Bio papers — 2026-09-22

Today's work focused on improving how existing predictive models can find active compounds more quickly when experimental data is limited. Researchers looked at fine-tuning the affinity heads of Boltz-2 using only a small set of binary activity labels. This approach showed that it can significantly boost early enrichment, specifically increasing the number of actives in the top one percent by a geometric mean of 1.77 times across eight different targets. Furthermore, this method improved average precision by two point one four times compared to not fine-tuning these heads at all.

This improvement in finding hits is also maintained when only a small subset of candidates is rescored. Restricting that rescoring to about ten percent of the evaluation set kept the hit recovery comparable to doing a full rescore. This suggests that this method for improving early enrichment is robust and can be applied practically.

Separately, there is ongoing effort to create new chemical entities using deep learning models trained on natural products. These models have successfully generated compounds whose distributions closely resemble those found in nature. This offers a way to reduce the time and cost associated with discovering novel natural product-like drug candidates.

In terms of computational efficiency, a new framework called Boltzina is being developed. This framework aims to use the high accuracy of Boltz-2 without its slow structure prediction step by directly predicting affinity from docking poses. This method achieved up to eleven point eight times faster screening throughput than previous methods like AutoDock Vina while still performing better than AutoDock Vina and GNINA on eight assays from the MF-PCBA dataset.

Today's papers

Adapting Boltz-2 with limited experimental activity data improves early enrichment in virtual screening: Fine-tuning the affinity heads of Boltz-2 using limited experimental data can significantly improve how well it prioritizes active compounds in virtual screening. [paper]
NPGPT: Natural Product-Like Compound Generation with GPT-based Chemical Language Models: This method uses language models trained on natural products to generate new molecules that resemble those found in nature. [paper]
Boltzina: Efficient and Accurate Virtual Screening via Docking-Guided Binding Prediction with Boltz-2: Boltzina is a framework that makes the accurate binding prediction from Boltz-2 much faster for large-scale virtual screening by skipping the structure prediction step. [paper]
Improving the adaptive and continuous learning capabilities of artificial neural networks: This study suggests that mimicking how biological systems use neuromodulators can help artificial neural networks learn better and avoid forgetting old information.
How Metacognitive Architectures Remember Their Own Thoughts: This review examines how artificial agents can remember their own thoughts to improve their performance and explainability, but notes a lack of standardization in this area.

The papers

Important terms

Boltz-2: This is a predictive model being fine-tuned to find active compounds quickly using limited experimental data. It was shown to significantly improve early enrichment and average precision.
Early Enrichment: This refers to the ability of a model to successfully identify active compounds when you only have very little experimental data available. The research showed this method is very effective for finding hits.
Rescoring Subset: This technique involves only re-evaluating a small portion, about ten percent, of the candidate molecules after an initial screening. This keeps the hit recovery high while saving computational time.
Boltzina: A new computational framework designed to be much faster than previous methods. It predicts drug affinity directly from docking poses without needing a slow structure prediction step.