INSearch: AI-native framework for large scale proteomic database search via contrastive joint embeddings and transformer-based scoring function
thesis · 2026
| Date | 2026-06-29 |
| Type | thesis |
| Publisher | MSc thesis |
| Contribution | adjacent |
| Supervisor | Konstantinos Kalogeropoulos, Jeroen Van Goey |
| Link | https://jeroen.vangoey.be/files/theses/ai-guided-database-search-proteomics.pdf |
Abstract
In mass spectrometry-based proteomics, identifying the peptide sequence that generated a specific spectrum is a complex challenge. The main method that is used for peptide-spectrum matching is searching against a reference database using traditional search engines; these search engines use a tailored score function to assign a score to each match based on certain criteria. However, as datasets grow in size, the search space expands combinatorially due to the inclusion of many organism proteomes or post-translational modifications, and the time complexity of these search engines is O(S × ρP) (Kalogeropoulos et al., 2026), where S and P are the number of query spectra and peptides, respectively. The aim of this project is to design a preliminary framework prototype called INSearch. The INSearch objective is to provide a sublinear candidate retrieval complexity O(S(log P + K)) while efficiently maintaining identification accuracy in large-scale database searches. INSearch utilises a dual-encoder architecture to jointly project the experimental tandem mass spectra and theoretical peptide sequences into a shared latent space so that the spectra and their corresponding ground truth peptides align. The standard contrastive loss was used as the objective loss function with an additional component, variance regularisation (Bardes et al., 2022), to prevent feature collapse. Another loss function was used as an alternative to the standard contrastive loss, alignment and uniformity loss (Wang and Isola, 2020) that optimises the contrastive learning objective asymptotically. The similarity search is conducted through an approximate nearest neighbour search algorithm to retrieve the top-k peptide candidates for a certain spectrum query, providing a sublinear time complexity. Since the similarity search is approximate, the retrieved candidate list undergoes a neural re-ranking stage: the decoder of InstaNovo (Eloff et al., 2025) is used in teacher-forcing mode as a scoring function to assign residue-specific log-probabilities, which are aggregated using the geometric mean into a final confidence score for each match. On the nine-species benchmark, the two-stage INSearch achieved strong recall on the held-out yeast test split: a Stage-1 Recall@1/5/100 of 60.0/78.9/91.9%, with the neural re-scoring stage lifting Recall@1 to 82.9%, indicating cross-species generalisation. On the more challenging S. brodae proteome the results were more moderate (a Stage-1 Recall@100 of 55.5%), indicating a further need for scaling and training on massive datasets.
Methods and tools
- INSearch: Prototype AI-native database search that replaces the combinatorial scan with retrieval. A dual encoder projects experimental spectra and theoretical peptide sequences into one shared latent space under a contrastive objective with variance regularisation to prevent feature collapse, and an alternative alignment-and-uniformity loss; approximate nearest-neighbour search then returns the top-k candidates, cutting retrieval from the O(S × ρP) of a conventional engine to O(S(log P + K)). Because retrieval is approximate, candidates are re-ranked by InstaNovo’s decoder run in teacher-forcing mode as a scoring function, aggregating residue log-probabilities by geometric mean. On the nine-species benchmark the held-out yeast split reaches Recall@1/5/100 of 60.0/78.9/91.9%, with neural re-scoring lifting Recall@1 to 82.9%; the harder S. brodae proteome reaches only 55.5% Recall@100, pointing to a need for larger-scale training.
- InstaNovo: Knapsack beam search