Making the research cycle learn
Every protein-design campaign produces two kinds of knowledge: evidence about the molecules being tested, and evidence about the decisions that led to those experiments. The second kind is easy to lose. Why did we choose this library? Which change helped the assay? What did a failed experiment reveal about the model?
Aureka's ambition is to make both kinds of knowledge productive. We are developing a system that uses biological feedback to improve its protein design models, its experimental systems, and the specialized LLM that helps develop both. Recursive self-improvement, or RSI, becomes a concrete research question: can experience make the next generation better at building the generation after it?
A thought experiment captures the ambition. Imagine 100 PhDs spending ten years and 1,000 iterations moving a conceptual score from 60 to 70. Now imagine an agent moving from 60 to 80 in one month and ten iterations. These are illustrative aspirations, not a benchmark or forecast. The score has no biological unit, and the iterations are conceptual research updates, not compressed biological cycles. The question is whether learning can increase the progress made with each research decision.
100 PhDs
Researcher-led improvement
- Time horizon
- 10 years
- Iterations
- 1,000
Specialized LLM (agent)
With researchers and lab infrastructure
- Time horizon
- 1 month
- Iterations
- 10
Different time horizons; conceptual scores, not LLM evaluation results. “Iterations” means envisioned R&D updates, not serial full biological cycles.
View the detailed figure ↗
Three capabilities that improve together
The architecture connects three objects of improvement. The protein design model generates and evaluates sequences. The experimental system determines what we can build, measure and learn. The specialized LLM diagnoses limitations, develops interventions and learns from their consequences.
The LLM's role extends inside the protein designer. It can propose changes to training data, objectives, optimization and sampling, then use experimental outcomes to evaluate the resulting model. Architecture changes are another possible development direction. The important distinction is that the mechanism producing the designs can itself change.
The same reasoning applies to biology. The LLM can develop yeast and plasmid changes, screening strategies, gating choices and inference updates. A better designer can propose more informative libraries; a better experiment can make model failures easier to detect. Their learning signals reinforce each other.
Better intelligence → better experiments → better data → better intelligence.
This also changes what counts as a good library. Some libraries should maximize promising binders. Others should explore uncertainty. An exploratory batch can have a lower immediate hit rate yet help train a better next-generation model.
Specialized LLM
Reason, plan, and learn through post-training.
Directly improves both capabilities below.
Build a better designer
Training data · objectives
Model updates · sampling
Build a better experiment
Yeast / plasmids · FACS / gating
NGS · measurement strategy
High-throughput biological feedback
Molecular measurements improve models and experiments.
Decision–outcome experience feeds LLM post-training.
↺ Experimental evidence and decision experience return to the LLM.
The LLM improves both the protein designer and experimental strategies. The designer and experimental system inform each other. Their biological measurements and decision outcomes generate feedback that improves models and experiments and trains the LLM, closing the loop.
View the detailed figure ↗
A high-bandwidth connection to biology
For protein-model developers, a useful learning environment must return more than a list of winners. It should reveal where predictions fail, which regions remain uncertain, and whether apparent binding failures are actually expression or measurement failures.
Our experimental paradigm connects target-specific libraries to yeast display, FACS and gating, next-generation sequencing, and modeling. Together, these steps produce sequence-linked readouts and inferred pseudo-KD values. Yeast and plasmid engineering belong inside the learning loop because they influence which designs can be observed and how reliably.
A limited Biacore SPR program provides reference measurements for calibration and independent evaluation. SPR helps assess the screening-derived estimates; it does not need to measure every library member. Keeping calibration separate from evaluation matters: a system should not receive credit for agreeing with the same references it repeatedly used to tune itself.
“High bandwidth” means useful feedback delivered to a learning decision. Sequencing reads, detected variants, usable affinity estimates and validated binders are different units. Uncertainty, missing observations and out-of-range measurements remain part of the record. More data becomes valuable when the system can understand what those data actually say.
- 01Design the libraryTarget-specific candidates and informative probes
- 02Display on yeastEngineer the host, plasmid and expression system
- 03FACS + gatingSelect populations and retain measurement structure
- 04SequenceLink molecular identities to screening readouts
- 05Infer pseudo-KDModel, calibrate and qualify usable feedback
View the detailed figure ↗
Build a better protein designer
Consider an illustrative scenario with synthetic outcomes. An LLM develops a designer update that combines training-data reweighting, a revised ranking loss and uncertainty-guided sampling. The original and updated models each submit 10,000 designs for three held-out targets. The evaluation assay stays fixed so that changing the measurement process cannot explain the model comparison.
