Protein solubility depends on centrifugation: Aiki-Sol, a per-regime predictor for E. coli

The paper introduces Aiki-Sol, a protein solubility predictor that overcomes the performance plateau of existing models by explicitly accounting for centrifugation regimes as a critical feature rather than noise, achieving significant accuracy gains on a newly released, stringency-annotated E. coli dataset.

The Gist

Imagine you are trying to teach a computer to predict whether a specific protein (a tiny building block of life) will dissolve nicely in water or clump up into a solid mess when made inside a bacteria called E. coli. For the last eight years, scientists have been using advanced AI to make these predictions, but they've hit a wall. The computers aren't getting any better, no matter how smart they get.

The Hidden Problem: The "Spin" Confusion
The paper argues that the computers aren't failing because they aren't smart enough; they are failing because they are being tricked by a hidden variable: centrifugation.

Think of making a protein like making a smoothie with chunks of fruit.

• If you put the smoothie in a blender and spin it slowly, the big chunks stay at the bottom, and the liquid on top looks clear. You call this "soluble."
• If you spin it super fast, even the tiny bits get forced to the bottom, leaving you with almost no liquid. You might call this "insoluble."

The protein itself hasn't changed. It's the same smoothie. But the method used to separate the liquid from the solids (the "centrifugation regime") changes the result.

For years, scientists have been feeding their AI models data where the "spin speed" was hidden. They just labeled everything as "soluble" or "insoluble." It's like trying to teach a student to predict the weather, but you hide the fact that some data comes from a sunny beach and some comes from a rainy mountain. The student gets confused because the rules seem to change randomly. The paper calls this a "latent confound"—a hidden trap in the data.

The Solution: Aiki-Sol and the New Dataset
The researchers fixed this by creating a massive new library of data called the Aiki-Sol Dataset. Instead of just saying "soluble" or "insoluble," they tagged every single protein with exactly how hard it was spun (the "stringency").

They organized this into three tiers:

1. The Benchmark: A strict, high-quality set of about 85,000 proteins where the spin speed is known.
2. The Extension: A larger set of about 147,000 proteins with just the basic labels.
3. The Research Pool: A huge collection of about 229,000 proteins from various sources.

The Results: It's About the Rules, Not the Brain
When they tested old AI models on this new, honest data, the results were shocking. On the "high-speed spin" group, the best existing models actually performed worse than random guessing (like flipping a coin). They were so confused by the hidden spin speeds that they got it wrong more often than right.

Then, they built a new model called Aiki-Sol.

• The Trick: Instead of trying to guess one single answer, Aiki-Sol is trained to give five different answers depending on how hard the protein is spun, plus one answer if the spin speed is unknown.
• The Surprise: They found that making the AI "bigger" (adding more brainpower or using complex 3D structures) didn't help. The magic wasn't in the architecture; it was in curation. By teaching the AI to pay attention to the "spin speed" rules, a standard-sized model suddenly became much smarter.

The Outcome
When tested on new groups of proteins that the AI had never seen before, Aiki-Sol jumped from a success rate of about 70% to over 82%. Even more impressively, on groups where the AI had zero prior knowledge of the specific proteins, it still improved by a huge margin.

In a Nutshell
The paper claims that for years, protein solubility predictors were stuck because they ignored the "spin speed" used in the lab. By creating a new dataset that respects these different lab conditions and teaching the AI to adapt its predictions based on them, they broke the performance plateau. The key wasn't building a bigger, more complex brain, but rather teaching the existing brain to understand the specific rules of the game.

Technical Summary

Technical Summary: Protein Solubility Depends on Centrifugation: Aiki-Sol, a per-regime predictor for E. coli

Problem Statement

Sequence-based predictors for recombinant protein solubility in Escherichia coli have reached a performance plateau, with independent-test AUC scores stagnating between 0.760 and 0.80 over eight years of protein language model (PLM) development. The authors identify a latent confound driving this stagnation: the centrifugation regime used to separate soluble from insoluble fractions is a hidden variable collapsed into a single binary "soluble" label. While the protein's biochemistry remains constant, the specific fraction of the lysate recovered as "soluble" varies significantly based on the centrifugation stringency (e.g., g-force). Existing predictors treat these varying regimes as label noise rather than a predictive feature, and sequence overlap between training and test sets masks the resulting failure modes, particularly when models are applied to regimes unseen during training.

Methodology

To address this, the authors introduce the Aiki-Sol Dataset, a tiered corpus of E. coli solubility data:

• Benchmark Tier: A ~85K dataset with explicit stringency annotations.
• Extension Tier: An Apache-licensed ~147K dataset adding binary-only labeled proteins.
• Research Tier: A ~229K pool incorporating non-commercially licensed sources.

The authors developed Aiki-Sol, a model that jointly supervises five distinct outputs corresponding to specific centrifugation stringencies, alongside a marginal output for proteins with unknown stringency. The architecture utilizes a fine-tuned ESM-2 650M backbone.

Key experimental design choices included:

• Partitioning: Evaluation was performed on sequence-cluster-disjoint partitions to prevent data leakage.
• Ablation Studies: The authors tested whether structure-aware predictors (using ESMFold structures) or increased model capacity (scaling to 3B parameters) improved performance over the sequence-only, conditioned 650M backbone.
• Baseline Comparison: The model was compared against the strongest published binary comparators.

Key Results

• Failure of Binary Models on High Stringency: On the 32,000 x g stratum of the benchmark, the strongest published binary comparator fell below chance performance (AUC 0.491 ± 0.020), indicating a fundamental inability to generalize to high-stringency regimes when trained on mixed data.
• Performance Gains: A fine-tuned ESM-2 650M backbone with five protocol-matched outputs increased the pooled AUC by +0.108 (paired-bootstrap CI lower bound ≥ +0.090).
• Architecture vs. Curation: The performance gain was attributed to data curation and regime conditioning rather than architectural complexity. Structure-aware predictors using ESMFold structures did not outperform the sequence-only frame, and scaling parameters to 3B did not exceed the performance of the conditioned 650M backbone.
• External Validation: On five external cohorts, Aiki-Sol lifted the cohort-mean AUC from a baseline of 0.69–0.70 to 0.825. Notably, on three cohorts with measurably zero training-pool overlap, the model achieved a lift of +0.10 to +0.16.

Significance

The paper claims that the stagnation in protein solubility prediction is not a limitation of sequence modeling capacity but a consequence of ignoring the experimental context (centrifugation regime) in the training labels. By explicitly modeling the dependency of solubility on centrifugation stringency, Aiki-Sol resolves the failure modes of previous binary classifiers. The work demonstrates that accurate prediction in this domain requires treating the experimental protocol as a primary feature rather than noise, enabling robust performance even on data with no sequence overlap with the training set.