Work

A Reported Featurizer Ranking Instability Is Noise Between Inseparable Arms

Year: 2026

Type: preprint

Abstract: Representation comparisons are commonly published from one split at one seed on uncurated public sets. Repeating one, the count of endpoints whose best featurizer changes with the split read three, one, one and three under two choices, seeds and deduplication, and every reorder lay between arms whose intervals overlap. Only the fingerprint baseline separated anywhere, by being worse; duplicates placed up to 3.8% of random-split test rows in training. A four-step remedy is given.

Source: ChemRxiv

DOI: 10.26434/chemrxiv.15009293/v1

Author: Phani Kumar Balagam

Language: English

Topic: Computational Drug Discovery Methods

Subfield: Drug Discovery

Field: Pharmacology, Toxicology and Pharmaceutics

Domain: Life Sciences

Open Access status: green

Funders: National Institutes of Health, National Center for Advancing Translational Sciences

OpenAlex ID: W7214062097

Cites: 8

Cited by:

Related to: 10