HELIXA

Turning Genomics into Decisions

KBU-MedLab · TEKNOFEST Oncology 3T Competition

---

What HELIXA is

HELIXA is a biomedical AI platform for glioblastoma molecular subtyping and biomarker discovery. It takes raw TCGA/GDC RNA-seq data, removes the technical artefacts that would otherwise corrupt the result, discovers molecular subtypes, classifies new patients against those subtypes, validates every conclusion against independent published literature — and, for each subtype, surfaces the genes that both mark it in real patients and that a glioblastoma cell cannot survive without, checked one at a time against the published literature and tiered by how actionable they actually are.

It is a working system, not a mock-up. Every number the web interface displays is read from a file produced by the analysis pipeline in this repository, and the "Analyze Patient" workflow runs the actual frozen model. When a patient is classified, the site now also shows the biomarker/drug-target panel for the predicted subtype, and a one-click, self-contained Patient Molecular Report (logo, classification, biomarkers, technical appendix) that the browser turns into a PDF with no server involved.

Cohort: 328 glioblastoma patients · Subtypes discovered: 6 · Classifier: 97.9% leave-one-out agreement across all 328 patients, with calibrated subtype-stability probabilities · Biomarker candidates: 61 literature-checked genes across the 6 subtypes, 17 of them tier-1 actionable.

---

The pipeline

Raw GDC RNA-seq  (60,660 gene loci)
        ↓  01  Data ingestion
Expression matrix  (log2 TPM+1)
        ↓  02  QC / low-expression filter          → 27,777 genes
        ↓  03  Protocol detection + batch correction → 25,738 genes, batch signal 11.07 → 0.0000
        ↓  04  Feature selection (MAD)             → 1,000 signature genes
        ↓  05  Dimensionality reduction (PCA)      → 5 components
        ↓  06  Consensus clustering (1,000 resamples) → k = 6, PAC = 0.2105
        ↓  07  Classification (multinomial logistic regression)
        ↓  08  Confidence & uncertainty            → core / boundary
        ↓  09  Biological validation               → 4 independent published sources
   Subtype · confidence · markers · pathway profile · therapeutic context
        ↓  10  Biomarker discovery (independent second pipeline, see below)
   Per-subtype drug-target panel · tiered by literature evidence · with sources

The quality-control step that mattered most

The strongest signal in the uncorrected cohort was not tumour biology — it was library preparation protocol. Two protocols were mixed: poly(A)-selected libraries contained 0.42% non-polyadenylated RNA, total-RNA libraries contained 55.45% — a 170× difference, and a single threshold separated them with 100% accuracy.

Left uncorrected, the pipeline would have "discovered" the sequencing protocol and presented it as a molecular subtype. A three-layer correction (gene-space filter → re-normalisation → within-batch standardisation) removes it to a residual ARI of -0.0044.

---

The six subtypes

#	Subtype	n	Verhaak identity	Independent verdict
0	MTC — Mitochondrial / OXPHOS	40	Mixed (Neural-leaning)	Confirmed (moderate) · 2.15×
1	PN — Proneural / Progenitor	49	Proneural-like	Confirmed (strong) · 2.69×
2	CL — Classical / EGFR	65	Classical-like	Partially supported · 1.96×
3	MES — Mesenchymal / Immune	63	Mesenchymal-like	Confirmed (strong) · 2.84×
4	INT — Intermediate / Mixed	67	Mixed	Partially supported · 1.36×
5	OLIGO — Oligodendrocytic / Myelin	44	Mixed (Proneural-leaning)	Confirmed (strong) · 2.32×

cluster_0 (MTC) independently matches the mitochondrial subtype published in Nature Cancer 2021 at 72.5% (2.9× chance, p = 2.7 × 10⁻¹¹) — a subtype reported to carry the most favourable prognosis and selective vulnerability to OXPHOS inhibitors.

---

Biomarkers & drug targets

Classifying a patient answers "which subtype is this?" A second, independent pipeline — the HELIXA biomarker engine — answers the question a classifier cannot: for this subtype, what could actually be drugged, and how sure are we?

It is deliberately built to avoid the mistake an earlier attempt made: taking a subtype's marker genes and asking a dependency database "is this gene essential?" without checking essential compared to what. That approach returned ten genes — mostly RNA-polymerase subunits and ribosomal proteins that every human cell needs equally — and left one subtype (MTC) with nothing at all.

