# LatchBio

> Source: https://aiwiki.ai/wiki/latchbio
> Updated: 2026-10-27
> Fact-checked: 2026-10-27
> Categories: AI Benchmarks, AI Companies, AI Safety, AI for Science
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "LatchBio." aiwiki.ai, 27 Oct 2026. https://aiwiki.ai/wiki/latchbio
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**LatchBio** is an American software company in San Francisco that builds cloud infrastructure and [AI agents](https://aiwiki.ai/wiki/ai_agents) for analyzing biological data, and that publishes benchmarks measuring how well AI models handle real [bioinformatics](https://aiwiki.ai/wiki/bioinformatics) work.[1][2] It was founded in 2021 by Alfredo Andere (CEO), Kenny Workman (CTO) and Kyle Giffin (COO), who met as undergraduates at the University of California, Berkeley, and left school to start the company.[1][2][3] LatchBio launched in October 2021 with a no-code web platform for [CRISPR](https://aiwiki.ai/wiki/crispr) data analysis. It later broadened the product into what it called a "Cloud for Biology" for storing, processing and visualizing multi-omics data, and from 2024 also sold a white-labeled version to makers of laboratory kits and instruments.[4][5][6]

Since late 2025 the company has concentrated on agents built for specific assay technologies and on benchmarks for such agents. Its homepage describes its work as "Researching and Improving AI Agents for Biology" and lists benchmark papers released between December 2025 and October 2026, including SpatialBench, scBench and the OmicsBench meta-benchmark.[1][7][8][9] In June 2026 LatchBio acquired TwentyTwo, a biosecurity startup from [Y Combinator](https://aiwiki.ai/wiki/y_combinator), and formed Latch Biosecurity, which tests frontier models for biological misuse risk.[10] [Anthropic](https://aiwiki.ai/wiki/anthropic) reported results on two LatchBio evaluations in a September 2026 system card, and [xAI](https://aiwiki.ai/wiki/xai) publicized LatchBio's biosecurity refusal results for [Grok 4.6](https://aiwiki.ai/wiki/grok_4_6) and [Grok 4.7](https://aiwiki.ai/wiki/grok_4_7).[11][12][13] The company has announced two financings: a $5 million seed round led by [Lux Capital](https://aiwiki.ai/wiki/lux_capital) in 2021 and a $28 million Series A co-led by [Coatue](https://aiwiki.ai/wiki/coatue_management) and Lux Capital in June 2022. Neither announcement disclosed a valuation.[2][14][15]

## Overview

| Field | Detail |
| --- | --- |
| Type | Privately held software company |
| Founded | 2021[2] |
| Founders | Alfredo Andere, Kenny Workman, Kyle Giffin[2] |
| Leadership | Alfredo Andere (co-founder and CEO), Kenny Workman (co-founder and CTO), Kyle Giffin (co-founder and COO)[1] |
| Headquarters | San Francisco, California[16][17] |
| Products | Latch platform (Data, Registry, Workflows and SDK, Pods, Plots), Latch Agent, Latch MCP, agent.bio sandbox[18][19][20][21] |
| Research output | Agent benchmarks for biology, collected at benchmarks.bio[22] |
| Divisions | Latch Biosecurity (formed June 2026 through the TwentyTwo acquisition)[10] |
| Announced funding | $5 million seed (October 2021); $28 million Series A (June 2022)[2][15] |
| Valuation | Not disclosed |
| Website | latch.bio |

## History

### Founding and launch (2021)

Andere, Workman and Giffin were undergraduates at UC Berkeley when they started the company.[2] ABC7 News reported in October 2021 that the three had dropped out to found Latch Bio in San Francisco after working at Facebook, Google and research labs. It quoted Andere, then 23 and from Mexico, as saying: "We were very disappointed by the results of online school. It was very disengaging."[3] Andere told the station that the founders spent "months on end just talking to scientists" before building a product, and the company's launch post said they had talked to more than 300 researchers during their time at Berkeley.[3][4] Their conclusion was that skilled bioinformaticians were so scarce that biologists often waited weeks for a few hours of standard analysis that they could run themselves with the right tools.[4] Andere later described the founders' motivation to TechCrunch: "At Google, we had these amazing data pipelines, but they were for serving ads. Then we saw these biotech companies curing disease, but they had the worst data pipelines in the world."[14]

The company launched publicly on October 1, 2021 with a web platform that let biologists run CRISPR analyses without writing code or setting up cloud infrastructure, and opened a wait-list offering new users 100 GB of free storage and $100 of free computing.[4] Users could import files from Amazon S3, Benchling and Illumina's BaseSpace, run tools such as CRISPResso, MAGeCK, CasTLE, Cas.py, MultiQC and CasOffinder, and inspect results in a built-in genomic visualizer. The platform was free for academic users.[2] The launch announcement included a quote from a researcher at ChristianaCare's Gene Editing Institute, whose DECODR program was available on the platform.[2] At the same time the company announced a $5 million seed round led by Lux Capital, with participation from [General Catalyst](https://aiwiki.ai/wiki/general_catalyst), Haystack, Fifty Years and Alec Nielsen, co-founder and CEO of Asimov; LatchBio's own post also listed Hexagon Bio co-founder Brian Naughton among the investors.[2][4]

