AI-Generated Biology Has a Provenance Problem: What Biotech Engineering Leaders Need to Know About Synthetic Data Integrity

The protein databases that underpin modern drug discovery and synthetic biology research were built on the assumption that their contents reflect observed reality. That assumption is increasingly strained. As generative AI tools produce millions of novel protein structures, sequences, and molecular designs at a pace no experimental validation pipeline can match, the question of whether a given biological record is computationally predicted, experimentally confirmed, or entirely synthetic has become a material engineering concern, not a philosophical one.

The Database Contamination Problem Is Already Here

Public repositories like the Protein Data Bank and UniProt have long served as ground truth for structural biology. When AI-generated structures enter these databases without clear provenance markers, downstream models trained on that data inherit the uncertainty. A model trained on a mix of experimentally resolved and AI-predicted structures will behave differently from one trained on verified data alone, and the difference may not surface until a compound reaches a wet-lab validation stage or, in worse cases, a regulatory submission.

The mechanism is straightforward: generative protein models can produce plausible-looking outputs that satisfy structural energy constraints without corresponding to anything that exists or could be synthesised safely. If those outputs are deposited into shared databases without adequate labelling, they become training data for the next generation of models. The contamination compounds silently.

This is not a theoretical risk. AlphaFold's release into the European Bioinformatics Institute's database added over 200 million predicted structures. The scientific value is undeniable. The provenance challenge is equally real, because predicted structures carry a fundamentally different epistemic status than crystallography or cryo-EM resolved entries, and that distinction is not always preserved as data moves between systems.

What Watermarking Approaches Actually Offer

Google DeepMind's SynthID Bio represents one of the more technically grounded attempts to address this problem. The approach embeds a statistical signal into AI-generated biological sequences, specifically protein sequences in the initial implementation, that persists through typical downstream processing and can be detected without access to the original model. The watermark is designed to be invisible to standard bioinformatic analysis while remaining recoverable by a detection algorithm.

The Signal Persistence Challenge

The core technical difficulty with biological watermarking is that biological sequences are not static artefacts. They get truncated, mutated, aligned, and recombined as researchers work with them. A watermarking scheme that degrades under standard sequence manipulation provides weaker guarantees than one that survives those operations. SynthID Bio's design targets this directly by embedding the signal in a way that is statistically recoverable even from partial sequences, though the robustness of that guarantee across all realistic manipulation scenarios remains an active area of evaluation.

What Watermarking Cannot Do

Watermarking addresses the detection problem for sequences generated by a specific model using a specific scheme. It does not address sequences generated by models that do not implement watermarking, which is currently the majority of deployed protein design tools. It also does not retroactively label the AI-generated content already present in public databases. Engineering teams should treat watermarking as one layer of a provenance architecture, not as a complete solution.

DNA Synthesis Screening Has a Structural Gap

The biosecurity dimension of this problem sits at the synthesis layer. DNA synthesis providers are expected to screen orders against databases of sequences of concern, including known pathogen genomes and toxin-coding sequences. The screening systems were designed for a world where sequences of concern are known and catalogued. Generative biology introduces a different threat model: novel sequences that do not match any existing entry but that encode functional properties that would be flagged if the screening system understood function rather than identity.

Current screening approaches are predominantly sequence-similarity based. A sufficiently divergent AI-designed sequence encoding a dangerous protein function may not trigger a match. The gap is not a failure of the screening vendors specifically; it reflects the fact that the threat model has changed faster than the regulatory and technical infrastructure designed to address it.

Engineering teams building synthesis workflows on top of AI-designed sequences should not assume that standard synthesis provider screening provides equivalent protection to what it offered five years ago. The risk surface has expanded, and the screening tools have not yet caught up.

What Data Lineage Actually Requires in a Biological AI Stack

Provenance in a biological AI context is not simply a matter of recording where data came from. It requires capturing what transformations were applied, which model version generated a given output, what confidence or uncertainty estimates accompanied that output, and whether the output has been experimentally validated at any stage. This is a data engineering problem as much as a biology problem.

Minimum Viable Provenance for Production Workflows

For engineering teams integrating biological AI outputs into research or commercial pipelines, a minimum viable provenance architecture should track the following at the record level:

• Generation source: model name, version, and whether the output is predicted, designed, or experimentally derived
• Validation status: whether any experimental confirmation exists and at what resolution or assay type
• Transformation history: any downstream processing steps that may have altered the original output
• Screening status: whether the sequence has been passed through biosafety screening and which screening version was used

Without this information at the record level, data governance becomes a post-hoc exercise rather than a structural property of the pipeline.

The Regulatory Exposure

Regulatory frameworks for AI-generated biological content are still forming, but the direction of travel is clear. Agencies are moving toward requiring explicit documentation of training data provenance for AI systems used in drug development. Engineering teams that build pipelines now without provenance infrastructure will face a retrofit problem later, and retrofitting data lineage into a production biological AI system is significantly more expensive than building it in from the start.