AI Development Company in Boston

Biotech / Precision Medicine

A Cambridge-based precision oncology company (Series B, $85 million raised, partnered with 3 top-10 US cancer centres including Dana-Farber) was developing a companion diagnostic — an AI system that predicts which cancer patients will respond to a novel immunotherapy based on tumour genomic profiling. The scientific challenge: only 22% of patients responded to the therapy in the Phase II trial (a typical response rate for immunotherapies). If a biomarker could identify likely responders, the Phase III trial could enrol an enriched population — dramatically improving the trial's probability of success and reducing the number of patients needed (from 800 to 300, saving $120 million in trial costs and 18 months of time). The company had a computational biology team that had identified 12 candidate biomarkers using traditional statistical methods (differential gene expression, pathway analysis). But single-biomarker approaches achieved only 58% accuracy in predicting response (AUC 0.64) — insufficient for clinical utility. The FDA requires companion diagnostics to demonstrate clinical validity: the biomarker must reliably distinguish responders from non-responders, with performance that changes the clinical decision. An AUC of 0.64 would not meet this standard.

We built a multi-omic biomarker discovery and validation pipeline that identifies a composite biomarker signature using machine learning. (1) Multi-omic data integration: the training data comprised 480 patients from the Phase II trial, each with: whole-exome sequencing (20,000+ genes, identifying mutations — the raw feature space is enormous), RNA-seq transcriptomics (gene expression levels for 20,000+ genes — capturing which genes are active in each tumour), tumour mutational burden (TMB — the total number of mutations, a known immunotherapy predictor), PD-L1 expression (a protein biomarker already used for immunotherapy selection — but with limited accuracy alone), tumour microenvironment composition (estimated from RNA-seq — the proportions of different immune cell types infiltrating the tumour), and clinical features (cancer type, stage, prior treatments, age, ECOG performance status). The total feature space: 60,000+ features per patient. With 480 patients, this is a classic high-dimensional, low-sample-size problem — the exact challenge that biomedical ML must solve. (2) Feature selection and dimensionality reduction: we implemented a three-stage feature selection pipeline. Stage 1: biological filtering — removing features with no biological plausibility for immunotherapy response (reducing 60,000 to ~8,000 features based on domain knowledge and published literature). Stage 2: statistical filtering — univariate analysis with multiple testing correction (Benjamini-Hochberg) to identify features individually associated with response (reducing 8,000 to ~400). Stage 3: ML-based feature selection — recursive feature elimination with cross-validation using a gradient-boosted ensemble (XGBoost), identifying the minimal feature set that maximises prediction accuracy. Result: a 47-feature composite biomarker signature spanning genomic mutations (12 genes), gene expression patterns (28 genes), TMB, PD-L1, and 5 tumour microenvironment features. (3) Prediction model: an ensemble combining XGBoost and a regularised logistic regression (the latter for interpretability — FDA reviewers prefer models where the contribution of each feature can be clearly understood). The model outputs a response probability score (0-1) for each patient. We used nested cross-validation (outer loop for performance estimation, inner loop for hyperparameter tuning) to provide unbiased performance estimates — critical for FDA submission where overfitting claims are closely scrutinised. (4) Validation strategy: the 480 patients were split into discovery (n=320) and validation (n=160) cohorts. The model was developed entirely on the discovery cohort and evaluated on the held-out validation cohort with no data leakage. Additionally, the company had access to an independent external cohort (n=120 patients from a separate clinical trial at a different cancer centre) for external validation — the gold standard for biomarker validation. (5) FDA pre-submission package: we prepared a comprehensive analytical validation package: intended use statement (the biomarker identifies patients likely to respond to the specific immunotherapy), analytical performance (sensitivity, specificity, PPV, NPV with 95% confidence intervals), clinical validity (evidence that the biomarker score correlates with clinical response), reproducibility (replicate testing showing consistent results from the same tumour samples), and software documentation (21 CFR Part 11-compliant development process, validated software, audit trails for all model development decisions).

Biomarker performance — discovery cohort: AUC 0.82, sensitivity 78%, specificity 81% (from single-biomarker AUC 0.64 — 28% improvement). The 47-feature composite signature dramatically outperformed any single biomarker. Validation cohort: AUC 0.79, sensitivity 74%, specificity 78% — the modest drop from discovery to validation demonstrates the model was not overfit (a critical FDA concern). External validation: AUC 0.77, sensitivity 71%, specificity 76% — performance held up in an independent patient population from a different clinical centre, demonstrating generalisability. Clinical utility: using the biomarker to enrich the Phase III trial population, the predicted response rate increases from 22% (unselected) to 54% (biomarker-positive) — a 2.5x improvement that transforms the trial design. Phase III trial impact: enriched trial requires 300 patients (from 800 unselected) — saving $120 million in direct trial costs and 18 months of enrolment time. The probability of Phase III success increases from ~45% (unselected, based on the 22% response rate) to ~78% (enriched, based on the 54% predicted response rate). FDA pre-submission: the FDA pre-submission meeting received positive feedback — the agency confirmed the analytical validation approach was appropriate and the clinical validity evidence was compelling. The companion diagnostic is expected to be reviewed alongside the Phase III BLA (Biologics License Application). Commercial value: if the drug is approved with the companion diagnostic, the biomarker test generates $50-80 million in annual diagnostic revenue (at $3,500 per test, 15,000-23,000 tests annually in the indicated population). The drug itself, enabled by the enrichment strategy, has a projected peak revenue of $1.2 billion. Total programme value: the biomarker enables a drug programme worth $1.2 billion in peak annual revenue. Development investment: $520,000 over 20 weeks. The biomarker AI represents perhaps the highest-leverage AI investment in any industry — a $520K investment that de-risks a multi-billion-dollar drug programme.