Jev predicts protein binder affinity zero-shot
Computational biologist Dr. Colby Ford published a study evaluating TypeSafe AI's "System One" decision engine architecture—via Featherless's open-source Simple Jev stack—to classify whether de novo computationally designed proteins bind their biological targets. Using Anthropic's dataset of 1,440 Claude-generated binder designs and Boltz-2 structural metrics, the zero-shot approach achieved balanced accuracies above 60% and recall exceeding 70% without biological fine-tuning, demonstrating how calibrated decision models can triage candidates prior to wet lab synthesis.
Using fast, calibrated System 1 decision-routing models like Jev as zero-shot meta-classifiers over raw structural metrics is a clever paradigm that proves general LLMs can parse complex numerical biophysics data, though high false-positive rates make wet lab triage far from solved.
* Out-of-Distribution Generalization: General foundation models (like Qwen) with no biology-specific training managed to interpret tabular structural metrics like ipSAE and DockQ to achieve an ROC-AUC of up to 0.68.
* High Recall, Low Precision: While binder recall exceeded 70% across both Adaptyv and Twist sets, poor precision at default thresholds indicates significant false-positive rates that require probability threshold tuning.
* Speed and Cost Efficiency: The Jev architecture's single-pass, structured probabilistic outputs avoid the massive token overhead and latency of autoregressive text generation when triaging thousands of candidate sequences.
* Benchmark Limits and Experimental Noise: The results highlighted the inherent messiness of protein design validation, as experimental binding assays from different vendors disagreed nearly 9% of the time.
DISCOVERED
1h ago
2026-09-22
PUBLISHED
1h ago
2026-09-22
RELEVANCE
AUTHOR
picocreator