomics into a forecast of how a tumor grows. It fuses three layers into one pipeline. First, an XGBoost risk model trained on TCGA-BRCA and METABRIC expression data — with SHAP explainability — converts a patient's molecular profile into a biological aggressiveness score. That score parameterizes a 3D Fisher–Kolmogorov reaction–diffusion PDE that evolves tumor cell density voxel-by-voxel. Critically, the simulation is calibrated against real two-timepoint MRI: we fit the tissue invasion coefficient to each lesion's observed follow-up, seeded from expert segmentation masks. Validated across a 298-patient UCSF glioma cohort and TCGA-BRCA imaging, Onco Pulse renders results in an interactive 3D volumetric viewer with true-scale growth analytics (doubling time, RECIST) and an LLM narrative layer. Two cancers, one biophysically-grounded growth engine — bridging genomics, radiology, and simulation.

Members: Praneeth Mandalapu, Jasim Yasser, Chaandan Saini, Philip Thomas

ProteinMaxx is a preprocessing pipeline that enables antibody design against underresearched target proteins. Our tool uses ESMFold to generate a target protein structure from sequences, as well as FreeSASA to select hotspot residues. This data then gets piped into Germinal, a generative AI framework developed by Stanford in collaboration with the Arc Institute for designing epitope-targeted de novo antibodies.

Members: Tiara Ahmad, Sharlene Chiu

ProteomicLens is an interactive, full-stack visual analytics platform engineered for the UCSF QBI Hackathon to demystify high-dimensional, highly sparse single-cell mass spectrometry data. Designed to streamline modern workflows like SCoPE2, ProteomicLens guides researchers through an end-to-end processing pipeline: filtering low-quality cellular debris, correcting robust experimental batch effects, and dynamically clustering distinct cell types with 84% accuracy against FACS-sorted biological controls.

By integrating fluid 2D UMAP visualizations, researchers can seamlessly partition, search, and analyze complex cellular subpopulations. The platform incorporates a built-in literature scanner and an AI-assisted research console to instantly cross-reference single-cell protein abundance markers with active functional pathways. Complete with an interactive step visualizer, data impact calculator, and educational glossary, ProteomicLens bridges the gap between raw quantitative proteomics and rapid discovery. It is fully responsive, leveraging a highly performant stack (React, Vite, Tailwind CSS, Express) to make complex multi-omics accessible for clinical researchers.

Members: Hironmoy Paul and Eisha Paul

A prompt-only system for ternary gene expression classification (up/down/none) that predicts how a genetic perturbation affects a target gene using pure LLM reasoning — no training, no fine-tuning, just inference. For each perturbation–gene pair, the model reasons through the underlying biology and emits a probability distribution over the three outcomes, parsed into p_up and p_down mass.

Built on a GPT-OSS-120B endpoint with concurrent request handling, the system optimizes the exact official metric: the mean of two micro-AUROCs — DE (movers vs. none, scored on p_up + p_down) and DIR (up vs. down among true movers, scored on p_up / (p_up + p_down)). It validates on doubly-blocked splits that simultaneously hold out a fraction of perturbations and genes, so the dev set contains only unseen pert × gene combinations, mirroring the competition's disjoint test set and preventing leakage into few-shot examples. Decoding is pinned to temperature 0 with a fixed seed for deterministic, reproducible runs, and a fault-tolerant parser recovers predictions from noisy model output. 

Members: Marlene Lin, Wenjiang Zhou, Ze Wang, Stephanie Shi, Zhangziyi Zhou

GlycoAssessor is an open-source computational platform built for the structural analysis of N-glycans, complex carbohydrates whose altered forms serve as cancer biomarkers. Glycan structures are highly complex and difficult to quantify and researchers have lacked an objective method to compare them across diseases. GlycoAssessor addresses this issue by converting glycan structures into standardized matrices, either via the Distance and Connectivity Index (DCI) or the Position and Composition Index (PCI), that represent the spatial and branching architecture of N glycans. We analyzed prostate, breast, and colon cancer glycans using GlycoAssessor and identified cancer-specific structural subtypes. Our structural metrics can be used as biomarkers for diagnosis and prognosis of common cancers, and PCA space resulting from DCI can become a weighting factor that can be multiplied to measures of relative abundance. Our methods bring the field one step closer to cracking the “sugar code” of biology.