---
title: Generative Biology
type: vocabulary
url: "https://www.envisioning.com/vocab/generative-biology"
summary: AI-driven generative design of proteins, genomes, and biological systems.
year: 2023
generality: 0.62
---

# Generative Biology

AI-driven generative design of proteins, genomes, and biological systems.
Generative biology is the application of generative modeling techniques — diffusion models, language models, flow-matching, and increasingly end-to-end learned simulators — to the design of biological systems: proteins, RNA structures, regulatory elements, viral capsids, and increasingly whole genomes and cellular subsystems. The term became widely used in 2023 as laboratory groups and companies demonstrated that AI-generated protein designs could be synthesized and shown to fold and function as specified, marking a transition from AI as a tool for analyzing existing biology to AI as a tool for producing new biology. Generative biology is practiced by academic groups including David Baker's laboratory and the Arc Institute, and by companies such as EvolutionaryScale, Profluent, Generate Biomedicines, Cradle, and Cradle Biosciences, each of which has developed proprietary or open generative models targeted at different biological modalities.

The mechanism, as practiced, follows a structure common to generative AI more broadly. A model is trained on a corpus of biological sequences and, increasingly, on paired sequence-and-function data, optionally with structural or fitness annotations. Once trained, the model can be sampled to produce novel sequences, conditioned on a desired property or function. The major shift in the 2020s has been from models that predict a property given a sequence to models that generate a sequence given a property. Protein-design models like RFDiffusion and Chroma demonstrated this inversion for individual proteins, and genome-level models like Evo 1 and 2 extended it to much longer sequences of DNA. The wet-lab half of the discipline has converged around synthetize-test-learn cycles: take a model's output, have it manufactured and assayed, feed the resulting data back as additional training signal. This mirrors the way large language models improve with feedback but is constrained by the cost and latency of biological measurement.

The tradeoffs with traditional structure-based or directed-evolution approaches to designing biological molecules are substantial on both speed and accessibility. Generative biology can produce thousands of plausible designs in minutes where rational design typically produces one or two, and where directed evolution requires cycling through many generations of selection. The approach inherits the failure modes of all learned generative models: outputs that look plausible but are not actually functional, difficulty with out-of-distribution design targets, and the challenge of evaluating the small fraction of functional designs among the large number generated. The field has also raised biosecurity concerns distinct from earlier synthetic biology debates because the design automation lowers the skill barrier for producing functional biological constructs that previously required substantial wet-lab expertise to produce.

Open questions include how to evaluate generative biology models well — standard benchmarks for language and image generation exist, but the corresponding ground-truth standards for protein, RNA, and genome function are still being established — and how the field will handle dual-use concerns as model capabilities improve. Whether generative biology will eventually produce truly novel biological functions that have no analog in natural biology, or whether its outputs will remain broadly within the envelope of what evolution has already explored, is a central open question for both scientific and governance reasons. Connections to other watch-list items — autonomous wet labs, biological foundation models, programmed life forms — are tight: generative biology is the design step in an emerging pipeline that, once coupled to autonomous labs, may approach the kind of self-sustaining cycle that some AI forecasts describe.

---
Source: Envisioning — Technology Research Institute (https://www.envisioning.com/vocab/generative-biology)