BioTech IP Review

Artificial intelligence has recently crossed a threshold in protein science: from predicting protein structure to designing entirely novel proteins with specified functions. Using deep learning models trained on massive protein sequence and structure datasets, researchers can now generate amino acid sequences that fold into predetermined three-dimensional shapes or catalyze desired chemical reactions. This development reframes proteins not merely as discovered biological entities, but as algorithmically engineered molecular machines, with significant implications for patent law, inventorship, and biotechnology regulation.

From Structure Prediction to De Novo Design

Early breakthroughs in protein AI focused on structure prediction, most notably deep neural networks capable of inferring a protein’s three-dimensional conformation from its amino acid sequence. Subsequent iterations expanded these capabilities into generative design. Instead of mapping sequence to structure, modern models invert the problem: given a target fold or biochemical function, they propose sequences predicted to realize that goal.

Technically, these systems employ architectures such as diffusion models, variational autoencoders, and transformer-based language models trained on curated protein databases. By learning statistical regularities governing folding and interaction, the models can explore regions of sequence space not sampled by evolution. The resulting proteins often lack close natural analogs, yet exhibit stability and function when synthesized and tested experimentally.

This inversion of the traditional design paradigm—specifying function first and deriving structure and sequence computationally—marks a departure from directed evolution and rational mutagenesis, which modify existing proteins. AI design instead treats proteins as outputs of an optimization problem constrained by physical chemistry.

Functional Scope and Applications

AI-designed proteins are being developed for diverse applications, including:

1. Therapeutic binding proteins, engineered to bind disease-associated targets with antibody-like specificity.
2. Enzymes, designed to catalyze reactions not known in natural metabolism.
3. Structural proteins, forming nanomaterials with programmable geometry.

Unlike antibodies or naturally occurring enzymes, which are products of evolutionary selection, these proteins are defined by computational objectives. Their novelty lies not merely in sequence divergence but in functional intentionality.

From a regulatory standpoint, such proteins may be classified as biologics or as novel chemical entities depending on their size and intended use. However, their origin in silico raises questions as to whether existing regulatory categories sufficiently capture their design process.

Patent Eligibility and Inventorship

The patentability of AI-designed proteins raises issues distinct from those associated with naturally derived biomolecules. Courts have historically excluded naturally occurring DNA and proteins from patent eligibility when claimed in isolation. However, proteins whose sequences are not found in nature and are created by algorithmic design may fall outside this exclusion.

The central legal inquiry shifts from whether the molecule exists in nature to whether it is the product of human ingenuity. In AI protein design, human contribution lies in defining objectives, curating training data, and validating outputs experimentally. The algorithm proposes candidate sequences, but selection and functional characterization remain human-mediated steps.

Inventorship analysis must therefore consider whether the AI system is merely a sophisticated tool or whether it substantively determines the inventive features. While current doctrine requires a human inventor, disputes may arise over whether specifying a desired protein function suffices as conception of the claimed sequence.

Claim Scope and Enablement

Drafting claims for AI-designed proteins presents challenges in both breadth and support. A single computational run may yield thousands of candidate sequences predicted to share a function. Attempting to claim this entire class risks running afoul of written description requirements, particularly where only a small subset is experimentally validated.

Claim strategies may include:

• Sequence-based claims to specific proteins tested in vitro or in vivo.
• Genus claims defined by structural motifs or functional assays.
• Method claims covering computational design workflows coupled with experimental validation.

Enablement requires that the patent specification teach how to make and use the claimed proteins without undue experimentation. Where the invention lies in an algorithmically defined sequence space, the sufficiency of disclosure may hinge on whether the model architecture and training parameters are described in enough detail to permit reproduction.

Relationship to Prior Art

AI-designed proteins blur traditional distinctions between discovery and invention. While they may resemble naturally occurring folds, their sequences may be statistically optimized rather than evolutionarily derived. Prior art searches must therefore address both known protein sequences and disclosed computational design methods.

This duality raises complex novelty questions. If a generated protein is structurally similar to a known protein but differs significantly in sequence, is it anticipated? Conversely, if a known protein shares a function but not a fold, does it render the AI-designed protein obvious?

These issues will likely require courts to grapple with what constitutes a “protein” as a claimed subject: its sequence, its structure, or its function.

Commercial and Competitive Implications

Companies developing AI protein design platforms increasingly emphasize their proprietary models rather than individual protein products. This shifts value from discrete compositions to algorithmic pipelines, echoing trends in synthetic biology and drug discovery software.

As a result, trade secret protection may become as important as patenting, particularly for model architectures and training datasets. Conversely, proteins intended for therapeutic use may still require composition-of-matter patents to justify development costs.

Licensing disputes may also arise where training datasets include patented protein sequences or structures, implicating questions of permissible use and derivative invention.

Conclusion

AI-designed proteins represent a new class of biologic invention, characterized less by evolutionary history than by computational intent. By transforming protein creation into an algorithmic exercise, they challenge established frameworks for patent eligibility, inventorship, and claim construction. These developments foreshadow a convergence of biotechnology and software law. As proteins become outputs of machine learning systems rather than products of natural selection, the legal system will be required to reassess what it means to invent a molecule. The resolution of these questions will shape not only intellectual property doctrine but the commercial architecture of next-generation biologics.