My company, Superbio.ai just launched Locus, free through July. It lets you work with large-scale data — genomics, proteomics, clinical or research datasets, 100GB+ — by asking questions in plain English, with no pipelines or infrastructure to manage. Biotech teams can collapse weeks of workflow setup into a single conversation; everyone else can use it to analyze their own health data or dig into the research behind a treatment. Free for the month.
Sincere thanks to all of you who made it to Superbio’s webinar yesterday, introducing our genomics Agent Locus. To try the product, navigate here: app.superbio.ai.
Here are some helpful time stamps for skipping ahead:
6:15 - Beginning of Ronjon’s presentation on Superbio’s mission.
16:22 - Beginning of Ronjon’s product demo highlighting use cases.
25:20 - Beginning of product & feature overview from Eva.
33:24 - Overview of Complete Genomics sequencing technologies and an end-to-end whole-genome workflow on Superbio, from Senior Bioinformatics Scientist, Biniam Feleke.
42:45 - Beginning of Q&A with Ronjon and Superbio team.
This is cool. Just messing around with this now but I’d be interested to see people who have their genetic data play with this for a bit and post interesting outputs here.
The most significant advantage is that it is specifically trained for biological tasks. Unlike other models, it provides actual protocols and interventions directly, without the standard annoying disclaimers.
As a biology-trained LLM, it generates Python notebooks to run various bioinformatic tools. Because these tools run on high-performance cloud servers with GPUs, tasks that previously took days on my MacBook Pro are now completed in just a few hours.
For example, I used it to normalize, re-sequence and merge my genome from two 30x sequencing runs. The process took only a few hours and resulted in a 67x decoded genome.
I am currently using it to generate an optimized Thymus rejuvenation protocol.
Hi everyone, our team collaborated on responses to your questions from last week’s webinar. Please find our answers below, and feel free to reach out with any further questions or comments. We’ve also attached a spreadsheet detailing current genomics capabilities from Locus. Thanks again for joining us!
Q: Can full LLM be run on a local machine for 100% confidentiality? Or is cloud interaction required?
A: Not at this time - Superbio’s compute instances and storage live in the cloud. Our customers work with large data sizes, so enabling remote transfer is desirable. However, our next product release will enable the use of any LLM, including open-source.
Q: How different from NotebookLM with a collection of Gemini agents?
A: Superbio’s multi-agent framework combines planning, critique, scientific knowledge bases, and tool access - going beyond a basic LLM setup. During pipeline execution, the agent operates within a unified bioinformatics environment for large-scale data transfer and high-performance compute. This means working with terabytes of data and GPU instances, unlike what NotebookLM has access to, and a specialized set of skills for working with these data types. It is not possible to use these skills at all without such infrastructure.
Q: Where are your SuperBio queries actually running?
A: Superbio currently uses Anthropic’s API for LLM inference (though, in the future we will support other providers) for chat queries and responses. All code execution occurs on Superbio’s own CPU and GPU infrastructure, inside a secure sandbox.
Q: Do you support bulk RNA-seq or single -cell- seq analysis?
A: Yes, Superbio supports a range of bioinformatics tasks uniquely enabled by our infrastructure compared to other coding agents. This includes whole-genome sequencing, long-read analysis, somatic and germline variant calling, pangenome-aware calling, single-cell -omics analyses, and bulk RNA-seq. A full list of Locus capabilities is attached to this email. You can also browse Superbio’s self-serve applications by scrolling here: Superbio.ai.
Q: Will Superbio ask for user clarification when given a vague question? For example asking for HER2 status in a breast cancer tumor
A: Yes, unless the prompt is very simple, or conversely has a high level of specificity already, Superbio’s Planning Agent will lead the user through a Q&A to obtain key details. This can include clarifications on specific disease indication, overall scientific objective, file format, sequencing chemistry, etc. - depends on the circumstance.
Q: If someone wants to test a certain chemical compound for cancer, how would they use the tool? Can they prompt for experiment?
A: Locus has access to a suite of high-throughput genomics tools - unique to publicly available agents. It can also access cheminformatics and protein design tools. So, to the extent that a compound can be tested using data science tools like molecular docking, PPIs, toxicity prediction, or pathway analysis - yes, users can expect the Agent to suggest (and execute!) a full computational experiment in response to queries about drug testing. Lab testing will require independent follow-up from the user.
Q: What data is already part of the superbio ecosystem, that’s not fetched live during user prompt.