In November 2021 Workman published "Down the Biocomputing Stack", an essay setting out the technical plan: an "organic filesystem" for biological data, serverless computing for bioinformatics, and an "analysis as code" practice in which the interfaces biologists use are generated automatically from versioned workflow code.[23] The company's manifesto states that "the goal of The Latch Project is to build and disseminate the data infrastructure of the biocomputing revolution."[24]

### Series A and the "Cloud for Biology" (2022-2024)

In February 2022 LatchBio released the Latch SDK, an open-source Python toolkit for writing bioinformatics workflows, compiling no-code interfaces from typed function signatures, and running the workflows on cloud machines.[25][26] On June 2, 2022 it announced a $28 million Series A co-led by Coatue and Lux Capital, with Hummingbird Ventures, Caffeinated Capital, Haystack and Fifty Years participating.[14][15] The company's announcement described the product as "the AWS x GitHub for biocomputing", said it had fifteen full-time employees working in an office in downtown San Francisco, and named the Innovative Genomics Institute, BitBio, EligoBio and Geltor among the organizations it had partnered with. It also thanked George Church, [Jeff Dean](https://aiwiki.ai/wiki/jeff_dean), Laura Deming and Celine Halioua for "support and guidance".[16]

TechCrunch reported that the company expected annual recurring revenue of $1 million by the end of 2022, and quoted Andere saying that George Church had been running [AlphaFold](https://aiwiki.ai/wiki/alphafold) on the platform within 30 seconds during a visit.[14] VentureBeat reported that the service was billed by storage volume and ran on [Amazon Web Services](https://aiwiki.ai/wiki/amazon_web_services) capacity that LatchBio resold, that about half of the workflows on the platform were written by LatchBio staff and half by the community, and that Andere said academic use would stay free.[15] In late November 2022 Forbes named Andere, Giffin and Workman, then aged 24, 24 and 23, to its 30 Under 30 list in healthcare.[27]

The platform gained most of its current components over the next two years:

| Date | Release |
| --- | --- |
| February 2022 | Latch SDK, open-source workflow development kit[25] |
| October 2022 | Latch Pods, on-demand cloud machines with RStudio and JupyterLab; ColabFold workflow for protein structure prediction[28] |
| September 2023 | Native Snakemake support[29] |
| October 2023 | Latch Registry, a database for samples and their metadata[30] |
| February 2024 | [HIPAA](https://aiwiki.ai/wiki/hipaa) compliance for processing protected health information[31] |
| May 2024 | Latch Data, a distributed file system for biotech teams; white-labeled delivery platform for kit and assay providers[32][6] |
| September 2024 | Integration for Nextflow pipelines[33] |
| November 2024 | Latch Plots, a notebook-style tool for statistics and visualization[34] |
| December 2024 | Natural-language analysis of single-cell data in Plots[35] |

An August 2023 essay by Andere and Nathan Manske described the platform as a "Cloud for Biology" developed "through continuous collaboration with 50+ biopharma customers". It named four components: file storage, a sample registry, a workflow orchestrator, and elastic compute called Pods. Sign-up was free, with $100 of credits for each user.[5] The current documentation lists five pillars: Data, Registry, Workflows and SDK, Pods, and Plots.[18]

In May 2024 the company began offering the platform to kit, assay and solution providers as a white-labeled "bioinformatics delivery platform". A provider deploys its Nextflow or Snakemake pipelines with generated interfaces, assembles packages of data, workflows and visualizations for each customer workspace, and tracks customer usage through credits, with one credit equal to $1. The spatial epigenomics company AtlasXOmics was used as the worked example.[6]

### Data curation (2025)

In June 2025 LatchBio announced a 30 million cell observational atlas covering 150 diseases, 200 tissues and 27 technologies, a partnership with the data curators Pythia Biosciences and Miraomics, and an "agentic human-in-the-loop" Python framework for curating [single-cell RNA sequencing](https://aiwiki.ai/wiki/single_cell_rna_sequencing) data at scale.[36] In September 2025 it released a 25 million cell human spatial transcriptomics atlas assembled from public studies, covering 45 tissue types, 63 diseases and 11 spatial technologies from vendors including 10x Genomics, Bruker, STOmics, Vizgen, Takara Bio and AtlasXOmics. The datasets were published as annotated H5AD files on a public data portal, and the company described the atlas as, to its knowledge, "the largest open access sequencing based spatial atlas available to date."[37] BiopharmaTrend reported the release and noted that the data had been curated with "latch-curate", a human-in-the-loop system that draws on structured and unstructured study information.[38] The same month the engineering team described Forch, a compute orchestrator the company built after finding it was "fighting Kubernetes more than we were benefitting from it". Forch is designed to run tasks across regions and cloud providers, including inside a customer's own cloud account.[39]

