From one litre of water to a species list, an invasion forecast and a record nobody can quietly edit.
BioRadar turns raw eDNA sequencing reads into biodiversity intelligence a forest officer can act on,
in English or Hindi, without catching, seeing or disturbing a single organism.
Watch the intro · Quick start · How it works · Intelligence layer · Validation · Run guide
BioRadar in eight seconds. Full-quality video.
Table of contents
A litre of estuary water carries DNA from everything living in it: shed skin, scales, mucus, waste. Sequence it and you can list the species present without trapping or disturbing any of them. It is the cheapest biodiversity survey method there is.
The hard part is what comes back from the sequencer: millions of short text strings. Turning those into "an invasive tilapia is establishing in Vembanad Lake, and here is the evidence" takes a bioinformatics pipeline, a curated reference database and somewhere to put the answer that a decision-maker can read. BioRadar is that whole path, end to end.
| SIH25042 needs | BioRadar delivers |
|---|---|
| Species identification from eDNA reads | A QIIME 2 pipeline (cutadapt, DADA2 or vsearch, naive Bayes) against an India-curated COI reference of 33,611 sequences and 8,020 species |
| Understanding of Indian waters | 43% of the reference comes from Indian records; six validated coastal and lake baselines |
| Invasive-species detection and risk | Establishment-risk model, streaming anomaly alerts and Indian legal mapping per species |
| Where it is heading | Biodiversity forecast and hydro-corridor spread prediction |
| Results people can trust | SHA-256 of every artifact, chained per sample into a Merkle ledger; byte-identical output across runs |
| Usable by non-specialists | A bilingual (English / हिन्दी) dashboard and an auto-written executive brief |
- Reads your data and configures itself. Marker, primers, read length and quality encoding are detected from the reads. You never type a truncation length.
- Refuses to waste your time. Pre-flight catches unusable data in about a second, instead of failing 40 minutes into a run.
- Recovers awkward datasets. Data DADA2 cannot denoise is routed to vsearch OTU clustering rather than rejected, and reversed R1/R2 mates are detected and swapped.
- Names species against an India-curated reference. Audited for cross-order collisions, with the India/global split recorded per phylum.
- Proves the result was not altered. A Merkle ledger over every artifact. Two independent runs produce byte-identical output.
- Speaks to the field. Interactive map, bilingual interface, Darwin Core Archive export and multi-channel alerts.
A single-page app on http://localhost:8080, bilingual, with a light eco-green theme and a dark theme.
| View | What you do there |
|---|---|
| Home | The story, the science and quick jumps into the tools |
| Analyze | Drop a folder of paired-end FASTQ. BioRadar shows what it detected, then runs |
| Monitor | Every pipeline stage in real time, with live logs and progress |
| Results | Species table with read counts and confidence, site map, composition charts, engine output |
| Compare | Shannon, Simpson and richness across sites and sampling rounds (the Time Machine) |
| Alerts | Invasive-species anomalies and biosecurity flags, with severity and reasoning |
| Settings | Pipeline configuration, theme and language |
Built on top of the species table, each engine answers one question a decision-maker would ask.
| Engine | Answers |
|---|---|
| Biodiversity weather forecast | Where is diversity heading over the next rounds? |
| Multi-agent stakeholder debate | What would an ecologist, an economist and a fisher each conclude? |
| Zero-shot taxonomy classifier | Can a sequence with no reference match still be flagged? |
| Streaming anomaly alerts | Is this reading a real departure from the site's baseline? |
| Sentinel-2 change detection | Did the habitat around this site physically change? |
| Invasive-species risk | How likely is this alien species to establish here? |
| Extinction-risk viability | What is the trajectory of a threatened species under each scenario? |
| Hydro-corridor spread | Which downstream sites are next, and when? |
| Sampling-site optimiser | Where should the next survey go to learn the most? |
| PINN source finder | Where upstream did this eDNA signal come from? |
| Explainable AI | Why did the model decide that? |
| Field verification (TFLite) | Does a field photo corroborate the eDNA call? |
| Executive brief (NLG) | Write the page a decision-maker will actually read. |
| Biodiversity NFT receipt | Can a sponsor hold a tamper-proof record of a survey? |
Mathematics and implementation notes: docs/ADVANCED_FEATURES.md.
FASTQ reads
│ pre-flight quality, primers, orientation, truncation
│ cutadapt remove primer sequences
│ DADA2 / vsearch reads → amplicon sequence variants
│ naive Bayes variants → species, against the India-curated reference
│ normalizer QIIME 2 output → BioRadar's frozen data contract
│ SHA-256 chain-of-custody Merkle ledger
▼ intelligence engines: forecast, risk, spread, anomalies, executive brief
species report · map · forecast · tamper-evident record
- Pipeline. The bioinformatics layer ships as a Docker image so a run is reproducible on any machine.
