Latent Labs
Last reviewed
Jun 7, 2026
Sources
16 citations
Review status
Source-backed
Revision
v1 · 1,822 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 7, 2026
Sources
16 citations
Review status
Source-backed
Revision
v1 · 1,822 words
Add missing citations, update stale details, or suggest a clearer explanation.
Latent Labs is a British and American artificial intelligence company building generative foundation models for protein and molecular design. It was founded in 2023 by Simon Kohl, a former research scientist on Google DeepMind's AlphaFold2 team who went on to co-lead DeepMind's protein design group. The company emerged from stealth on 13 February 2025 with $50 million in disclosed funding and describes its goal as making biology "programmable," meaning the design of novel proteins from a target specification rather than the prediction of existing structures. In July 2025 it released Latent-X, an all-atom generative model for de novo protein binder design, made available through a no-code web platform and an API. Unlike drug-discovery firms that develop proprietary medicines, Latent Labs positions itself as a model and platform provider whose tools are used by academic, biotech, and pharmaceutical researchers. It maintains offices in London and San Francisco and remains an early-stage research company.
Latent Labs was founded by Simon Kohl, who serves as chief executive. Kohl completed a PhD at the German Cancer Research Center (DKFZ) in Heidelberg, finishing in December 2019 with a dissertation on the semantic segmentation of ambiguous images. During that work he developed the Probabilistic U-Net, a generative approach to handling uncertainty in medical image segmentation that was presented as a spotlight at NeurIPS 2018. He subsequently joined DeepMind, where he was a research scientist on the core AlphaFold2 team and, according to his own biography, built AlphaFold2's widely used pLDDT confidence metric, the per-residue score that reports how reliable a predicted structure is. After AlphaFold2, Kohl spent roughly two years co-leading DeepMind's protein design team and helped establish the company's wet lab at the Francis Crick Institute in London.
AlphaFold2, published in Nature in 2021, contributed to the 2024 Nobel Prize in Chemistry awarded to Demis Hassabis and John Jumper for protein structure prediction. Kohl left DeepMind to start Latent Labs, with reporting placing the beginning of the effort in late 2022 and the company's London incorporation in mid-2023. Early hires reported at launch included other DeepMind alumni, a senior Microsoft engineer, and Cambridge PhDs, with a team of roughly 15 people split between the two offices. Latent Labs draws a deliberate contrast with Isomorphic Labs, the DeepMind spinout led by Hassabis that develops its own drug pipeline; Kohl has described Latent Labs instead as a service provider that gives external researchers direct access to protein design models.
Latent Labs' flagship model is Latent-X, described in a July 2025 technical report titled "Latent-X: An Atom-level Frontier Model for De Novo Protein Binder Design" (arXiv 2507.19375). According to the report, Latent-X is an all-atom protein design model: given a target protein epitope, it jointly generates the all-atom structure and the amino acid sequence of both the binder and the target, directly modelling the non-covalent interactions that drive specific binding. This contrasts with two-stage pipelines that first generate a backbone and then design a sequence onto it. The authors report that the end-to-end process is approximately one order of magnitude faster than existing multi-step computational pipelines, in part because structure and sequence are co-sampled rather than produced in separate steps. The work is credited to the Latent Labs team, with named contributors including Alex Bridgland, Jonathan Crabbe, Henry Kenlay, Daniella Pretorius, and Sebastian M. Schmon.
The company reports two therapeutic modalities. For macrocyclic peptides, the technical report states that Latent-X achieved experimental hit rates exceeding 90 percent across all evaluated benchmark targets. For mini-binders, the company reports that the model consistently produced potent candidates against all evaluated targets, with binding affinities reaching the low nanomolar and picomolar range, which the company characterizes as comparable to approved therapeutics, alongside high specificity on mammalian display. These results are reported by the company in its own technical report and have not, at the time of writing, been independently peer reviewed, so the performance figures should be attributed to Latent Labs rather than treated as established. The report further states that the laboratory validation used as few as 30 to 100 designs per target, a small number relative to the thousands of candidates often screened in conventional campaigns.
In head-to-head comparisons reported in the paper, Latent-X is benchmarked against state-of-the-art methods including AlphaProteo (from DeepMind), RFdiffusion, and RFpeptides, both descendants of the protein design lineage associated with David Baker's laboratory. The authors report that Latent-X produced binders with higher hit rates, stronger binding affinities, and more structurally diverse designs, including complex beta-sheet folds. Public descriptions characterize the underlying approach as a generative model that operates over atomic coordinates; the company has aligned its method with the broader family of generative structure models that includes diffusion and flow-based approaches, though the precise training objective is detailed only in the technical report.