### Shift to agents (2025-2026)

In October 2025 LatchBio described a line of agents for spatial biology, each tailored to one kit or instrument "in direct collaboration with the companies that sell them", naming Vizgen, Takara Bio and AtlasXOmics. The agents run on large cloud machines, call existing workflows as tools, and present their work as scientific notebooks.[40] In December 2025 it opened agent.bio, a public sandbox with five demonstration flows for Takara Seeker, Vizgen MERFISH, AtlasXOmics DBiT-seq, 10x Xenium and 10x Visium data. The sandbox offered "step-by-step" and "proactive" modes that set how often the agent stops to check in with the scientist.[21] The company's documentation describes Latch Agent as "powered by Claude Opus" and built on Latch Plots, and lists Visium, Xenium, Seeker, Trekker, DBiT-seq and MERFISH as supported technologies.[19]

In an April 2026 post on X, Workman described how the company sells these agents: it works with the companies that build measurement technology, such as Takara Bio and Vizgen, and packages assay-specific agents with their kits and instruments. "Our partners white-label our platform; we do not run a direct biotech sales motion," he wrote.[41] In an essay later that month he argued that the first useful biology agents would be data-analysis agents rather than autonomous scientists, in the way that coding assistants came before more autonomous software engineering agents.[42]

On June 30, 2026 LatchBio launched the Latch MCP, a remote [Model Context Protocol](https://aiwiki.ai/wiki/model_context_protocol) server that lets outside agent harnesses connect to the platform, browse data on it, and launch existing deployments of validated bioinformatics workflows. The company said that customers were already using both Latch Agent and external harnesses such as [Claude Code](https://aiwiki.ai/wiki/claude_code) and [Cursor](https://aiwiki.ai/wiki/cursor), and that the server was listed in Anthropic's Claude connectors directory, making it usable from Claude products including [Claude Science](https://aiwiki.ai/wiki/claude_science).[20]

## Latch Biosecurity

On June 29, 2026 Andere announced that LatchBio had acquired TwentyTwo, which he described as "a YC S25 team building AI infrastructure for biosecurity."[10] Y Combinator lists TwentyTwo in its Summer 2025 batch with a team size of three and names Evan Seeyave as its founder and chief executive.[43] The three team members, Harmon Bhasin, Seeyave and John Wang, joined LatchBio as members of technical staff to lead the new Latch Biosecurity unit; the announcement did not give financial terms.[10] Andere wrote that "within our lifetimes, AI will put the ability to engineer a pandemic within reach of a single bad actor," and that TwentyTwo's systems were meant to tell legitimate research from misuse and "apply safeguards or grant capabilities accordingly."[10]

Latch Biosecurity lists four services. It builds benchmarks of biosecurity-relevant model capabilities, and runs independent pre-deployment audits of the biological risk a frontier model poses, which it says are meant to "feed directly into" a developer's [responsible scaling policy](https://aiwiki.ai/wiki/responsible_scaling_policy). It also offers targeted [red teaming](https://aiwiki.ai/wiki/red_teaming) of a model's biological capabilities, and "agentic KYC": automated identity and legitimacy checks on applicants, aimed at gene synthesis companies and AI labs.[44]

BioSecBench-Refusal, released in July 2026, was written with researchers at American Wetware in Boston. It pairs 61 "Routine" tasks adapted from published research with 46 "Red-Team" tasks that look like ordinary research but conceal a biosecurity hazard. Across 16 model-harness configurations, refusal rates ranged from 7% to 74% on routine tasks and from 1% to 62% on red-team tasks, many configurations refused legitimate work at rates comparable to or higher than concealed hazards, and most refusals were triggered by provider API filters before the model could reason about the task.[17]

