From hypothesis to validation: AI translation for peptide order processing in under 60 days

Custom peptide orders rarely arrive in neat, standardized formats. They come through emails, written in the language of scientists, labs, and institutions, each with their own abbreviations, conventions, and assumptions.

For one specialty peptide manufacturer, that variability created a costly bottleneck. Every order had to be manually interpreted by an internal expert before it could be translated into the company’s proprietary peptide code and used for cost calculation.

In under 60 days, we helped validate an AI translation approach that could convert free-form peptide order requests into structured internal codes, reducing the need for manual interpretation and proving the viability of domain-specific AI for scientific order processing.

The translation gap in peptide manufacturing

Custom peptide synthesis depends on precision, but the language used to order peptides is not governed by a single, universally enforced standard. Across academic institutions, pharmaceutical companies, and research laboratories, customers use different abbreviations, naming conventions, annotations, and sequence formats to describe the same scientific intent.

This makes peptide order intake difficult to standardize. The challenge is not that customers provide incomplete information. It is that valid requests can arrive in many different forms, each shaped by the conventions of the sender, the lab, or the institution.

When every peptide quote depends on manual translation 

For the manufacturer, the lack of standardized peptide notation created a critical dependency in the quoting process. Before an order could be costed, an internal specialist had to read the customer’s request, identify the intended peptide sequence, interpret any contextual instructions, and translate the result into the company’s proprietary peptide code nomenclature.

This was not just a data entry task. A request such as “please synthesize the following formula, where P stands for...” requires understanding the sender’s implied convention before the sequence can be translated accurately. Without that contextual reading, automation attempts struggled to produce reliable results.

At scale, this manual translation step slowed response times, increased the risk of transcription errors, and limited how many quote requests the team could process efficiently.

“The real bottleneck was not a lack of automation tools. It was the absence of a system that could interpret scientific intent from unstructured natural language, the way a trained chemist would.”

Thibault Viglino, Software Architect @ Visium

The client wanted to explore whether AI could support this translation step and reduce the dependency on manual interpretation. The goal was not to launch a full production system immediately, but to validate a clear hypothesis: could a domain-specific AI model understand real peptide order language and convert it into the structured internal codes required for quoting? 

From manual translation to AI-powered peptide code conversion 

We approached the problem as a specialist translation challenge. Instead of translating between human languages, the AI model was trained to translate between two scientific “languages”: the varied, informal language used in customer order emails and the manufacturer’s structured proprietary peptide code.

To do this, we adapted a sequence-to-sequence translation architecture for peptide nomenclature conversion. The model learned from historical order data, using past customer emails and the corresponding internal codes that specialists had already produced manually.

This training data allowed the system to learn more than simple format conversion. It captured recurring patterns in how different customers, laboratories, and institutions described peptide sequences, including shorthand, annotations, and context-dependent conventions.

Within approximately two months of project initiation, the model produced results strong enough to validate the approach. In test examples, it successfully converted complex, informally written peptide descriptions into structured internal codes that could be used for downstream cost calculation.

Demonstrating AI translation from unstructured peptide requests to cost-ready codes 

The proof of concept showed that AI could interpret the variability of real-world peptide orders and convert them into the manufacturer’s structured internal code with minimal expert intervention at that step.

This mattered because the model was not only handling clean, predefined inputs. It was working with the complexity of actual customer emails, including informal descriptions, context-dependent notation, and sender-specific conventions. In doing so, it demonstrated that machine learning could support a task that had previously depended on specialist scientific judgment.

Within the agreed timeline, the engagement answered its core question: could a domain-specific AI model translate unstructured peptide order language into a format ready for downstream cost calculation? The validated results showed that it could.

Moving into production would require further work on data maturity and process standardization on the client side, but the project clearly established the technical viability of the approach and the speed at which a focused AI solution can move from concept to validated output.

‍