Constrained Submodular Optimization for Vaccine Design

Authors: Zheng Dai, David K. Gifford

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We supplement the theoretical improvement with an empirical comparison of vaccines designed using our approach and previous designs. Our proposed framework also contributes the following desirable properties: it makes explicit the utility of having redundancy, does not discount the benefits of being covered without redundancy, and is able to reason with uncertainty. We find that Optivax-P outperforms the baselines in most settings (Figure 2).
Researcher Affiliation Academia Zheng Dai, David K. Gifford Computer Science and Artificial Intelligence Laboratory, Massachusetts Institute of Technology Cambridge, Massachusetts 02139 USA zhengdai@mit.edu, gifford@mit.edu
Pseudocode Yes Algorithm 1: Optivax-P
Open Source Code Yes Code, data, appendices, and demo are accessible from https://gifford-lab.github.io/DiminishingReturns.
Open Datasets Yes To attain well calibrated credences, we make use of publicly available datasets that were used to validate Net MHCpan, which contain no overlap with the dataset used to train Net MHCpan (Reynisson et al. 2020). Liu et al. (2022) have also published a set of genotypes and their frequencies which we use.
Dataset Splits Yes A calibration curve of the calibrated predictions of the validation dataset is shown in Figure 1B.
Hardware Specification Yes Our implementation can generate designs of size k 102 over a peptide set of size |P| 103 with |M| 106 genotypes in approximately 5 minutes when parallelized over 8 Titan RTX GPUs.
Software Dependencies Yes We use well established state-of-the-art neural network based models (Net MHCpan4.1 and Net MHCpan II4.0 (Reynisson et al. 2020)) to generate predictions that we use to derive Pr(display(p, m)).
Experiment Setup Yes Designs were constrained such that no pair of peptides can be within 3 edits (insertions, deletions, or substitutions) of each other for the MHC Class I design, and 5 edits for the MHC Class II design. Designs were optimized for the objective FUT for T between 1 and 20 inclusive, where UT is defined as: UT (x) = min(x, T)