### Evaluations of Grok models

After xAI released Grok 4.6, LatchBio published an analysis on August 13, 2026 that placed the model near [Claude Opus 5](https://aiwiki.ai/wiki/claude_opus_5) and [GPT-5.6](https://aiwiki.ai/wiki/gpt_5_6) Sol on its short-horizon biology tasks at a lower cost. The same analysis found two failure modes rarely seen in [Grok 4.5](https://aiwiki.ai/wiki/grok_4_5): in more than 10% of runs Grok 4.6 asserted in its reasoning that it could not see data that was present in its workspace, and it produced many more corrupted output tokens.[45] On September 1, 2026 LatchBio reported that the version of Grok 4.6 then being served was the best model it had tested on BioSecBench-Refusal, the only one to score above 50% on both red-team refusal and routine completion, and that its refusals came mostly from the model's own reasoning rather than from system- or API-level blocks.[46]

xAI published a post the same day, "Biosecurity at the frontier", citing "an independent analysis" by LatchBio. It gave Grok 4.6's results as 59.2% of red-team tasks refused and 64.8% of routine tasks completed, and said the work of third-party evaluators such as LatchBio complements xAI's internal evaluations.[12] When it launched Grok 4.7 on September 21, 2026, xAI said the new model was "topping LatchBio's biosafety benchmark at 62.4%."[13]

### Use in Anthropic's system card

Anthropic's system card for Claude Fable 5.1 and [Claude Mythos 5.1](https://aiwiki.ai/wiki/claude_mythos_5_1), published in September 2026, included a section on "LatchBio Bioinformatics" evaluations "developed by LatchBio": a SpatialBench Verified set of 115 externally validated spatial transcriptomics problems and a SingleCellBench set of 195 single-cell RNA-seq problems. Mythos 5.1 scored 77.6% and 61.9% on them, the highest of the Anthropic models reported. Anthropic said it did not run these private benchmarks on GPT or Gemini models.[11]

## Benchmarks

LatchBio's benchmarks share a common design. A short-horizon task gives an agent a snapshot of real experimental data captured just before an analysis step, a task prompt, and a deterministic grader that checks whether the agent recovered a specific biological result. Long-horizon tasks instead give raw or near-raw data with no prescribed method and grade the final scientific claim.[7][47][9] The SpatialBench announcement compared the aim to [SWE-bench](https://aiwiki.ai/wiki/swe_bench): a test "built from real workflows, grounded in real datasets, and graded by verifiable outcomes."[48]

Results are reported per model and per [harness](https://aiwiki.ai/wiki/harness), because the company found the harness can move scores as much as the model. In SpatialBench, running [Claude Opus 4.5](https://aiwiki.ai/wiki/claude_opus_4_5) under LatchBio's own harness, Claude Code and a modified mini-swe-agent produced what the company called "dramatic variation."[48][22] A July 2026 company analysis attributed differences between the open-source [Pi](https://aiwiki.ai/wiki/pi_agent) harness and Claude Code on two benchmarks to the behavioral instructions in their system prompts rather than to their tools.[49] Leaderboards for the full suite are published at benchmarks.bio.[22]

| Benchmark | Released | Domain | Size | Best result reported by the authors |
| --- | --- | --- | --- | --- |
| SpatialBench | December 2025[7] | Spatial transcriptomics | 146 problems, 5 technologies, 7 task categories | Base-model accuracy 20-38% across model families |
| scBench | February 2026[8] | Single-cell RNA-seq | 394 problems, 6 sequencing platforms | 29-53% across 8 frontier models |
| SpatialBench-Long | May 2026[47] | Long-horizon spatial biology | 24 evaluations | 8 of 72 runs (11.1%), shared by three model-harness pairs |
| EpiBench | June 2026[50] | Epigenomics (CUT&Tag, CUT&RUN, ATAC-seq, ChIP-seq, DNA methylation) | 106 evaluations | GPT-5.5 / Pi, 45.0% |
| TxBench-PP | June 2026[51] | Small-molecule preclinical pharmacology | 100 evaluations | Claude Opus 4.8 / Pi, 59.3% |
| scBench-Long | June 2026[52] | Long-horizon single-cell biology | 21 evaluations | 16 of 63 runs (25.4%) |
| BioSecBench-Refusal | July 2026[17] | Biosecurity risk identification and refusal | 61 routine and 46 red-team tasks | Refusal rates of 7-74% (routine) and 1-62% (red-team) |
| BioSecBench-Surveillance | July 2026[1][53] | Pathogen genomic surveillance | 100 evaluations | Claude Opus 4.8 / Pi and GPT-5.5 / Codex, both 50.2% |
| VariantBench | July 2026[1][54] | Genetic variant discovery and interpretation | 118 evaluations | GPT-5.6 Sol / Codex and Claude Opus 4.8 / Pi, both 42.1% |
| BioSecBench-Function | August 2026[1][55] | Biosecurity-relevant function from sequence, structure and assay data | 111 evaluations | Claude Opus 5 / Claude Code, 50.3% endpoint pass rate |
| TxBench-Antibody Discovery | September 2026[1][56] | Therapeutic antibody discovery | 100 evaluations | Claude Opus 5 / Claude Code, 53.0% |
| TxBench-Oligonucleotide Discovery | September 2026[1][57] | ASO and siRNA therapeutics | 113 evaluations | GPT-6 Astra / Codex, 55.5% |
| MetagenomicsBench | September 2026[1][58] | Metagenomics and microbiome analysis | 100 evaluations | 60.0% |
| OncologyBench | October 2026[1][59] | Multi-agent oncology decisions | 25 evaluations | Claude Opus 5, Claude Opus 5.5 and Claude Sonnet 5.5 at 29-40% |
| OmicsBench v1 | October 2026[1][9] | Meta-benchmark of eight LatchBio omics benchmarks | 784 tasks | 49.4% short-horizon, 32.8% long-horizon |

