Closed-loop genetic algorithms surface the earliest mutation signatures that destabilise P-53 (the "guardian of the genome") before malignant cascades take hold. Award-winning research that connects evolutionary computation and cancer genomics.
About half of all cancers carry a broken copy of a single gene called p53, the cell's built-in safety switch that tells damaged cells to shut down. This project uses evolution-inspired algorithms to spot the very first signs that p53 is breaking, long before cancer takes hold. The work won Best Research Project at UBC Vantage.
P-53 operates as the "guardian of the genome," halting cell division when DNA damage is detected. Mutations in this gene contribute to over 50% of all human cancers, making it a critical target for early detection strategies.
Think of P-53 as the cell's quality inspector. When it spots damaged DNA, it stops the cell from dividing so the damage cannot spread. When P-53 itself breaks, that safety check fails. A broken P-53 shows up in more than 50% of all human cancers, which is why catching the break early matters so much.
Traditional approaches focus on already-malignant sequences. This study identifies predictive patterns in pre-malignant mutations, potentially enabling intervention before cancer emerges.
Most cancer studies look at DNA that is already cancerous, like inspecting a fire after the house has burned. This study looks earlier, at the DNA just before it turns cancerous, hunting for warning patterns that could let doctors step in before the disease takes hold.
Combining genetic algorithms with self-organizing maps to trace deterministic pathways from healthy to malignant P-53 sequences, revealing early-warning biomarkers for clinical application.
The project pairs two tools. Genetic algorithms simulate how DNA changes step by step over many cell divisions. Self-organizing maps then sort those changes into groups, so we can see which tiny DNA edits show up just before a healthy P-53 turns into a cancer-driving one. Those edits become the early warning signs.
Concept Overview
Awarded Best Research Project at UBC Vantage College Capstone Conference for innovative genetic algorithm approaches to identify DNA characteristics leading to P-53 cancerous mutations.
Identify recurring mutation motifs that precede carcinogenic behaviour in the P-53 tumor suppressor gene by simulating mitotic propagation under controlled conditions.
The SOM surfaced six high-risk pentamer motifs (cagcc, agcca, cccag, ccagg, ttttt, ctttt) with an optimal 0.451 silhouette score under a 1×6 matrix configuration.
Curated 25 wild-type and cancerous P-53 DNA strands (2,509 bases each) from the NCBI repository. Pre-processed to remove non-nucleotide characters and aligned pathological/parental pairs.
First, we needed examples to study. We pulled 25 P-53 DNA sequences from the public NCBI database, some healthy, some cancerous. Each one is 2,509 letters long. We cleaned up stray characters and lined up each cancerous sequence next to the healthy version it came from, so we could compare them directly.
Spawned a binary tree representing mitotic bifurcation. Each node stores generation index, DNA composition, and malignancy state. Recursion continues to generation 14 to emulate tumour initiation depth.
Cells grow by splitting in two, again and again. We modelled that as a tree: one cell becomes two, two become four, and so on. Each branch remembers which generation it belongs to, what its DNA looks like, and whether it has turned cancerous. We let the tree grow 14 generations deep, roughly the depth where a tumour usually gets started.
Depth-first traversal extracts generational paths, evaluates mismatch rates via Levenshtein similarity, and flags the highest-drift ancestors preceding malignant nodes.
Next we walked each branch of the tree from root to tip. For every step, we measured how much the DNA had drifted away from its parent using Levenshtein similarity, a score that counts how many letters changed. The branches that drifted the most just before turning cancerous are the ones we flagged as suspects.
Calculated log₄(L) to choose k = 4, transforming each strand into a 1×1024 feature vector representing nucleotide frequency. Result: 73,475 length-adjusted rows.
DNA is a long string of just four letters. To turn that into something a computer can compare, we chopped each sequence into overlapping 4-letter chunks (k = 4, chosen with log₄(L)) and counted how often each chunk appeared. That gives every sequence a 1,024-number fingerprint. In total we had 73,475 fingerprints to compare.
Applied SOM grids ranging 1×2 to 1×11. Correlation analysis reduced dimensional redundancy before finalising a 1×6 lattice that maximised separation with minimal distortion.
Finally we used a self-organizing map, a method that groups similar fingerprints together, like sorting socks by colour. We tried sorting them into anywhere from 2 to 11 groups. Six groups turned out to give the cleanest split: each pile was tight and clearly different from its neighbours.
The SOM revealed six distinct mutation clusters with clearly differentiated nucleotide signatures. Clusters enriched in thymine-heavy motifs surfaced consistently in malignant branches, providing early warning indicators for cancer development.
Achieved optimal 0.451 silhouette score with 1×6 grid configuration, indicating well-separated cluster centroids with minimal overlap between mutational trajectories.
Highlighted motifs align with known loss-of-function trajectories for P-53, establishing a computational pipeline to monitor early mutational convergence in other cancer datasets.
Six clusters emerged as reliable precursors to malignant outcomes. Each cluster possesses a signature nucleotide fingerprint, reinforcing that mutation progression is pathway-dependent, not random.
Silhouette scores climbed steadily from 1×2 through 1×6 matrices before dropping sharply. The 1×6 configuration maintains separation without sacrificing interpretability.
Limitations include simulated rather than patient-specific conditions and a relatively small number of base sequences. Framework is portable for expanding datasets.
Scaling this methodology to other tumour suppressor genes could expose similar early-warning mutation signatures, guiding screening pipelines before clinical symptoms manifest.
Genetic algorithms plus SOM clustering expose deterministic shifts toward malignancy. Six pentamer motifs consistently precede malignant conversion, providing high-clarity monitoring targets.
The big takeaway is that cancer is not pure bad luck at the DNA level. Cells seem to follow a predictable path on the way to going rogue. Six short DNA patterns (each five letters long) keep showing up just before that switch flips. That gives doctors and researchers a clear list of things to watch for.
Encoded feature space remains interpretable, enabling rapid clinician dialogue. The motif shortlist now feeds wet-lab validation and computational monitoring pipelines.
A nice bonus: the maths stays readable. A clinician can look at the groups and understand what each one means, instead of trusting a black box. The shortlist of suspect DNA patterns is already being passed on for lab testing and for software that scans patient samples.
Future work will expand datasets with longitudinal patient samples and fuse expression-level data to link mutational motifs with phenotypic impact.
What comes next: track real patients over time instead of simulated cells, and add data on which genes are actually switched on or off. That way we can connect each DNA pattern to what it actually does to a person's health.
Di Leo, A., et al. (2007). p-53 gene mutations as a predictive marker in advanced breast cancer. Annals of Oncology, 18(6), 997-1003.
Vogelstein, B., Lane, D., & Levine, A. J. (2000). Surfing the p53 network. Nature, 408, 307–310.