The three gates a gene has to pass, in order:

1. Identity — it marks the subtype in the 328 real patients (Welch t-test, FDR < 0.05, Cohen's d > 0.5, and a specificity margin > 0.15z — the gene's mean expression in this subtype must beat the best of the other five, not their average).
2. Tumour dependency, with a safety window — using DepMap CRISPR gene-dependency screening across 1,208 cancer cell lines (53 glioblastoma, 1,118 from outside the brain), the gene must be a real dependency of glioblastoma specifically — high dependency in the 53 GBM lines, and low dependency in the 1,118 non-brain lines. 1,329 of the screened genes are dependencies of essentially every cell line in existence; those are discarded outright, because a drug against them would harm the patient as much as the tumour.
3. Literature check — every gene that survives both gates is looked up individually in PubMed, the Human Protein Atlas, Open Targets, DrugBank and ClinicalTrials.gov. This is the step that catches what arithmetic cannot: two of the genes that pass both statistical gates are established tumour suppressors — inhibiting them would help the tumour, not hurt it — and are kept visible, marked EXCLUDED with the reason, rather than silently dropped.

Final tiers, shown on the site and in every patient report:

Tier	Meaning
Tier 1 · Actionable	passes both data gates, published support, a drug against it already exists (possibly for another cancer)
Tier 2 · Credible	passes both data gates, published support, nothing built against it yet
Tier 3 · Hypothesis	passes both data gates, but the literature is thin — a real finding, unproven
Excluded	the gene is a tumour suppressor, or its biology otherwise points the wrong way

Result: 359 genes pass both statistical gates; 61 of the strongest were checked against the literature; 17 are tier-1 actionable, 18 tier-2 credible, 21 tier-3 hypotheses, and 5 were excluded. Two of the tier-1 targets already have drugs approved in other cancers (FGFR1 → pemigatinib, PTK2 → defactinib); two others already failed glioblastoma trials (ITGB5 → cilengitide, phase III CENTRIC; CDK6 → palbociclib, terminated for futility) — that failure is recorded on the same row rather than left out.

The full six-stage pipeline — patient-side differential expression, DepMap dependency screening, intersection & scoring, literature evidence collection, evidence integration, and the reporting figures — with every intermediate table and every literature source, lives in BIOMARKER_ENGINE/ in this repository. EXPLANATION_STEP_BY_STEP.txt there is the narrative walkthrough; 05_Tables/final_targets.csv is the exact file the website's data/biomarkers.json (and every biomarker table on the site) is generated from.

On the site: after a patient is classified, a "Biomarkers & drug targets for <subtype>" panel appears automatically underneath the result, and a Download / Print Patient Report button generates a self-contained report (see below) for that patient and subtype.

Same caveat as the classifier, stated plainly: these are computational candidates from an unpublished research pipeline, not treatments. Nothing on this list has been tested in a laboratory by this project. The dependency evidence comes from cell lines, not patients — cell lines lack an immune system, a blood-brain barrier and a tumour microenvironment, which is a specific reason the MES (immune-infiltrated) panel may be understated.

---

Patient Molecular Report

Every classified patient can be turned into a one-click, fully self-contained report — the HELIXA logo, the patient/sample label, the predicted subtype and confidence, all six probabilities, the genes that drove the decision, and the full biomarker/drug-target table for that subtype, with a plain disclaimer that this is a research prototype and not a diagnosis.

There is no PDF library and no server involved: Download / Print Patient Report opens the report as a normal page in a new browser tab (website/frontend/js/report.js), built from the exact same result object the on-page card already rendered and the exact same data/biomarkers.json dataset, so the report can never disagree with what the page showed. The tab's own "Print / Save as PDF" button (or the browser's native Ctrl/Cmd + P) hands the page to the browser's print engine, which is how the patient/clinician saves it as a PDF — the same mechanism a normal web page uses to become a PDF, with no upload and no third-party service anywhere in the path.

---

Model evaluation

The classifier was rebuilt in September 2026. Version 2 is trained on all 328 patients against the consensus profile — the fraction of 1,000 resampled clusterings in which a tumour co-clusters with each subtype — rather than on the 273 clean patients against hard labels. See prediction_model/ for the full write-up.

Metric	Value
Leave-one-out agreement, all 328 patients	97.87%
on CORE (n = 273)	99.63%
on BOUNDARY (n = 55)	89.09%
Agreement with the consensus profile (out-of-sample, 328 × 6 cells)	r = 0.9656
Profile RMSE	0.0766
Mean absolute error on the reported percentage	± 0.113
Median reported probability	0.7370
Permutation null test (300 shuffles)	real 99.63% vs 19.30% random, p = 0.005
Protocol detector	99.695% (327/328)

Twelve model families were compared out-of-fold against the consensus profile — polynomial expansions, kNN at four neighbourhood sizes, RandomForest and ExtraTrees. The plain linear model beat all eleven alternatives, so the residual error is irreducible with five principal components rather than a failure to fit.

