Residual adaptation improves sequential screening hit recovery rates
Historical-Feature Residual Adaptation for Sequential Screening
Machine Learning
Summary
Sequential screening tries to find important results using limited tests and feedback. The authors introduced a small correction that uses past information combined with initial scores to better rank candidates before new feedback arrives. Their method improved the fraction of hits found in tests compared to previous approaches. It also worked well when added to different base models, showing it can adapt pretrained methods effectively.
What this means in practice
- •For drug discovery teams: Improve compound ranking accuracy during sequential testing to identify promising drug candidates more efficiently within limited experimental budgets.
- •For machine learning engineers: Enhance pretrained ranking models by adding a lightweight residual adaptation using fixed historical features to improve decision making without retraining the whole model.
Authors
Yixiang Hong, Yuchen Zhang
Abstract
Sequential screening policies must use acquired feedback effectively under a fixed experiment budget. We study acquisition-time residual adaptation: a small ranking correction built from fixed historical features and anchored to base scores recorded before feedback is revealed. The interface accepts a history-conditioned scorer and a candidate feature table; its primary instantiation augments AssayFormer's ten-dimensional representation with 64 historical directions while retaining the pretrained model. In retrospective evaluation on 20 official-test screens from 15 recorded sources, a comparison fixed before target-outcome access increased source-macro fraction of hits recovered from 22.58% to 28.61%, exceeding all six prespecified comparators. A same-feature ordinary ranking head achieved 25.73%. Supplementary strong logistic, rank-fusion, and unprojected controls achieved 26.42%, 26.02%, and 26.78%, respectively. A separate model-independent pilot on 13 reused development screens reconstructs the features without AssayFormer assets. Across three seeds, residual adaptation improves random-forest hit recovery from 28.87% to 31.27%, while logistic recovery changes from 30.22% to 29.78%. The results show effective pretrained-policy adaptation and useful transfer to an independent nonlinear base, with gains that depend on the base model.