A: Locus has access to any connected user S3 buckets, and a container registry for genomics workflows created by Superbio. The container registry includes semantic information + metadata to enhance Agent decision-making. Agent also has pre-loaded reference genomes and integrations with the following: Ensembl, UniProt, NCBI, AlphaFold, GO, KEGG, ARCHS4, CELLxGENE, OpenTargets, cBioPortal, COSMIC, NCBI Virus. Other data integrations are spun up at runtime.
Q: How is superbio different from some of the other ai co-scientist, such as benchsci’s EMET, phylo’s biomni lab, Edison’s Kosmos, etc. I can see the individual applications as part of the utilities in the app. What else?
A: Today, Locus UX is a Jupyter IDE, enabling users to view and edit code, as well as analysis outputs. This is distinct from the chat-only interfaces common with other platforms. We also support vast datasets for ingestion, vs. Phylo as an example - their upload limit is 25GB, effectively precluding analyses like whole-genome sequencing (100GB data per sample), which Superbio supports. We have not found Claude Science to be stable with these data types, either.
Q: What about code optimization? Often Agents and AI code writing systems produce code that ‘works’ but often isn’t fully tested or optimized for speed or memory conservation. The importance of how to craft the prompt becomes very important, which still requires some expertise to know what to ask. To what extent are there any internal systems utilized to augment a prompt for any optimization?
A: Most of Locus’ tools are heavily optimized already, containerized for reproducibility reasons, and equipped with the optimal infrastructure to support performance. In practice, this means our algorithms can boast up to 30X time acceleration vs legacy tooling for some use cases, such as whole-genome sequencing. This time saving can enable experiments that were highly cumbersome or not previously possible. For example, Claude Science estimated a whole-genome sequencing data analysis workflow that takes 45 minutes on Superbio (and runs reproducibly each time) would take 1-2 days on Claude Science, with no subsequent reproducibility guarantee. Additionally, Locus will make suggestions during the Q&A flow about which data types are optimal to use, with compute and time consideration in mind.
I started using superbio.ai several days ago and positively impressed.
It allowed me to upload my 10 GB .vcf genome file, recognized its somewhat older/non-standard format (other services I tried failed to do so) and was able to interactively answer all my questions like “check for apoe4 and other risk variants”, “check for hypothyroidism risk”, “check for MTHFR Gene Mutation”, “check for other common health risks”.
Then I uploaded all my self-tracking data (food, sleep, exercise, bloodwork etc.) from Health Hub and got answers, reports, correlations, graphs for “analyze my self-tracking data from Sergey Vlasov.md file and suggest optimizations for better health”, “how much goitrogens in food currently consumed by 100 g” etc.
It looks authoritative, through and smart for biological research tasks.
Yes but are you sure it’s a FASTA? Generally it’s for reference genomes.
I uploaded 2 30x FASTQ full genome from 2 sequencing companies and had SuperBio normalize them. After that it ran the full pipeline and gave me a 67x merged BAM and annotated VCF.
I did both DeepVariant and HC variant calling on the BAM.
Takes some time though but so much faster than my MBP M4.
Sure. I had SuperBio run Parabricks BamMetrics QC (NVIDIA)
WGS Coverage Metrics (sample: MB)
Metric
Value
Assessment
Mean coverage
67.4x
Excellent (>2× clinical 30x standard)
Median coverage
75x
High-depth WGS
SD of coverage
26.9
Moderate variability (expected for WGS)
PCT_1X
93.3%
93.3% genome covered at ≥1×
PCT_10X
92.2%
92.2% covered at ≥10×
PCT_20X
91.2%
91.2% covered at ≥20×
PCT_30X
89.5%
PASS (≥80% at 30x is standard)
PCT_50X
81.1%
Very high coverage
Duplication exclusion
3.4%
Low dup rate
MAPQ exclusion
6.7%
Normal (low-mapq reads filtered)
Total bases excluded
19.9%
Acceptable
HET SNP sensitivity
92.9%
High sensitivity for heterozygous variants
HET SNP Q
12
Phred-scaled sensitivity score
BAM QC Summary (combined with earlier flagstat/idxstats)
Metric
Value
Assessment
Mapping rate
96.66%
PASS
Duplication rate
3.58%
PASS
Properly paired
95.28%
PASS
Mean coverage
67.4x
PASS
Coverage at 30x
89.5%
PASS
Inferred sex
XY (male)
From Y/X read ratio
QC verdict: All metrics pass WGS best-practice thresholds.
The sample is high-quality at ~67× depth — well above the 30× minimum for confident variant calling.
)
Excellent (>2× clinical 30x standard)
