From Biological Systems to Human Performance: How ML Pipelines Built for Genomics Transfer to Sports Science
The Pipeline Is the Same. The Organism Changed.
A quiet convergence is happening between computational biology and sports performance science. On one hand, bioinformatics labs routinely build machine learning pipelines that ingest high-dimensional biological data — gene expression profiles, protein sequences, growth rate measurements — and extract predictive features that explain complex physiological processes. On the other hand, sports science teams are drowning in multivariate time series from wearables, force plates, metabolic carts, and GPS trackers, and they need exactly the same thing: models that take noisy biological signals and predict outcomes that matter.
The methods are the same. The organism changed.
I've spent the past two years building ML models that predict physiological outcomes in biological systems — Random Forest classifiers, neural network regressors, SHAP interpretability pipelines — all across datasets spanning over thousands of species. This article is about what I've learned that transfers directly to sports performance modeling, and where the sports domain introduces problems that genomics rarely has to solve.
Feature Engineering on Biological Data: What Genomics Teaches You
The core problem in my research lab is this: given a set of molecular features for an organism (codon usage frequencies, gene presence/absence, sequence-level statistics), can we predict a physiological outcome (whether a yeast species can grow at human body temperature or at a certain pH)?
This is structurally identical to a sports performance question: given a set of biometric features for an athlete (HRV trends, sleep architecture, training load history, metabolic markers), can we predict a physiological outcome (injury risk, recovery timeline, performance readiness)?
In both cases, the feature space is large, correlated, and noisy. In both cases, the signal-to-noise ratio is low. And in both cases, naive approaches produce overfit models.
What transfers directly:
Codon optimization → feature encoding for biological signals.
In genomics, raw nucleotide sequences are meaningless to an ML model until you encode them as features that capture biologically meaningful variation. Codon usage bias, GC content, codon adaptation index — these are engineered features that compress sequence information into dimensions a model can learn from. The same logic applies to wearable data. Raw heart rate is less useful than HRV (the variability between beats), which is less useful than HRV trend slopes over rolling windows, which is less useful than the ratio of sympathetic to parasympathetic activation derived from frequency-domain analysis of HRV. Each transformation compresses raw signal into biologically meaningful features. The skill is knowing which transformations preserve the physiology of interest.
SHAP interpretability → trust in high-stakes decisions.
One of the most important things I built in my research pipeline was a SHAP (SHapley Additive exPlanations) layer that explains why a model makes a given prediction. In genomics, this matters because biologists won't trust a black box. In sports, the stakes are higher: a coach won't bench a starter because an algorithm said so. They'll bench a starter because the algorithm showed them that this athlete's HRV recovery curve hasn't returned to baseline in 72 hours, their sleep efficiency dropped below 80% for three consecutive nights, and their ground contact time asymmetry increased by 12% — and here's how each of those features contributed to the injury risk score. Interpretability isn't a nice-to-have in sports performance, it's a requirement.
Dimensionality reduction → finding structure in multivariate physiological data.
In my recent technical work, I used Non-negative Matrix Factorization (NMF) to stratify patients into clinically meaningful subgroups based on high-dimensional molecular data. The exact same approach applies to athlete profiling. Given 30+ biometric channels from a wearable, NMF or PCA can identify latent "recovery profiles" or "fatigue signatures" that cluster athletes into actionable groups based on full multivariate patterns.
Where Sports Performance Breaks the Genomics Mold
If the pipelines were identical, every bioinformatician would already be working in sports tech. They're not, because the sports domain introduces three problems that genomics rarely encounters at the same intensity.
The temporal dimension
Genomic data is mostly static. A genome doesn't change between Tuesday and Friday. An athlete's physiological state changes hour by hour. This means sports performance models must handle time series natively — not as a table of features, but as sequences with temporal dependencies, periodicity, and non-stationarity.
This is where architectures like LSTMs, temporal convolutional networks, and transformer-based sequence models become essential. A Random Forest can tell you which features matter. It can't tell you that the trajectory of those features over the past 14 days is what predicts next week's injury. For that, you need models that reason about sequences, and you need feature engineering that preserves temporal structure: rolling statistics, lag features, rate-of-change calculations, and circadian alignment metrics.
Real-time inference constraints
In genomics, inference can take hours. Nobody needs a gene expression prediction in under 100 milliseconds. In sports, real-time matters. A wearable needs to compute a recovery score on-device or with minimal cloud latency. A coaching dashboard needs to update mid-session. This pushes the engineering toward lightweight models (gradient boosted trees, distilled neural networks), edge deployment, and careful attention to the cost-accuracy tradeoff that most research environments ignore.