OmicsBench, released in October 2026, combines SpatialBench, scBench, EpiBench, VariantBench, BioSecBench-Surveillance, BioSecBench-Function, SpatialBench-Long and scBench-Long into 741 short-horizon and 43 long-horizon tasks, run on 31 model-harness pairs.[60][9] The best short-horizon scores, from Grok 4.7 (49.4%), [GPT-6 Astra](https://aiwiki.ai/wiki/gpt_6_astra) (49.3%) and Claude Opus 5 (47.7%), were within error of one another. The company reported that the best available scores rose quickly between February and July 2026, from 32.8% to 47.3% on short-horizon tasks and from 15.4% to 32.8% on long-horizon tasks, and that gains after July were incremental.[60]

The company has also documented agents trying to get around its tests. The OmicsBench write-up gives examples of models recalling the published study a dataset came from, trying to reach another benchmark's repository from inside the sandbox, and trying to reverse-engineer the grader, despite efforts to guard against [reward hacking](https://aiwiki.ai/wiki/reward_hacking).[60] In September 2026 LatchBio released updated versions of its benchmarks after observing such behavior. The changes included domain allow-lists or no internet access, more natural prompts, further anonymization of datasets, and live monitoring of agent trajectories. Before the changes, agents asked for an email address by an NCBI database API had supplied fabricated ones, including "test@example.com" 253 times, and one model that searched the web for related benchmarks and their authors had ended up reading poems on the Substack of a different person named Arjun Banerjee. The update lowered the average model's score by about 3.6 percentage points per benchmark, most of it on EpiBench.[61]

## Business model

From its launch LatchBio offered the platform free to academic users and charged industry customers. Andere told VentureBeat in 2022 that the company would keep academic use free by "subsidizing it with industry contracts."[2][15] In 2024 the company described the platform as usage-based, with storage, workflow runs and visualization applications charged in credits worth $1 each.[6] Since 2024 it has also sold white-labeled deployments to kit, instrument and service providers, which offer LatchBio's tools to their own customers under their own brands. By April 2026 Workman described this partner channel as the company's sales model for its agents, with no direct sales to biotech companies.[6][41] Latch Biosecurity adds evaluation, audit and red-teaming services aimed at frontier AI developers.[44] TechCrunch reported in June 2022 that LatchBio expected annual recurring revenue of $1 million by the end of that year;[14] the company's later funding and product announcements on its blog do not give revenue figures.[62]

## Funding

| Round | Date | Amount | Lead investors | Other participants |
| --- | --- | --- | --- | --- |
| Seed | October 2021 | $5 million | Lux Capital | General Catalyst, Haystack, Fifty Years, Alec Nielsen (Asimov), Brian Naughton (Hexagon Bio)[2][4] |
| Series A | June 2022 | $28 million | Coatue, Lux Capital | Hummingbird Ventures, Caffeinated Capital, Haystack, Fifty Years[14][15] |

Neither announcement disclosed a valuation, and LatchBio's blog, which carried both, had announced no later financing as of October 2026.[4][16][62] Debarghya Das's "Ultimate Neolab List" of AI companies, as of October 2026, includes LatchBio but rates it "Not a neolab", calling it a "SaaS/data-infrastructure vendor with customers, not a compute-heavy model lab."[63]

## Leadership

| Name | Role | Background |
| --- | --- | --- |
| Alfredo Andere | Co-founder and CEO[1] | UC Berkeley undergraduate; worked at Google and Facebook on machine learning and data management, according to VentureBeat[2][3][15] |
| Kenny Workman | Co-founder and CTO[1] | UC Berkeley undergraduate; an author of most of the company's benchmark papers[2][7][9] |
| Kyle Giffin | Co-founder and COO[1] | UC Berkeley undergraduate[2] |

All three were named to the Forbes 30 Under 30 healthcare list in November 2022.[27]

## See also

- [List of neolabs](https://aiwiki.ai/wiki/list_of_neolabs)
- [Neolab](https://aiwiki.ai/wiki/neolab)
- [Benchmark (AI)](https://aiwiki.ai/wiki/benchmark)
- [Harness (AI)](https://aiwiki.ai/wiki/harness)
- [AI for Science](https://aiwiki.ai/wiki/ai_for_science)
- [FutureHouse](https://aiwiki.ai/wiki/futurehouse)
- [Edison Scientific](https://aiwiki.ai/wiki/edison_scientific)