Latent-X was launched publicly on 21 July 2025 and is offered through a no-code web platform at platform.latentlabs.com, along with programmatic access via an API. The platform lets a user upload a target protein, select hotspot residues on the target surface, and generate candidate macrocycles or mini-binders, then computationally score and rank the resulting designs within minutes. The interface includes structure overlays, hotspot selection, and ranking tools intended to let researchers prioritize candidates without maintaining their own machine learning infrastructure. The company frames this browser-based delivery as a way to broaden access to protein design beyond groups with large computational teams.
At launch, Latent-X was free to use. In interviews around the release, Kohl indicated that the company intends to eventually charge for advanced features and capabilities, consistent with a platform and licensing business model rather than internal drug development. Kohl has drawn a sharp line between Latent-X and AlphaFold, noting that AlphaFold lets a researcher visualize existing structures but does not generate new proteins, whereas Latent-X is designed to produce entirely novel binders. The stated intended users span academic institutions, biotech startups, and pharmaceutical companies.
Latent Labs disclosed $50 million in total funding when it emerged from stealth in February 2025. This comprised a previously unannounced $10 million seed round and a $40 million Series A. The Series A was co-led by Radical Ventures and Sofinnova Partners, with participation from Flying Fish and Isomer alongside the existing seed investors. The seed round came from 8VC, Kindred Capital, and Pillar VC. Named angel investors include Google chief scientist Jeff Dean, Cohere co-founder and Transformer co-author Aidan Gomez, and ElevenLabs co-founder Mati Staniszewski. Aaron Rosenberg of Radical Ventures, a former DeepMind head of strategy and operations, was reported among the backers; some coverage also lists Anthropic chief executive Dario Amodei as a supporter, a single-source claim that is noted here but not independently corroborated.
| Item | Detail |
|---|---|
| Founded | 2023 (London incorporation; effort began late 2022) |
| Founder and CEO | Simon Kohl (former DeepMind AlphaFold2 scientist) |
| Headquarters | London, United Kingdom; San Francisco, California |
| Stealth exit | 13 February 2025 |
| Total disclosed funding | $50 million |
| Seed round | $10 million (8VC, Kindred Capital, Pillar VC) |
| Series A | $40 million, co-led by Radical Ventures and Sofinnova Partners |
| Series A participants | Flying Fish, Isomer, plus seed investors |
| Notable angels | Jeff Dean, Aidan Gomez, Mati Staniszewski |
| Flagship model | Latent-X (released 21 July 2025) |
| Model type | All-atom de novo protein binder generator (structure plus sequence) |
| Modalities | Macrocyclic peptides; mini-binders |
| Reported macrocycle hit rate | Above 90% on evaluated targets (company report) |
| Reported mini-binder affinity | Low nanomolar to picomolar (company report) |
| Access | No-code web platform and API |
Latent Labs sits within a wave of companies extending the AlphaFold lineage from structure prediction toward generative design, an effort sometimes grouped under AI for science. Its founder's direct provenance on AlphaFold2, including authorship of the pLDDT confidence score, gives the company unusually close ties to that work. The field it competes in is anchored by two research traditions: DeepMind's structure-prediction work that produced AlphaFold and AlphaProteo, and David Baker's laboratory at the University of Washington, whose RFdiffusion and related tools pioneered diffusion-based de novo protein design and whose contribution was also recognized in the 2024 Nobel Prize in Chemistry. Latent-X is presented by its authors as competitive with or ahead of methods from both traditions on the specific task of binder design, though those claims rest on the company's own technical report rather than independent replication.
The company's business model is its other notable feature. Where Isomorphic Labs and Xaira Therapeutics aim to develop proprietary medicines, and where many drug discovery startups keep their models internal, Latent Labs has chosen to expose its frontier model directly to outside researchers through a browser, an approach closer to how general-purpose AI models are distributed. If the reported wet-lab hit rates hold up under broader and independent testing, the combination of all-atom joint generation, fast turnaround, and open web access could lower the barrier to protein engineering for smaller laboratories. As of mid-2026 Latent Labs remains an early-stage research company, and the durability of its performance claims and its platform business will depend on validation and adoption that extend beyond its initial disclosures.