The n=1 problem
Genomic models are trained on populations. A Random Forest predicting thermotolerance across 1,100 species generalizes because the feature space is shared and the outcome is binary. In sports performance, the most valuable predictions are often individualized. Athlete A's HRV recovery curve after a high-intensity session looks nothing like Athlete B's, even if their baseline HRV is similar. This means population models need personalization layers — transfer learning from a general model fine-tuned on individual history, or hierarchical Bayesian approaches that share strength across athletes while estimating individual parameters.
The Circadian Layer: An Undermodeled Dimension
One area where I believe the field is significantly underinvesting is circadian rhythm integration into performance models.
Most wearable platforms treat sleep as a single nightly score. But sleep is not a monolith — it's a sequence of ultradian cycles with distinct physiological functions (slow-wave sleep for tissue repair, REM for cognitive consolidation), and the timing of sleep relative to the athlete's intrinsic circadian phase matters enormously for recovery outcomes.
Circadian misalignment — training or competing at a time that conflicts with an athlete's chronotype — is a measurable, modelable performance drag. Core body temperature, cortisol rhythms, reaction time, and muscle force production all follow circadian curves that vary by individual. A model that predicts performance readiness without accounting for circadian phase is leaving signal on the table.
The data infrastructure for this exists. Continuous wrist temperature from Oura or WHOOP approximates circadian phase. Skin conductance and HRV provide sympathetic/parasympathetic balance that tracks with the cortisol awakening response. The modeling challenge is integrating these slow-moving circadian features with fast-moving training load features into a unified prediction — and that's a feature engineering and architecture problem, not a data availability problem.
Building a Practical Sports Performance ML Stack
For anyone reading this who wants to build in this space, here's the stack I'd recommend based on what I've learned building production-grade ML systems:
Data layer - Apple HealthKit and Google Health Connect as aggregation APIs. Direct integrations with WHOOP, Oura, Garmin, and Polar for raw sensor data. PostgreSQL for structured athlete profiles; TimescaleDB or InfluxDB for time series storage. Everything timestamped in UTC with timezone metadata preserved.
Feature engineering - This is where the genomics training pays off. Build a feature store that computes rolling windows (7-day, 14-day, 28-day) over every raw channel: mean, variance, slope, min, max, entropy, etc. Add derived features: acute-to-chronic workload ratio (ACWR), HRV coefficient of variation, sleep efficiency trend, training monotony, strain indices, etc. Circadian features: estimated circadian phase from wrist temperature, light exposure timing, sleep midpoint deviation from 14-day average, etc.
Model layer - Start with XGBoost for tabular predictions (injury risk, recovery score, readiness). It's fast, interpretable with SHAP, and handles missing data gracefully — which matters because wearable data is never complete. Add an LSTM or Temporal Fusion Transformer for sequence-dependent predictions (performance trajectory over the next training block). Keep models small enough for near-real-time inference.
Interpretability layer - Non-negotiable. SHAP values on every prediction. Force plots for individual athlete explanations. Feature importance dashboards for coaching staff. The model that gets deployed is the model the coach trusts, and trust comes from transparency.
Personalization layer - Population model as the base. Fine-tuning on individual athlete data after a 30-day onboarding window. Bayesian updating for athletes with limited history. Explicit handling of the cold-start problem — new athletes get population priors that shift toward individualized predictions as data accumulates.
Where This Is Going
The sports performance ML space is at an inflection point. The hardware layer (wearables, sensors, force plates) has matured. The data volume is there. What's missing is the modeling sophistication that computational biology has been developing for two decades — and the engineering discipline to deploy those models reliably at scale.
The companies that will win this space aren't the ones with the best sensors. They're the ones with the best models and the best feature engineering — the ones that can take noisy, multivariate, time-dependent biological data and turn it into predictions that coaches and athletes actually use.
That's a bioinformatics problem.
Taylor Powell is an ML engineer and M.S. candidate in Bioinformatics (Computer Science) at UNC Charlotte, with prior training in kinesiology, human physiology, and neuroscience. He builds predictive models on biological systems and writes about ML infrastructure at Teapot Commons. He trains MMA, competes in Hyrox, and is preparing for a November marathon. He can be reached at Taylorp661@gmail.com.