Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
ToDD: Topological Compound Fingerprinting in Computer-Aided Drug Discovery
Authors: Andaç Demir, Baris Coskunuzer, Yulia Gel, Ignacio Segovia-Dominguez, Yuzhou Chen, Bulent Kiziltan
NeurIPS 2022 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform extensive numerical experiments in VS, showing that our To DD models outperform all state-of-the-art methods by a wide margin (See Figure 1). |
| Researcher Affiliation | Collaboration | Andac Demir Novartis EMAIL Baris Coskunuzer University of Texas at Dallas EMAIL Ignacio Segovia-Dominguez University of Texas at Dallas Jet Propulsion Laboratory, Caltech Yuzhou Chen Temple University Yulia Gel University of Texas at Dallas National Science Foundation Bulent Kiziltan Novartis EMAIL |
| Pseudocode | No | The paper describes the process in numbered steps but does not provide a formal pseudocode block or algorithm listing. |
| Open Source Code | No | The paper does not contain an explicit statement about releasing the source code for the described methodology or a direct link to a code repository for their work. |
| Open Datasets | Yes | Cleves-Jain: This is a relatively small dataset [26] that has 1149 compounds.* There are 22 different drug targets, and for each one of them the dataset provides only 2-3 template active compounds dedicated for model training, which presents a few-shot learning task. ... *Cleves-Jain dataset: https://www.jainlab.org/Public/SF-Test-Data-Drug Space-2006.zip DUD-E Diverse: DUD-E (Directory of Useful Decoys, Enhanced) dataset [67] is a comprehensive ligand dataset with 102 targets and approximately 1.5 million compounds.* ... *DUD-E Diverse dataset: http://dude.docking.org/subsets/diverse |
| Dataset Splits | Yes | The performance of all models was assessed by 5-fold cross-validation (CV). |
| Hardware Specification | Yes | Training time of To DD-Vi T and To DD-Conv Ne Xt for each individual drug target takes less than 1 hour on a single GPU (NVIDIA RTX 2080 Ti). |
| Software Dependencies | No | The paper mentions software components like 'Adam optimizer' and 'Conv Ne Xt_tiny models' but does not specify their version numbers or the versions of underlying libraries like Python or PyTorch. |
| Experiment Setup | Yes | Transfer learning via fine-tuning Vi T_b_16 and Conv Ne Xt_tiny models using Adam optimizer with a learning rate of 5e-4, no warmup or layerwise learning rate decay, cosine annealing schedule for 5 epochs, stochastic weight averaging for 5 epochs, weight decay of 1e-4, and a batch size of 64 for 10 epochs in total led to significantly better performance in Enrichment Factor and ROC-AUC scores compared to training from scratch. |