concept · intermediate · partly written
Many representations of one object, chosen per task or per query
Several encodings of the same molecule or entity are kept and one is picked for the question; per-task winners and concatenations are documented, a trained per-query chooser was not found.
His idea, in his words
"The solution may be to represent data in many different ways, so it exists in the superposition of many different states. You sample the data for different representations, and you take the one that's most similar at a certain point." 1
Amit's hypothesis
The idea stated so it can be wrong more
A selector over several encodings of one object beats the best single fixed encoding when the choice is made on training data only, on a split that withholds structure. Our reading of his words. 23
Our inference
What already exists more
MoleculeNet's authors conclude that learnable models broadly offer the best performance, that below about 3,000 samples kernel and tree methods are more robust, and that on quantum-mechanical sets physics-aware featurisations can matter more than the learning algorithm. These are per-task winners chosen by the benchmark's author, not by a selector. Concatenating several fingerprints and descriptors improved results on 22 ADMET tasks in one preprint (abstract only), which supports using many at once, not picking one. TopExpert routes molecules to topology-specific graph experts (abstract only; baselines not seen). 456
Reported, not yet replicated
Evidence that the fixed simple choice is hard to beat more
Riniker and Landrum compared 12 fingerprints plus two simple baselines in virtual screening with the similarity method fixed; only the baselines differed significantly. Praski and colleagues tested 28 pretrained embedding models on 25 datasets and report that all neural models except CheMeleon failed to outperform the classical ECFP count fingerprint. 23
Reported, not yet replicated
What remains open (our inference) more
A trained router, nearest view or learned, against the best single fixed representation, on a scaffold or temporal split, with the choice made inside nested validation: not found in the sources opened. This is an absence in what was read, not a measured negative for selection. 43
Our inference
Where it would fail more
Selecting on the evaluation set inflates results, and the measured pattern is that a tuned fixed fingerprint is hard to beat; most multi-view gains found were fusion against a single view, not selector against fixed. 35
Our inference
Cheapest test (our inference) more
Five to ten MoleculeNet or Therapeutics Data Commons sets on a scaffold split: choose among ECFP count, RDKit descriptors and one pretrained embedding per task with a nested cross-validation selector; compare with ECFP count given the same tuning budget, and with the oracle that picks the best per task. If selector minus ECFP sits inside run-to-run noise, selection has shown nothing. 43
Our inference
Still open here
- Does any per-query representation chooser survive nested validation and a scaffold split?
Not yet written: Plain answer, Why it matters, Equation, Real numbers.
Linked from
Sources
- Amit's dictation of 2026-11-18: how biology data is represented, and where ontologies can be innovated. Kept in the atlas repository; not published.
- Riniker S, Landrum GA. Open-source platform to benchmark fingerprints for ligand-based virtual screening. Journal of Cheminformatics (2013). paper; opened 2026-11-18 all records citing this
- Praski M, Adamczyk J, Czech W. Benchmarking Pretrained Molecular Embedding Models For Molecular Representation Learning. arXiv (2025). preprint, not peer reviewed; opened 2026-11-18 all records citing this
- Wu Z, Ramsundar B, et al.. MoleculeNet: a benchmark for molecular machine learning. Chemical Science (2018). paper; opened 2026-11-18 all records citing this
- Notwell JH, Wood MW. ADMET property prediction through combinations of molecular fingerprints. arXiv (2023). preprint, not peer reviewed; opened 2026-11-18 all records citing this
- Kim S, Lee D, Kang S, et al.. Learning Topology-Specific Experts for Molecular Property Prediction. arXiv (AAAI 2023) (2023). paper, not peer reviewed; opened 2026-11-18 all records citing this
Where it sits on the map
History of this record (5)
This record has been changed 5 times, on the dates below. The note kept against each change is written for the people who maintain the atlas, in their own shorthand, so it is not shown here; it is in the record itself.
- Status
- partly written
- Last updated
- 2026-11-19
- Record
- content/concepts/many-representations-chosen-per-task.json