ML maturity ladder for biologics discovery and development
Many organizations confine machine learning in molecular discovery and design to isolated proof-of-concept projects. We propose a 5-level maturity ladder to illustrate how teams advance towards autonomous systems that learn from every experiment and guide discovery across entire pipelines.
Stef van Grieken · 3 min read
Machine learning (ML) is reshaping how biologics are engineered, accelerating discovery cycles, enabling smarter decisions, and opening new paths through an unimaginably vast design space. But the pace and breadth of change can feel overwhelming. What does “good” practice even look like for a discovery and development organization today? What might it look like tomorrow?
The promise of ML in biologics R&D is to deliver safe, effective therapeutics in less time, but achieving that potential will require clarity on how maturity is defined and measured.
In our work with leading pharmaceutical and industrial biotech organizations, we’ve seen a clear need for a shared language to describe ML maturity, that helps teams and executives assess where they stand and how they can advance.
So we’re proposing a straightforward “maturity ladder” for biologics discovery and development. Inspired by the self-driving car industry’s SAE levels – which range from purely manual driving to full self-driving capability – our ladder charts progress as a climb toward greater autonomy, enabled by smarter sequence generation and faster learning from increasingly automated experimental cycles.
Our aim is to provide a simple reference that leaders can use to benchmark progress, align strategy, and guide investment. More importantly, we hope it sparks discussion across the R&D community about how the field can mature together.
Just as the automotive industry defines levels of self-driving capability, from manual control to high automation, biologics R&D is seeing its own progression in how machine learning contributes to discovery.
Level 0: No machine learning used
Akin to SAE Level 0: full manual control, where the human does all the driving and the system offersano automation.
Traditional protein engineering that relies on semi-rational or rational approaches. This might involve introducing mutations, observing their effects, and combining beneficial changes, or applying expert knowledge and bioinformatics to guide where and how to make changes to sequences.
Level 1: Human assessment
Akin to SAE Level 1: driver-assist features like lane-keeping or cruise control – the system provides guidance, but the human remains fully responsible for every action.
Tools such as AlphaFold are used to help scientists understand the likely structure or properties of a protein, but decisions about what to design or test still rest entirely with humans.
Level 2: Ranking and selection
Akin to SAE Level 2: partial automation, where the system can control both speed and steering under human supervision – ML handles defined tasks while humans oversee and decide what to test next.
Individual computational scientists begin developing models. ML models play a more active role, predicting how well different known candidate proteins might perform on single properties, such as binding affinity, aggregation or immunogenicity. Each sequence is given a predicted-performance score for ranking, which scientists use to build heuristics for selecting the most promising candidates to test in the lab.
Level 3: Ranking and selection
Akin to SAE Level 3: conditional automation – the system can drive itself under limited conditions but still expects the human to monitor and take over when needed.
At this stage, organizations either begin developing their own enterprise-grade ML platforms or adopt existing ones to support scalable discovery workflows. ML models go beyond evaluating existing sequences, generating entirely new candidates for ranking and testing. This can be done without project-specific training data (i.e. zero-shot) or using prior assay results from hit-identification or lead optimization.
At the top of this level, the ML models consider multiple goals at once, such as binding strength, stability, toxicity, and manufacturability, and down-select sequences that balance those trade-offs. The software becomes part of an active learning system: it proposes new designs, selects the most promising set of candidates for testing, receives real-world experimental feedback, and learns from the results, closing the loop between ML systems and the wet lab.
Beyond the learning loop for individual programs, deeper organizational embedding of ML systems bring learning and accelerated progress across programs and departments as programs leverage each other's data. Scientists still decide what experiments to run, which programs to continue, and what additional assays to develop.
Level 4: Agents and Decision Making
Akin to SAE Levels 4 and 5: largely autonomous operation within defined limits, with optional human oversight.
At the final level the software goes beyond designing sequence libraries and deciding on experimental setup, taking on a more agentic role, meaning it can act on its own analyses to make defined decisions. This can include choosing which programs to advance based on target product profiles, allocating experimental bandwidth, or refining assays to improve understanding of the target. The level of human-in-the-loop can vary by organization and program, but at this stage, required coordination is minimal and you have a fully integrated and automated system for engineering biomolecules.
What it really takes to ladder up
For autonomous vehicles, the highest level of maturity is a car that can drive itself anywhere, in any conditions, without human intervention. In biologics, we see the equivalent as a discovery system that can design new proteins, select the best candidates, trigger lab experiments, learn from the results, and continually improve with minimal human coordination. At that point, machine learning becomes more than a tool. It becomes a strategic guide for scientific discovery.
Reaching that level of autonomy depends not only on building smarter models, but on how effectively teams connect them to experimental science and structure their organizations around those feedback loops. In a future companion piece, we’ll explore the three foundational strands that make this climb possible – scope of ML application, data practices, and organizational integration – and show how the real race to ladder up in biologics is to advance all of them together.
Cradle AI-readiness webinar
Rewatch the Cradle Webinar with CEO Stef van Grieken on AI readiness for drug discovery.
Recent posts
Subscribe and get new posts and updates from Cradle straight to your inbox.