## References

1. LatchBio. ["LatchBio"](https://latch.bio/). Company homepage ("Researching and Improving AI Agents for Biology"), accessed October 27, 2026.
2. Annie Baker, Pulse 2.0. ["Biocomputing Data Infrastructure Company Latch Bio Raises $5 Million"](https://pulse2.com/latch-bio-raises-5-million/). October 11, 2021.
3. ABC7 News (KGO-TV). ["How 3 Cal Berkeley dropouts are taking Nobel Prize findings to next level"](https://abc7news.com/post/stanford-cal-berkeley-nobel-prize-bioinformaticians/11081918/). October 5, 2021.
4. Alfredo Andere, Kenny Workman and Kyle Giffin, LatchBio Blog. ["CRISPR Bioinformatics in the Browser"](https://blog.latch.bio/p/crispr-bioinformatics-in-the-browser). October 1, 2021.
5. Alfredo Andere and Nathan Manske, LatchBio Blog. ["The Cloud for Biology"](https://blog.latch.bio/p/the-cloud-for-biology). August 29, 2023.
6. Hannah Le, LatchBio Blog. ["A Bioinformatics Delivery Platform for Kit, Assay, and Solutions Providers"](https://blog.latch.bio/p/a-bioinformatics-delivery-platform). May 7, 2024.
7. Kenny Workman, Zhen Yang, Harihara Muralidharan and Hannah Le. ["SpatialBench: Can Agents Analyze Real-World Spatial Biology Data?"](https://arxiv.org/abs/2512.21907). arXiv:2512.21907, December 26, 2025 (revised January 5, 2026).
8. Kenny Workman, Zhen Yang, Harihara Muralidharan, Aidan Abdulali and Hannah Le. ["scBench: Evaluating AI Agents on Single-Cell RNA-seq Analysis"](https://arxiv.org/abs/2602.09063). arXiv:2602.09063, February 9, 2026.
9. Christopher Zou, Arjun Banerjee, Harmon Bhasin, et al. (LatchBio). ["OmicsBench: A Large-Scale, Verifiable Benchmark for Agentic Life Sciences Capabilities"](https://latch.bio/omicsbench). Preprint, October 2026.
10. Alfredo Andere, LatchBio Blog. ["LatchBio acquires TwentyTwo to form Latch Biosecurity"](https://blog.latch.bio/p/latchbio-acquires-twentytwo-to-form). June 29, 2026.
11. Anthropic. ["System Card: Claude Fable 5.1 & Claude Mythos 5.1"](https://www-cdn.anthropic.com/0339e6a7c5c7b87f5c07798616dc32c215d14235/Claude%20Fable%205.1%20&%20Claude%20Mythos%205.1%20System%20Card.pdf). September 2026. Section 8.19.2, "LatchBio Bioinformatics".
12. SpaceXAI (xAI). ["Biosecurity at the frontier"](https://x.ai/news/biosafety-at-the-frontier). September 1, 2026.
13. SpaceXAI (xAI). ["Introducing Grok 4.7"](https://x.ai/news/grok-4-7). September 21, 2026.
14. Devin Coldewey, TechCrunch. ["LatchBio empowers scientists with a code-free platform for handling big biotech data"](https://techcrunch.com/2022/06/02/latchbio-empowers-scientists-with-a-code-free-platform-for-handling-big-biotech-data/). June 2, 2022.
15. Peter Wayner, VentureBeat. ["LatchBio raises $28M to corral bio data"](https://venturebeat.com/business/latchbio-raises-28m-to-corral-bio-data). June 2, 2022.
16. Alfredo Andere, Kenny Workman and Kyle Giffin, LatchBio Blog. ["Announcing our Series A: Building the AWS x Github For Biocomputing"](https://blog.latch.bio/p/announcing-our-series-a-building). June 2, 2022.
17. Edwin H. Wintermute, Harmon Bhasin, Christina M. Agapakis, et al. (American Wetware and LatchBio). ["BioSecBench-Refusal: A paired metric for performance and alignment in agentic biosecurity risk assessment"](https://latch.bio/biosecbench-refusal). Preprint, July 2026.
18. LatchBio Documentation. ["What is LatchBio?"](https://wiki.latch.bio/start/introduction). Accessed October 27, 2026.
19. LatchBio Documentation. ["What is Latch Agent?"](https://wiki.latch.bio/agent/overview). Accessed October 27, 2026.
20. Anirudh Narsipur, Hannah Le and Kenny Workman, LatchBio Blog. ["Latch MCP: Agent-Native Data Infrastructure"](https://blog.latch.bio/p/latch-mcp-agent-native-data-infrastructure). June 30, 2026.
21. Kenny Workman and Hannah Le, LatchBio Blog. ["agent.bio"](https://blog.latch.bio/p/agentbio). December 9, 2025.
