NCI Division of Cancer Biology Junior Investigators Meeting (Abstract): Submission #5

Submission information
Submission Number: 5
Submission ID: 192257
Submission UUID: 1f2b0c9c-bf77-4fd0-b339-64de85977add
Submission URI: /dcb/ji-meeting/abstract

Created: Wed, 08/26/2026 - 12:09
Completed: Wed, 08/26/2026 - 12:09
Changed: Wed, 08/26/2026 - 12:09

Remote IP address: 10.208.24.244
Submitted by: Anonymous
Language: English

Is draft: No
First Name Aakash
Middle Initial
Last Name Saha
Degree(s) Ph.D.
Position/Title/Career Status Postdoctoral Research Scientist
Organization Columbia University
Organization Address New York
Email as7656@columbia.edu
Abstract Category Consider my abstract for a Methodology/Technology presentation
Abstract Keywords Protein–protein interactions, Structural bioinformatics, Parallel computing, Cancer signaling networks, Structure-based interaction prediction
Abstract Title A User-Accessible Python-Based PrePPI Framework for Protein-Protein Interaction Prediction
Abstract Protein–protein interactions (PPIs) underlie cellular processes, yet proteome-scale identification of functional interactions remains computationally challenging. All-to-all analyses may require evaluation of billions of protein and domain pairs, a burden amplified when considering interactions across species, including hosts and pathogens. PrePPI makes proteome-scale searches for binary interactions tractable by using evolutionarily conserved structural cues to prioritize pairs that can be modeled with known interaction templates.
PrePPI-SM evaluates interactions between structured protein regions using structural similarity, interface coverage, and conserved contacts within a naïve Bayesian framework. PrePPI-SLiM identifies interactions in which a structured domain recognizes a short peptide-like motif in a partner protein.
We rewrote the Perl-based PrePPI-SM pipeline as an end-to-end Python workflow and redesigned Skan, its structural-neighbor search, for scalable execution. Skan partitions a database of representative protein structures into balanced shards and splits large sequence clusters to prevent bottlenecks. A shared queue dynamically assigns these shards among 30 CPUs as workers become available, improving utilization across uneven workloads. The original workflow repeatedly searched directories and opened small interface files, creating input/output overhead. We consolidated these data into indexed SQLite databases so workers retrieve only required records, reducing file-system traffic and unnecessary data loading. Across 447 human proteins and domains, the redesigned structural-neighbor search achieved a median 13.4-fold end-to-end speedup without altering the structural scoring framework. Now comparable to Foldseek in runtime, Skan complements Foldseek’s efficient close-match search by capturing more diverse structural neighbors. This helps identify a binary complex with an appropriate interface for predictive modeling.
These approaches could contribute to collaborative cancer research by prioritizing interactions involving cancer-associated proteins, connecting poorly characterized proteins to signaling networks, and generating structural hypotheses for experimental testing. This workflow could complement experimental proteomics and functional studies, helping collaborators select candidate interactions and interpret their potential roles in cancer biology.