# 06 — Neural Biological Simulators ("World Models of Life") That Agent Swarms Wield An AI agent swarm calling `AlphaFold` or `CRISPResso2` in 2024 was limited to folding single proteins or checking guide alignments. Between 2025 and 2026, four monumental **deep-learning foundation models** transformed the *bottom layer* of the stack—giving agent swarms **in-silico simulators** for DNA design, custom CRISPR enzymes, non-coding regulatory mutations, and whole-cell Perturb-seq state transitions. Think of these models as the **"physics engines / unit-test oracles"** that bio agents query *before* spending weeks in a wet lab. --- ## 1. Evo 2: Generative Genome & CRISPR-Operon Foundation Model (*Nature*, 2026) * **Citation:** Brixi, G., Durrant, M. G., Ku, J., ..., Hsu, P. D., & Hie, B. L. *"Genome modelling and design across all domains of life with Evo 2."* ***Nature*** (March 2026), DOI: `10.1038/s41586-026-10176-5` (Preprint: *bioRxiv*, February 2025; building on *Evo 1*, Nguyen et al., *Science*, 2024). * **Institutions:** Arc Institute, Stanford, UC Berkeley, NVIDIA. ### Architecture & Scale (For AI Engineers) * **40B Parameters (`7B` + `40B` variants) on `StripedHyena 2`:** Instead of standard quadratic-cost Attention ($O(N^2)$), Evo 2 uses a hybrid convolutional/state-space architecture (**StripedHyena 2**) trained autoregressively at **single-nucleotide (1-token = 1 DNA base)** resolution. * **1 Million Base-Pair Context Window:** Allows Evo 2 to attend across huge genomic spans—capturing long-range promoter-enhancer switches or holding an entire viral/microbial genome inside one prompt. * **Training Data (`OpenGenome2`):** ~**9 trillion nucleotides** spanning bacteria, archaea, phages, plants, fungi, and animals (including human). ### What It Enables for CRISPR & Synthetic Biology * **Co-Designing Complete CRISPR-Cas Systems:** Earlier protein models designed *either* a protein *or* an RNA. Because Evo operates directly in DNA space across huge operons, it generates **co-adapted Cas proteins + guide RNAs + flanking sequences together** in one autoregressive rollout. * **De Novo Bacteriophages:** Using prompt engineering/fine-tuning around lytic phage templates ($\Phi$X174), the team generated **functional, living bacteriophage genomes from scratch** that replicated in wet-lab bacterial hosts, lysed target bacteria, and even outperformed natural phages in competition assays. * **Zero-Shot Variant Scoring:** Scores coding and non-coding mutations (indels, SNVs) across human genomes without task-specific fine-tuning. --- ## 2. OpenCRISPR-1: First AI-Generated Human Gene Editor (*Nature*, 2025) * **Citation:** Ruffolo, J. A., Nayfach, S., Gallagher, J., ..., Cameron, P., & Madani, A. *"Design of highly functional genome editors by modelling CRISPR–Cas sequences."* ***Nature*** (Published online July 30, 2025), DOI: `10.1038/s41586-025-09298-z`, PMID: `40739342`. * **Company:** Profluent Bio. ### Key Details * **Training Corpus (`CRISPR-Cas Atlas`):** Mined over **26 terabases** of genomic/metagenomic data to curate **1M+ natural CRISPR operons**, training protein language models (ProGen2 family) on the evolutionary grammar of gene editors. * **De Novo Leap:** **OpenCRISPR-1** is a functional Type II Cas9 editor that is **400+ amino-acid mutations away** from standard *S. pyogenes* Cas9 (`SpCas9`) and **nearly 200 mutations away from any known natural protein**. * **Wet-Lab Performance:** Achieved high on-target editing efficiency in human cells with **reduced off-target cutting** compared to wild-type SpCas9, works seamlessly with base-editor fusions, and was released open-source—allowing custom CRISPR IP generation for therapeutics. --- ## 3. AlphaGenome & AlphaGenome 2: Reading the 98.5% "Dark Genome" (*Nature*, 2026) * **Citation:** Avsec, Ž., Latysheva, N., Cheng, J., Novati, G., Kohli, P., Hassabis, D., et al. *"AlphaGenome: advancing regulatory variant effect prediction with a unified DNA sequence model."