What changed from v1, and why

v1 reported a median confidence of 0.9993, with 178 of 328 patients at ≥ 99.9%. It was trained only on the 273 CORE patients — the subset the clustering had already separated cleanly — which are nearly linearly separable in 5-PC space, so the softmax saturated. The 55 intermediate tumours were excluded entirely and the model had never seen ambiguity.

	v1	v2
trained on	273 CORE	328 (all)
median probability	0.9993	0.7370
patients at ≥ 99.9%	178 / 328	1 / 328
mean probability on BOUNDARY tumours	0.774	0.474

The percentage now means subtype stability: "73% CL" means this expression profile co-clusters with Classical in about 73% of resampled clusterings. It is not a diagnostic probability.

---

Repository layout

HELIXA/
├── NEW_START_RESULTS/            the clustering/classification project (unmodified)
│   ├── 01_Data_Preparation_extra/
│   ├── 02_Batch_Diagnosis/       the library-prep artefact, diagnosed
│   ├── 03_Clustering_Sweep/      726 configurations evaluated
│   ├── 04_Consensus_Clustering/  final assignments + embedding
│   ├── 05_Clusters/              one folder per subtype, with its patients
│   ├── 06_Biological_Validation/ marker genes, signature & pathway scores
│   ├── 07_Plots/                 23 original figures
│   ├── 08_Evaluation_Report/     global metrics + master workbook
│   ├── 09_Scripts/               every analysis script
│   ├── 10_Prediction_Model/      frozen model + predict_new_patient.py
│   └── 11_Evaluation_From_Internet/  independent literature validation
│
├── BIOMARKER_ENGINE/              the biomarker/drug-target discovery project
│   ├── EXPLANATION_STEP_BY_STEP.txt   narrative walkthrough, start here
│   ├── 01_Engine/                 the 5 scripts, run in order (01 → 05)
│   ├── 03_Evidence/               literature_evidence.{csv,json} — every source URL
│   ├── 04_Plots/                  the 10 official reporting figures
│   └── 05_Tables/                 candidates_all.csv, final_targets.csv (the
│                                   exact file website/frontend/data/biomarkers.json
│                                   is generated from), dependency_side_depmap.csv
│
├── website/
│   ├── backend/                  Starlette API wrapping the real model
│   │   ├── app.py
│   │   ├── helixa_engine.py      thin wrapper over the frozen artifacts
│   │   ├── export_static_data.py results (+ BIOMARKER_ENGINE) → JSON for the frontend
│   │   └── requirements.txt
│   └── frontend/                 zero-dependency SPA (no build step)
│       ├── index.html
│       ├── css/helixa.css
│       ├── js/                   router, data layer, SVG charts, views,
│       │                         report.js — the Patient Molecular Report
│       ├── data/                 17 JSON datasets, incl. biomarkers.json
│       └── assets/               logos + 27 scientific figures
│
└── .github/workflows/deploy.yml  GitHub Pages deployment

---

Running locally

Frontend only (all recorded results, no live prediction)

cd website/frontend
python3 -m http.server 5173
# open http://127.0.0.1:5173

No build step, no npm install — the frontend is native ES modules and hand-written CSS.

With the live model (enables "Analyze Patient")

cd website/backend
pip install -r requirements.txt
uvicorn app:app --port 8000

Reload the site. It auto-detects the API at 127.0.0.1:8000 and the upload workflow becomes active. Override the base URL with ?api=http://host:port or the field on the Analyze page.

Classify a patient from the command line

cd NEW_START_RESULTS/10_Prediction_Model
pip install scikit-learn joblib pandas numpy
python3 predict_new_patient.py  /path/to/patient.rna_seq.augmented_star_gene_counts.tsv

Re-export the web data after re-running the science

cd website/backend
HELIXA_RESULTS=../../NEW_START_RESULTS \
HELIXA_BIOMARKER_ENGINE=../../BIOMARKER_ENGINE \
python3 export_static_data.py

This regenerates every JSON in website/frontend/data/, including biomarkers.json from BIOMARKER_ENGINE/05_Tables/final_targets.csv. If HELIXA_BIOMARKER_ENGINE does not point to a real final_targets.csv, the script skips that one file and says so — it never fabricates biomarker data.

---

Architecture

Browser ── HELIXA frontend (static, GitHub Pages)
                │
                ├── /data/*.json ........ recorded results, always available
                │     including biomarkers.json (from BIOMARKER_ENGINE/05_Tables/final_targets.csv)
                │
                ├── /model/*.bin ........ frozen classifier weights, fetched once
                │     └── js/engine.js — the classifier's forward pass, in the browser
                │
                ├── js/report.js ........ builds the Patient Molecular Report
                │     (new tab, browser print-to-PDF — no server, no PDF library)
                │
                └── HTTP ── Starlette API (local) ── helixa_engine.py
                                                          │
                                              model_artifacts.joblib
                                              (frozen PCA + classifier +
                                               batch means + gene masks)
                                                          │
                                                   real prediction

The scientific pipeline is never embedded in the UI. The frontend renders; the backend delegates to the frozen artifacts; the artifacts came from the scripts in NEW_START_RESULTS/09_Scripts/.