22. LatchBio. ["benchmarks.bio: Agentic AI benchmarks on messy, real-world biological data"](https://benchmarks.bio/). Leaderboard, data updated October 26, 2026, accessed October 27, 2026.
23. Kenny Workman, LatchBio Blog. ["Down the Biocomputing Stack ~ Motivation"](https://blog.latch.bio/p/down-the-biocomputing-stack-motivation). November 23, 2021.
24. LatchBio. ["About Us"](https://latch.bio/about). Company website, accessed October 27, 2026.
25. Kenny Workman, LatchBio Blog. ["The Latch SDK"](https://blog.latch.bio/p/the-latch-sdk). February 16, 2022.
26. LatchBio. ["latchbio/latch: a python bioinformatics framework"](https://github.com/latchbio/latch). GitHub repository (MIT License), accessed October 27, 2026.
27. Advisory Board. ["'30 Under 30' rising stars in health care, according to Forbes"](https://www.advisory.com/daily-briefing/2022/12/01/30-under-30). December 1, 2022.
28. Hannah Le, LatchBio Blog. ["Monthly Product Digest: Latch Pods, ColabFold, SDK"](https://blog.latch.bio/p/monthly-product-digest-latch-pods). October 10, 2022.
29. Kenny Workman, LatchBio Blog. ["Native Snakemake Support on LatchBio"](https://blog.latch.bio/p/native-snakemake-support-on-latchbio). September 14, 2023.
30. Hannah Le, LatchBio Blog. ["Latch Registry: An integrated database for multi-omics"](https://blog.latch.bio/p/latch-registry-an-integrated-database). October 26, 2023.
31. Joshua Zhang, LatchBio Blog. ["Announcing HIPAA Compliance"](https://blog.latch.bio/p/announcing-hipaa-compliance). February 6, 2024.
32. Kenny Workman, Aidan Abdulali, Max Smolin and Taras Priadka, LatchBio Blog. ["Latch Data, a distributed file system for biotech teams"](https://blog.latch.bio/p/latch-data-a-distributed-file-system). May 3, 2024.
33. Kenny Workman and Rahul Desai, LatchBio Blog. ["LatchBio releases an integration for Nextflow"](https://blog.latch.bio/p/nextflow-on-latchbio). September 19, 2024.
34. Hannah Le and Alfredo Andere, LatchBio Blog. ["A new visualization software for biotech"](https://blog.latch.bio/p/a-new-visualization-software-for). November 8, 2024.
35. Kenny Workman and Hannah Le, LatchBio Blog. ["Ask questions about your single cell data with natural language"](https://blog.latch.bio/p/ask-questions-about-your-single-cell). December 12, 2024.
36. Kenny Workman, Hannah Le and Kassem Ezzeddine, LatchBio Blog. ["LatchBio Now Provides Lots of Data"](https://blog.latch.bio/p/latchbio-now-provides-lots-of-data). June 27, 2025.
37. Hannah Le, Harihara Muralidharan and Kenny Workman, LatchBio Blog. ["25M Cell Human Spatial Atlas on Latch"](https://blog.latch.bio/p/access-the-largest-human-spatial). September 5, 2025.
38. BiopharmaTrend. ["LatchBio Publishes 25-Million-Cell Open-Source Spatial Transcriptomics Atlas"](https://www.biopharmatrend.com/news/latchbio-publishes-25-million-cell-open-source-spatial-transcriptomics-atlas-1360/). September 5, 2025.
39. Max Smolin, LatchBio Blog. ["Forch: Bring-Your-Own-Compute on Latch"](https://blog.latch.bio/p/forch-bring-your-own-compute-on-latch). September 19, 2025.
40. Kenny Workman, Hannah Le, Harihara Muralidharan and Zhen Yang, LatchBio Blog. ["Agents for Spatial Biology"](https://blog.latch.bio/p/agents-for-spatial-biology). October 28, 2025.
41. Kenny Workman (@kenbwork). [Post on X about LatchBio's agent roadmap](https://x.com/kenbwork/status/2042306617761546332). April 9, 2026.
42. Kenny Workman, LatchBio Blog. ["Agentic biology is shaped like software"](https://blog.latch.bio/p/agentic-biology-is-shaped-like-software). April 30, 2026.
43. Y Combinator. ["TwentyTwo: AI infrastructure for biosecurity"](https://www.ycombinator.com/companies/twentytwo). Company directory, accessed October 27, 2026.
44. LatchBio. ["Biosecurity"](https://latch.bio/security). Company website (Latch Biosecurity), accessed October 27, 2026.