- Contract. Everything downstream codes against one frozen schema, so engines never parse raw taxonomy strings.
- Intelligence and provenance. Engines read the normalized table; every artifact is hashed into the ledger.
Needs Docker Desktop, about 40 GB of free disk (the scientific image is 11.7 GB) and 8 GB of RAM. Set Docker's memory to 6 GB or more, or DADA2 will be killed mid-run.
docker pull ghcr.io/omtawde09/bioradar-pipeline:v1.0
git clone https://github.com/omtawde09/BioRadar.git
cd BioRadar
./scripts/setup.sh # Git Bash on Windows
docker compose up appOpen http://localhost:8080. Step-by-step with the Docker Desktop GUI: docs/RUNNING.md.
The bundled demo is 12 samples from 6 coastal sites, sampled twice, and takes about 3 minutes. It is simulated and labelled as such: every read comes from a real COI reference sequence, but which species occur where is invented. See docs/DEMO_DATASET.md.
Measured on an in silico mock community with known ground truth.
| Measure | Result |
|---|---|
| Detections recovered | 80 of 80, at exact read counts |
| False positives | 0 |
| Named species | 12 of 12, 0 unnamed records |
| Invasive growth between rounds | 1.9×, so the Time Machine has real change to detect |
| Unit tests | 76, about 8 s, no Docker needed |
| Integration checks | 15 |
The validation also caught a mislabelled record in the public NCBI reference data: two sequences one base apart, assigned to species in different taxonomic orders. Details in docs/DEMO_DATASET.md.
python -m pytest tests/ -q # 76 tests
./ci/check_integration.sh --fast # integration checks, skipping DockerBioRadar needs paired-end FASTQ files, two per sample.
MySample01_S1_L001_R1_001.fastq.gz
MySample01_S1_L001_R2_001.fastq.gz
Other layouts such as SRR123_1.fastq.gz are recognised and renamed. For the map, drop a samples.csv
next to the reads:
sample_id,site_id,latitude,longitude,collected_at
MySample01,MANDOVI,15.4989,73.8278,2026-01-15Then Analyze → Select folder → Upload & analyze. A run takes 3 to 40 minutes depending on depth.
| Layer | Technology |
|---|---|
| Bioinformatics | QIIME 2, DADA2, cutadapt, vsearch, Snakemake, packaged in Docker |
| Backend | Python 3.10+, standard-library HTTP server |
| AI and analytics | scikit-learn, NumPy, SciPy, PINN and PVA models, TFLite, LLM agents |
| Provenance | SHA-256 Merkle ledger, Solidity receipt contract |
| Web app | Vanilla-JS single-page app, Leaflet, custom charts, EN/HI i18n |
| Landing page | React 18, TypeScript, Tailwind CSS, Framer Motion, Lenis |
bioradar/ integration layer: webapp, contract, preflight, pipeline runner, ledger, exports
ai/ forecast, debate, zero-shot, IAS risk, PVA, PINN, spread, optimiser, XAI, NLG
analytics/ diversity forecasting and streaming anomaly detection
satellite/ Sentinel-2 change detection
blockchain/ receipt contract and generative-art engine
src/ landing page (React + TypeScript + Tailwind)
bioradar-pipeline/ Layer-1 bioinformatics image
data/ docs/ tests/ ci/
Everything came back "Unassigned". Is it broken?
No. The pipeline worked; the reference simply does not cover those organisms. That is a real result, and unassigned rows keep their read counts because they are often the most interesting ones.
Why did DADA2 fail?
Almost always memory. Raise Docker's limit to 6 to 8 GB.
Is the demo data real field data?
No, and it must not be presented as such. Classifications are genuine, but the site occupancy is simulated.
| Document | Covers |
|---|---|
| docs/RUNNING.md | Every way to run the system |
| docs/TESTING.md | Verifying each layer |
| docs/CONTRACTS.md | Data formats between components |
| docs/PIPELINE.md | The bioinformatics itself |
| docs/ADVANCED_FEATURES.md | The advanced engines, logic and mathematics |
| docs/DESIGN.md | Design system: tokens, components, accessibility |
BioRadar was a team entry for Smart India Hackathon 2026 (SIH25042) from K.C. College of Engineering, Thane, maintained by Om Tawde. I helped build the platform, including the species-to-intelligence layer and the validation that surfaced the mislabelled NCBI record.
Apache-2.0, see LICENSE. The Layer-1 pipeline derives from the eDNA-Container App by Wheeler, Brancalion, Kumar and Lintermans (NSW Department of Primary Industries), doi:10.3390/app14062641. Species assignment uses MIDORI2 and NCBI data; QIIME 2 is BSD 3-Clause. Full attribution in NOTICE.
Built for the Smart India Hackathon 2026 · SIH25042