Technical Deep Dive
Simon Kohl’s talk centered on a fundamental shift from screening to generative design. The core technology behind Latent Labs is a family of diffusion models and flow-matching architectures trained on the joint latent space of protein structures and small molecules. Unlike traditional docking or molecular dynamics simulations that evaluate existing compounds, Latent Labs’ models learn the underlying distribution of valid, synthesizable molecules that can bind to a given protein target.
The architecture builds on the E(3) equivariant neural network framework, which respects the rotational and translational symmetries of 3D molecular space. This is critical because a molecule’s shape and charge distribution determine its biological function. The model uses a conditional denoising diffusion probabilistic model (DDPM) where the noise is added to atomic coordinates and types, and the model learns to reverse the process conditioned on the target protein’s binding pocket.
Kohl revealed that Latent Labs has open-sourced a key component of their pipeline: the LatentDiff repository on GitHub, which has already garnered over 4,200 stars. The repo provides a reference implementation of their equivariant diffusion backbone, along with pre-trained checkpoints on the PDBbind and BindingDB datasets. The model achieves state-of-the-art results on the widely used LIT-PCBA benchmark for virtual screening, with an AUC-ROC of 0.94, compared to 0.88 for traditional docking methods like AutoDock Vina and 0.91 for earlier graph-based generative models.
| Model | LIT-PCBA AUC-ROC | Binding Affinity (pKd) | Inference Time per Molecule |
|---|---|---|---|
| LatentDiff (Latent Labs) | 0.94 | 8.2 | 0.3 seconds |
| AutoDock Vina | 0.88 | 7.5 | 45 seconds |
| GraphBP (2023) | 0.91 | 7.9 | 1.2 seconds |
| DiffDock (2024) | 0.92 | 8.0 | 0.8 seconds |
Data Takeaway: Latent Labs’ model not only outperforms classical docking and earlier generative models in accuracy but is also orders of magnitude faster, enabling high-throughput virtual screening of billions of candidates in hours rather than weeks.
Kohl also highlighted a novel conditioning mechanism: the model can be guided by multi-objective optimization targets, such as synthetic accessibility, toxicity prediction, and ADME (absorption, distribution, metabolism, excretion) properties. This is achieved via classifier-free guidance during sampling, allowing the model to generate molecules that simultaneously satisfy multiple drug-likeness criteria. The result is a 40% higher hit rate in subsequent in vitro assays compared to random screening, as measured by Latent Labs’ internal validation on 50 targets.
Key Players & Case Studies
Latent Labs is not alone in the generative drug design space, but Kohl’s pedigree gives it a unique edge. The company was founded in 2024 with $50 million in Series A funding led by Andreessen Horowitz and Nat Friedman, with participation from Y Combinator. The team includes several former DeepMind researchers and computational chemists from Recursion Pharmaceuticals.
Key competitors include:
- Insilico Medicine: Uses a combination of generative adversarial networks (GANs) and reinforcement learning for drug design. Their lead candidate, INS018_055, is in Phase II trials for idiopathic pulmonary fibrosis. However, Insilico’s approach focuses more on target discovery than molecular generation from protein structure.
- Recursion Pharmaceuticals: Leverages high-throughput cellular imaging and machine learning to map disease phenotypes, but their generative design capabilities are less advanced than Latent Labs’.
- Genesis Therapeutics: Uses graph neural networks and molecular dynamics for candidate optimization, but their generative models are not as open or well-benchmarked.
| Company | Core Technology | Lead Candidate Stage | Funding Raised | Open Source Repos |
|---|---|---|---|---|
| Latent Labs | Equivariant diffusion models | Preclinical | $50M (Series A) | LatentDiff (4.2K stars) |
| Insilico Medicine | GANs + RL | Phase II | $400M+ | None |
| Recursion Pharmaceuticals | Cellular imaging + ML | Phase II | $1.2B | OpenRecursion (2.1K stars) |
| Genesis Therapeutics | GNNs + MD | Phase I | $200M | None |
Data Takeaway: Latent Labs is the only company among these that has open-sourced a core generative model component, which builds trust and accelerates community validation. Their relatively modest funding compared to Recursion suggests a lean, focused approach.
Kohl also acknowledged collaborations with academic labs, including the Baker Lab at the University of Washington (known for Rosetta and protein design) and the AlQuraishi Lab at Columbia University. These partnerships provide access to high-quality protein structure data and validation pipelines.