Live vs static mode

GitHub Pages serves static files only, so the Python model cannot run there. The frontend detects this and says so plainly — it never fakes a prediction. All cohort data, evaluations and figures remain fully available; only the upload workflow requires the local API.

---

Environment variables

Variable	Default	Purpose
HELIXA_RESULTS	../../NEW_START_RESULTS	Where the scientific outputs live
HELIXA_CORS	localhost:5173,127.0.0.1:5173	Allowed browser origins
HELIXA_DATA	../frontend/data	JSON dataset directory

Copy website/backend/.env.example to .env and edit. Never commit .env.

---

Known limitations

These are stated on the platform itself, not buried here.

1. No external patient cohort has been tested. Everything derives from the same 328-patient TCGA/GDC cohort. Generalisation to another hospital or study is unproven.
2. No clinical outcome has been linked. Survival data has not been analysed, so these are molecular groupings — not validated prognostic classes.
3. This is a computational classification, not a diagnosis.
4. Agreement figures measure learnability, not diagnostic accuracy. The classifier learns to reproduce labels derived by clustering on this same cohort, so those labels are not independent ground truth. The leave-one-out and permutation tests prove there is no leakage — but this measures "the subtype boundaries are cleanly learnable", not clinical diagnostic accuracy. This is also why the reported probability is tied to the consensus profile rather than to the labels: calibrating against the labels would be circular.
5. Two subtypes (CL, INT) are only partially supported by independent sources, and are labelled as such everywhere they appear.
6. Verhaak 2010 is not an independent validator here — it was used inside the clustering optimisation. Only Neftel 2019, Garofano 2021, MSigDB Hallmark and the lineage panels are independent.
7. The biomarker/drug-target panel lists computational candidates, not treatments. Nothing on it has been tested in a laboratory by this project.
8. The dependency evidence comes from cell lines, not patients, and cell lines were not classified into the 6 subtypes (no DepMap expression data was available to do so) — so a gene's dependency evidence is "selective to glioblastoma", not "selective to this exact subtype". Cell lines also lack an immune system, a blood-brain barrier and a tumour microenvironment, which may specifically understate the MES (immune-infiltrated) panel.
9. The literature check covered the top ~10 ranked candidates per subtype (61 genes total), not all 359 that passed the statistical gates. A gene ranked just below the cutoff was never checked, however promising the raw numbers.
10. Two of the tier-1 targets already have drugs that failed in glioblastoma trials (cilengitide/ITGB5, palbociclib/CDK6) — the target may still be biologically valid; that specific molecule was not effective in that trial. This is recorded on their row, not hidden.

Next step that would matter most: link these subtypes to survival data from GDC and run a Kaplan-Meier analysis; and, for the biomarker panel, download DepMap's cell-line expression data to classify the 53 glioblastoma lines into the 6 subtypes directly, turning "selective to glioblastoma" into "selective to this exact subtype."

---

References

• Neftel C. et al. An Integrative Model of Cellular States, Plasticity, and Genetics for Glioblastoma. Cell 2019;178(4):835-849.
• Garofano L. et al. Pathway-based classification of glioblastoma uncovers a mitochondrial subtype with therapeutic vulnerabilities. Nature Cancer 2021;2:141-156.
• Verhaak RGW et al. Integrated genomic analysis identifies clinically relevant subtypes of glioblastoma. Cancer Cell 2010;17(1):98-110.
• Liberzon A. et al. The Molecular Signatures Database Hallmark gene set collection. Cell Systems 2015;1(6):417-425.
• DepMap, Broad Institute. Cancer Dependency Map — genome-wide CRISPR gene-dependency screening across 1,208+ cancer cell lines. depmap.org
• Every gene-level claim in the biomarker panel carries its own source URL (PubMed, Human Protein Atlas, Open Targets, DrugBank, or ClinicalTrials.gov) in BIOMARKER_ENGINE/03_Evidence/literature_evidence.json and in every patient report.

---

Team — KBU-MedLab

Abdulla ASKAR — Team Leader	B.Sc. Biomedical Engineering — KBÜ · B.Sc. Electrical and Electronics Engineering — KBÜ · M.Sc. Computer Science — İTÜ
Mohammad Alsagher — Literature	Medical Engineering
Leyan Mamou — Algorithms & Technical Development	Computer Engineering
Alı ALTHAHER — Algorithms & Other Technical Development	B.Sc. Biomedical Engineering — KBÜ
Mariam Mariam — Media & Other Contributions	Medical Engineering

Contact · [email protected] · 0 538 314 9253

---

HELIXA — Turning Genomics into Decisions KBU-MedLab