Qualified screening hits increase from 720 of 30,000 designs to 2,040 of 30,000: 2.4% to 6.8%. A qualified hit here means pseudo-KD at or below 100 nM plus expression and quality-control criteria. The denominator includes every submitted design.
Prediction ranking against a separate SPR reference panel improves from a Spearman correlation of 0.41 to 0.66. Among the top 24 candidates selected from each model for SPR, the best affinity changes from 12 nM to 2.8 nM. That selected subset does not independently validate the full-library screening hit rate, and stronger affinity alone does not establish better developability.
The point is to connect a change inside the designer to better biological outputs and predictive understanding. Finding a stronger molecule and learning how to build a stronger design model are related, separately testable achievements.
Make more of biology measurable
A second illustrative scenario isolates experimental improvement. Both workflow versions screen the same 1.25-million-variant repertoire. Yeast/plasmid expression optimization, gating changes and inference updates increase accepted coverage from 48% to 80%: 600,000 usable records become one million.
The feedback also becomes more reliable. On the same 120 independent SPR evaluation identities, mean absolute log-affinity error decreases from 0.42 to 0.16. Agreement within twofold increases from 47.5% to 85.0%. Expression-positive viable singlets rise from 62% to 89%, while normalized median display rises separately from 1.0 to 1.8.
Time from an induced, display-ready library to quality-controlled data release falls from 21 to 7 days. This is the screening interval; model training, library construction and other development work need their own accounting.
These endpoints matter for different reasons. Coverage determines how much of the library teaches us something. Agreement determines how much we can trust it. Turnaround determines how soon we can act. Better expression improves observability, without necessarily changing a molecule's affinity. Together, they determine the learning value of an experiment.
Let consequences teach the LLM
The third illustrative scenario asks whether the LLM becomes better at improving the other two systems. Molecular labels alone cannot answer this. The learning record must connect the available evidence, a concise rationale, the intervention, its outcome and its resource cost. Failed interventions are useful experience too.
In this synthetic comparison, post-training uses experience from development campaigns. Evaluation covers 48 tasks within three held-out target campaigns. Success means improving the assigned endpoint within its resource and quality constraints.
The starting LLM succeeds on 41.7% of tasks. Giving it retrieval access to the accumulated experience raises success to 52.1%. The post-trained LLM reaches 68.8%. The latter comparison is especially useful: retrieval and post-training have access to the same development experience under comparable evidence budgets, helping distinguish learning from simply seeing more information.
A further test starts the LLM versions from the same designer, experimental workflow and resources, then evaluates the improvements each produces. That prevents a stronger inherited protein model from being mistaken for a smarter LLM. With tasks nested within only three campaigns, these illustrative fractions support no statistical-significance or general-intelligence claim.
The economics of the next generation
The illustrative acquisition economics make repeated learning cycles easier to imagine. At an assumed $100 per individual SPR datum, one million measurements cost $100 million. The updated screening scenario yields one million accepted records for $10,000, or one cent each: a nominal 10,000-fold acquisition-cost difference.
Nominal 10,000-fold acquisition-cost difference. These records contain different information; the cost boundary and exclusions are stated below.
The comparison does not imply equivalent information or a 10,000-fold improvement in speed. Screening-derived pseudo-KD and individual SPR measurements serve different roles. The screening figure also excludes library synthesis, protein production and SPR, model training, capital expenditure and overheads. The opportunity is to make useful experimental feedback frequent enough to support model and LLM development.
Aureka has a proof of concept and has completed LLM post-training and a before/after comparison. The numbers, cohorts and controls described here remain illustrative; they are not the measured results of that work. The empirical case depends on the actual interventions, comparisons and transfer to new campaigns.
That leaves a precise north-star question: Does each generation become better at building the next? For us, the answer must emerge from better protein design models, better experiments, and an LLM increasingly capable of improving both.
Background reading
Adams et al. Measuring the sequence-affinity landscape of antibodies with massively parallel titration curves. eLife (2016).
Rapp, Bremer & Romero. Self-driving laboratories to autonomously navigate the protein fitness landscape. Nature Chemical Engineering (2024).