Inspiration
Antibody design workflows often require multiple specialized tools, manual file conversion, and difficult-to-interpret outputs. We wanted to make the computational path from antibody CDR inputs to a reviewable candidate report more transparent, reproducible, and easier to demonstrate.
What it does
Antibody Labmate is a computational antibody-design workflow under active development.
Users provide six IMGT-defined antibody CDR sequences and a replaceable protein antigen structure. The workflow validates inputs, generates candidate VH/VL sequences, predicts candidate structures, runs an authorized protein-protein docking backend, analyzes interface residues, and produces a ranked candidate table plus an offline HTML report.
The product has three execution modes:
- Replay: a stable, clearly labeled replay of verified fixtures for demos and testing.
- Live Local: real computation on the user's own machine after tool and resource checks pass.
- Live Remote: real computation through an authorized remote worker using the same artifact schema.
Replay never presents stored outputs as new calculations for different inputs. Live capabilities remain unavailable until their dependencies, licenses, and end-to-end checks have passed.
How we built it
We designed the project around a shared artifact contract rather than a single monolithic script. The workflow records normalized inputs, hashes, tool versions, model versions, parameters, chain mappings, stage status, and output artifacts in a manifest.
The planned stack includes Python, Pydantic data models, Streamlit for the interface, BioPython-based structural parsing, Jinja2 for offline HTML reporting, and optional open-source PyMOL rendering. Candidate generation, folding, docking, and visualization are implemented as replaceable adapters so that the same workflow can run in Replay, Local, and Remote modes.
Challenges we ran into
The main challenge is scientific and technical honesty. Six CDR sequences alone are not a complete antibody structure, so the CDR-only generation path requires a valid VH/VL carrier and must be experimentally verified against the selected generation tool.
We also needed to separate local computation from truly offline computation: local folding may still use an external MSA service unless models and databases have been prepared locally. Docking software licensing and redistribution restrictions also affect which backend can be used in a public or remote deployment.
Finally, docking scores and predicted structures are computational prioritization signals, not proof of binding affinity, specificity, safety, or therapeutic value.
What we learned
We learned that reproducibility is not only about rerunning a command. It also requires explicit input hashes, chain mappings, versions, licensing status, failure states, and an honest distinction between replayed, cached, and live-computed results.
We also learned to design a stable demo path without disguising it as real-time scientific computation.
What's next for Antibody Labmate
Next, we will finish the Replay MVP, including input validation, ranking, interface residue analysis, and offline report generation. We will then validate the CDR-only generation adapter, local structure prediction, and a license-compliant docking provider before enabling Live Local and Live Remote modes.
Log in or sign up for Devpost to join the conversation.