* ***Nature*** (January 2026), DOI: `10.1038/s41586-025-10014-0` (Preprint: *bioRxiv*, June 2025). * **Institution:** Google DeepMind. ### Biology Context (For AI Engineers) 98.5% of human DNA doesn't encode proteins—it's regulatory logic (like feature flags and config files). Most disease-risk mutations discovered in GWAS studies sit in this non-coding "dark genome," making it notoriously hard for drug target agents to know *which gene* a mutation actually breaks. ### Key Specs & Impact * **1 Megabase Input $\rightarrow$ 1-bp Output Resolution:** Reads 1,000,000 DNA letters and predicts thousands of genomic readout tracks simultaneously across **11 modalities**: gene expression (`RNA-seq`, `CAGE`, `PRO-cap`), chromatin accessibility (`ATAC-seq`, `DNase-seq`), 3D DNA folding contact maps, transcription factor binding / histone marks (`ChIP-seq`), and **RNA splicing** junctions. * **AlphaGenome Atlas (`1 Petabyte`):** DeepMind pre-computed the **AlphaGenome Variant Impact (AVI)** score for **all ~9 billion possible single-letter mutations** in the human reference genome—turning variant triage from expensive GPU inference into an instant database lookup for agent swarms. * **AlphaGenome 2:** Extended to **single-cell resolution** (trained on `100M+` cell profiles), predicting cell-type-specific variant effects (`sc-eQTLs`) that bulk tissue models wash out. --- ## 4. Arc State & Tahoe-100M: The "Virtual Cell" Perturbation Simulator * **Citation:** Adduri, A. K., Roohani, Y., Hsu, P., et al. *"Predicting cellular responses to perturbation across diverse contexts with State."* ***bioRxiv*** (2025–2026) + **Tahoe-100M** (*bioRxiv*, early 2025). * **Institutions:** Arc Institute, Tahoe Therapeutics (formerly Vevo). ### Why It Matters for Functional Genomics & Swarms If a swarm like *PerTurboAgent* or *Virtual Biotech* asks, *"What happens inside a T-cell or lung cancer cell if I CRISPR-knockout Gene X or apply Drug Y?"*, **State** acts as the **neural forward simulator** of cellular behavior. ### Architecture & Scale * **Two-Module Transformer Operating on Sets of Cells:** 1. **`State Embedding (SE)`:** Pretrained on **`167 million` observational cells** to learn clean, batch-corrected representations of baseline cell states. 2. **`State Transition (ST)`:** Self-attention transformer trained on **`100M+` perturbed cells**—heavily anchored by **Tahoe-100M** (mapping 50 cancer models $\times$ ~1,100–1,200 small-molecule drugs) plus Parse cytokine datasets and genome-scale Perturb-seq screens—using an **energy-distance loss** to model population-level cell heterogeneity. * **Benchmarking (`Cell-Eval` & Virtual Cell Challenge):** Improved perturbation effect discrimination by **`+30%` to `+50%`** and achieved **`>2x` accuracy in identifying differentially expressed (DE) genes** compared to existing baselines, generalizing **zero-shot** to unseen cell contexts. --- ## 5. Frontier Lab Ecosystem Moves (2025–2026) * **Google DeepMind `TxGemma` & `Agentic-Tx`:** Open-weight therapeutic LLMs (`2B`, `9B`, `27B` Gemma derivatives trained on Therapeutics Data Commons) predicting binding, ADMET toxicity, and trial outcomes, paired with Gemini-driven `Agentic-Tx` orchestration. * **OpenAI + Retro Biosciences (`GPT-4b micro`):** Domain-adapted GPT-4o protein design model applied to the **Yamanaka cellular-reprogramming factors**—yielding redesigned proteins with **`>50x` higher stem-cell reprogramming marker expression** and improved DNA repair *in vitro*. * **Anthropic (`Claude for Life Sciences`):** Enterprise agent platform integrating Claude directly via **MCP connectors** into standard pharma infrastructure (**Benchling**, **10x Genomics**, **PubMed**, **BioRender**, **Synapse**), alongside partnerships with Bristol Myers Squibb, Generate Biomedicines, and BigHat Biosciences. * **Lila Sciences (Flagship Pioneering):** Pairing foundation reasoning models (NVIDIA Nemotron/BioNeMo) with fully robotic **"AI Science Factories"** (liquid handlers, flow cytometers, plate readers) for autonomous closed-loop physical experimentation. --- ## Synthesis: How All 6 Files Connect When you put the whole 2024–2026 stack together, a clear engineering narrative emerges: 1. **Layer 1 (`Evo 2`, `OpenCRISPR-1`, `AlphaGenome`, `Arc State`)** gives AI systems **high-fidelity neural simulators** of DNA, enzymes, mutations, and cell-state transitions. 2. **Layer 2 (`Biomni`, `OriGene`, `STELLA`)** wraps `150–600+` bio software packages, MCP servers, and genomic databases into executable tool environments. 3. **Layer 3 (`CRISPR-GPT`, `AutoScreen`, `Virtual Biotech`, `AI Co-Scientist`, `Robin/Kosmos`, `PerTurboAgent`)** coordinates swarms of specialized planners, critics, tournament rankers, and code-execution workers to formulate hypotheses, filter out noise, and design optimal wet-lab screens. 4. **Layer 4 (`LabOS`, `Lila`, human organoids, Perturb-seq)** connects back to **the wet-lab verifier**—the indispensable, slow, physical ground truth where cells still take 48–72 hours to respond.