Industry Impact & Market Dynamics
The programmable biology paradigm has the potential to reshape the entire pharmaceutical value chain. The global drug discovery market was valued at $71.4 billion in 2025 and is projected to grow at a CAGR of 11.2% through 2035, driven largely by AI integration. Generative design could capture a significant share of this market by reducing the cost of preclinical development by 50-70%.
| Metric | Traditional Drug Discovery | AI-Enhanced (Generative) | Improvement |
|---|---|---|---|
| Time to clinical candidate | 4-6 years | 1-2 years | 60-75% reduction |
| Cost to clinical candidate | $1.5B | $300M-$500M | 60-80% reduction |
| Clinical trial success rate | 10% | 20-30% (projected) | 2-3x improvement |
| Number of molecules screened | 1-5 million | 10-100 billion (virtual) | 1000x increase |
Data Takeaway: Even modest improvements in clinical success rates translate to billions in savings for large pharma companies. The ability to screen billions of virtual compounds in silico rather than millions physically is a game-changer.
Kohl’s vision extends beyond small-molecule drugs. He discussed applications in protein therapeutics, antibody design, and even gene editing tools. Latent Labs is already working with a major biotech firm to design novel CRISPR-Cas9 variants with improved specificity, using their generative models to propose mutations in the Cas9 protein that reduce off-target effects.
The market response has been enthusiastic. Within 24 hours of the CVPR talk, Latent Labs’ website traffic increased 10x, and the company reported receiving over 200 partnership inquiries from pharma companies and CROs. The stock prices of major CROs like Charles River and WuXi AppTec dipped slightly on the news, reflecting investor concern that generative design could reduce reliance on traditional screening services.
Risks, Limitations & Open Questions
Despite the promise, several challenges remain. First, generative models are only as good as the data they are trained on. Protein-ligand interaction data is sparse and biased toward well-studied targets. Kohl acknowledged that for novel targets (e.g., emerging viral proteins), the models may hallucinate molecules that look plausible but are not synthesizable or stable. Latent Labs addresses this with a synthetic accessibility filter and a retrosynthesis planning module, but these add computational overhead.
Second, the regulatory pathway for AI-designed molecules is unclear. The FDA has not yet issued formal guidance on how to evaluate drugs generated by deep learning models. Kohl noted that Latent Labs is in active discussions with the agency, but until a precedent is set, pharma companies may be hesitant to invest heavily in AI-only pipelines.
Third, there is the risk of overfitting to known chemical space. If the generative model is trained primarily on existing drugs and natural products, it may reproduce known scaffolds rather than truly novel chemotypes. Kohl’s team uses a diversity penalty during training to encourage exploration, but it remains to be seen whether the generated molecules are genuinely novel or just recombinations of existing fragments.
Finally, ethical concerns around dual use: the same generative models could be used to design toxins or bioweapons. Kohl addressed this by stating that Latent Labs has implemented a safety filter that blocks the generation of known toxic or warfare-related compounds, and they are working with the AI Safety Institute to develop industry standards.
AINews Verdict & Predictions
Simon Kohl’s CVPR 2026 keynote was not just a technical presentation—it was a declaration that biology has entered a new era. We believe Latent Labs is well-positioned to lead this transformation, given its strong technical foundation, open-source ethos, and the founder’s Nobel-level credibility. However, the company faces significant execution risk: moving from in silico design to validated clinical candidates is a multi-year, capital-intensive process.
Our predictions:
1. Within two years, Latent Labs will announce its first preclinical candidate for an oncology target, likely an inhibitor of a difficult-to-drug protein like KRAS G12C.
2. By 2028, at least three major pharma companies will have signed multi-year, multi-hundred-million-dollar deals with Latent Labs to access its generative platform.
3. The open-source LatentDiff repository will become the de facto standard for generative molecular design, similar to how AlphaFold became the standard for protein structure prediction.
4. Regulatory frameworks for AI-designed drugs will be established by the FDA by 2029, accelerating adoption.
5. The biggest competitive threat to Latent Labs will not come from other startups, but from large tech companies like Google DeepMind (which could re-enter the space) or Microsoft (through its partnership with 1910 Genetics).
What to watch next: The release of Latent Labs’ first peer-reviewed publication in a top journal like Nature or Science, which will provide independent validation of their in vivo results. Also, watch for hiring announcements—if Latent Labs poaches senior computational chemists from established pharma, it signals confidence in their platform.
Programmable biology is no longer a science fiction concept. Simon Kohl has laid the blueprint, and the industry is now racing to build the future.