Beyond Prediction: Why Protein Design Needs a Decision Layer

The release of AlphaFold 2 in 2021 reset computational biology. We moved rapidly from struggling to predict basic protein structures to generating entirely non-natural sequences de novo. What followed was a flood of specialised models for structure prediction, enzyme engineering, antibody design, and developability assessment. Keeping up with them quickly became a job in itself. For modern computational protein R&D, prediction and generation are no longer the hard part. Deciding what to do with everything those models produce is.

The Hidden Tax on Every Protein Researcher

Before a biologist can chain tools into anything resembling a real workflow, they have to become a software engineer.

New state-of-the-art models appear on GitHub faster than most teams can evaluate them. Each one comes with its own installation instructions, dependency stack, and version requirements that frequently conflict with what is already on the system. Getting a new model up and running can consume days of troubleshooting before any real research has begun. And all of this assumes the researcher has access to a GPU capable of running inference at a useful scale in the first place. The friction does not end with installation, either. Even a single workhorse tool can demand a serious learning curve: mmseqs2, the de facto standard for fast sequence search and clustering, ships with dozens of subcommands and hundreds of tunable parameters. Reading its manual closely enough to use it well can take several days before you have searched a single sequence.

Even setting up access aside, the landscape itself is overwhelming. A researcher trying to design an antibody (or engineer a more thermostable enzyme, annotate an uncharacterised proteome, or rank candidate biomarkers) faces dozens of competing model choices across every sub-task: structure prediction, inverse folding, binding affinity estimation, immunogenicity screening, stability assessment. Each model has its own interface, its own parameter conventions, its own output format. Every update risks breaking the workflow you spent days building. Papers recommend one model; practitioners in forums recommend another. The “best” tool for a given task changes every few months.

Picking the model is a rabbit hole of its own. Take protein language models: ESM2 alone comes in flavours spanning 8 million to 15 billion parameters, with Ankh, ProtT5, and a steady stream of newcomers alongside it. Bigger is not automatically better. The right embedding model depends on your task, your sequences, and, frankly, the GPU you can actually get time on. Choosing well is a research question, and the answer keeps moving.

The result is that a significant portion of a computational biologist’s time that should go toward scientific reasoning is consumed by infrastructure management, version pinning, environment debugging, and format conversion. This is not a skills gap. It is a structural tax levied on everyone who tries to work at the frontier of protein AI.

The Illusion of Progress in the Dry Lab

Even for teams with dedicated engineering support, the fragmentation problem runs deeper than tooling overhead.

It is now computationally inexpensive to generate thousands of novel protein sequences and corresponding predicted structures. But PDB and FASTA files were never the point. The goal of biotechnology is a functional, stable, synthesisable molecule.

When computational tools exist in silos - generative algorithms over here, sequence analysers over there, researchers are forced to manually piece together fragmented data. This lack of integrated context means critical decisions about which specific variants to advance are often based on incomplete evidence, leading to costly late-stage failures. Generating a million candidate sequences (binders, enzyme variants, redesigned scaffolds) is cheap; turning them into a decision requires a rigorous, systematic method to find the few that will actually express, fold, and function.

There is also a reproducibility crisis hiding in plain sight. Most computational protein workflows live in ad-hoc scripts, undocumented notebooks, and pipelines that exist only on a single researcher’s laptop. When a project is handed off, or when a key person leaves, the institutional knowledge embedded in those workflows disappears with them. Results cannot be audited, extended, or built upon. Teams routinely restart from scratch because they cannot confidently trace how a previous result was produced.

What Is a Decision Layer?

Addressing this bottleneck requires rethinking how computational tools are deployed. Instead of treating individual ML models as standalone utilities, the R&D pipeline needs an agentic “decision layer” to orchestrate them.

A decision layer doesn’t just run a prediction; it orchestrates the R&D loop from question to candidate. Agentic AI makes this possible. It chains the necessary tasks together, working out which tools are relevant and in what order they run, managing the handoffs between them, and aggregating scores and signals across the pipeline into a ranked, evidence-backed shortlist for wet-lab validation. Critically, every step remains fully traceable, so the reasoning behind any candidate decision can be interrogated, revisited, and shared.

Enter ProtXplain: Orchestrating the R&D Loop

This is exactly what we built ProtXplain to be. Designed as an agentic integration and decision platform, ProtXplain directly addresses pipeline fragmentation by unifying discovery, design, and optimisation into a cohesive decision layer.

ProtXplain does this through two integrated systems. ProtXplore is a conversational research agent where scientists ask questions in natural language and get back structured, citation-grounded hypotheses rather than raw database outputs. ProtFlow is the execution layer with a visual workflow engine that turns those hypotheses into reproducible, auditable pipelines connecting the tools teams already rely on, without requiring a single line of code. It spans the full range of protein work: homology search and clustering (mmseqs2, Foldseek), structure prediction, functional annotation, enzyme engineering, biomarker discovery, and, yes, antibody and binder design too.

The Future Belongs to Decision-Makers

The open-source protein AI ecosystem is not going to slow down. New models will keep emerging, and the pressure on research teams to evaluate, adopt, and integrate will keep growing. As the tooling landscape grows more complex, that decision-making infrastructure stops being optional. ProtXplain is built to be that infrastructure. It is also meant to grow with the science. The decision layer that helps you choose between today’s models is being extended to benchmark protein language models head-to-head on your own data, and before long to train and fine-tune your own. Soon the harder question won’t be which tool to trust, but which one to build.