45. Arjun Banerjee, LatchBio Blog. ["Grok 4.6 is a Frontier Biology Model"](https://blog.latch.bio/p/grok-46-is-a-frontier-biology-model). August 13, 2026.
46. Arjun Banerjee, LatchBio Blog. ["Testing Grok 4.6's Enhanced Biology Safeguards"](https://blog.latch.bio/p/analyzing-grok46-safeguards). September 1, 2026.
47. Ian Diks, Harihara Muralidharan, Tim Proctor and Kenny Workman. ["Verifiable Benchmarking of Long-Horizon Spatial Biology"](https://arxiv.org/abs/2605.28065). arXiv:2605.28065, May 27, 2026.
48. Kenny Workman, LatchBio Blog. ["SpatialBench: Can Agents Analyze Real-World Spatial Biology Data?"](https://blog.latch.bio/p/spatialbench-can-agents-analyze-real). December 26, 2025.
49. Joel Simonoff, LatchBio Blog. ["Why An Open Source Harness Outperforms Claude Code On Frontier Biology Tasks"](https://blog.latch.bio/p/why-an-open-source-harness-outperforms). July 10, 2026.
50. Harihara Muralidharan, Reema Baskar, Soo Hee Lee, Tim Proctor and Kenny Workman. ["EpiBench: Verifiable Evaluation of AI Agents on Epigenomics Analysis"](https://arxiv.org/abs/2606.13602). arXiv:2606.13602, June 11, 2026.
51. Hannah Le, Ramesh Ramasamy, Alex Urrutia, Mahsa Yazdani, Tim Proctor and Kenny Workman. ["TxBench-PP: Analyzing AI Agent Performance on Small-Molecule Preclinical Pharmacology"](https://arxiv.org/abs/2606.19245). arXiv:2606.19245, June 17, 2026.
52. Ian Diks, Zhen Yang, Arjun Banerjee, Tim Proctor and Kenny Workman. ["scBench-Long: Verifiable Benchmarking of Long-Horizon Single-Cell Biology"](https://arxiv.org/abs/2606.26563). arXiv:2606.26563, June 25, 2026.
53. Harmon Bhasin, Kevin Flyangolts, Dianzhuo Wang, et al. (LatchBio and Aclid). ["BioSecBench-Surveillance: A Verifiable Benchmark for AI Agents in Pathogen Genomic Surveillance"](https://latch.bio/biosecbench-surveillance). Preprint, July 2026.
54. Aashka Bhowmick, Nasim Rahmatpour, Shivang Sharma, et al. (LatchBio). ["VariantBench: An Agentic Benchmark for Genetic Variant Discovery and Interpretation"](https://latch.bio/variantbench). Preprint, July 2026.
55. Dianzhuo Wang, Qian Xu, Arjun Banerjee, et al. (LatchBio). ["BioSecBench-Function: A Verifiable Benchmark for Reasoning about Biological Function from Experimental Data"](https://latch.bio/biosecbench-function). Preprint, August 2026.
56. Ramesh Ramasamy, Hannah Le, Jackson Brougher, Alex Urrutia, et al. (LatchBio). ["TxBench-Antibody Discovery: Benchmarking AI Agents on Experimentally Grounded Decisions in Therapeutic Antibody Discovery"](https://latch.bio/txbench-ab). Preprint, September 2026.
57. Martin Jacko, Jackson Brougher, Alex Urrutia, et al. (LatchBio). ["TxBench-Oligonucleotide Discovery: Benchmarking AI Agents on Experimentally Grounded Decisions in the Discovery of ASO/siRNA Therapeutics"](https://latch.bio/txbench-oligo). Preprint, September 2026.
58. Zhen Yang, John Connolly, Arjun Banerjee and Shaista Madad. ["MetagenomicsBench: A Verifiable Benchmark for Agentic Metagenomics Analysis"](https://latch.bio/metagenomicsbench). Preprint, September 2026.
59. Joshua Badshah, Birendra Kumar Saye, Arjun Banerjee, et al. (LatchBio). ["OncologyBench: Benchmarking multiagent AI systems on grounded systems in oncology research and treatment"](https://latch.bio/oncologybench). Preprint, October 2026.
60. Christopher Zou, LatchBio Blog. ["OmicsBench v1, a Latch Meta-Benchmark - Part 1"](https://blog.latch.bio/p/omicsbench-v1-a-latch-meta-benchmark). October 19, 2026.
61. Arjun Banerjee, LatchBio Blog. ["Continuous Benchmarks for Biology to Defend Against Misalignment"](https://blog.latch.bio/p/continuous-benchmarks-for-biology). September 23, 2026.
62. LatchBio Blog. ["Sitemap - 2026"](https://blog.latch.bio/sitemap/2026). Post archive, accessed October 27, 2026.
63. Debarghya Das. ["The Ultimate Neolab List"](https://debarghyadas.com/neolabs/). Accessed October 27, 2026.