Machine learning
Tree ensembles that capture thresholds, interactions and messy plant data.
- XGBoost
- LightGBM
- Gradient boosting
- Random forest
- Extra trees
BIOMEDANA MODEL LIBRARY
2,500+ bioprocess models. One governed library.
From Raman spectra to off-gas, from cell culture to small-molecule API: a ready-to-train library of soft sensors and AI models built for bioprocessing.
2500+ready-to-train model configurations
139soft sensors across the bioprocess
18machine-learning model families
18measurement technologies
What is the BioMedAna Model Library?
The BioMedAna Model Library is a ready-to-train collection of bioprocess AI models: 139 soft sensors that can each be trained with 18 machine-learning model families, plus process and analysis models, for more than 2,500 model configurations. Each model is trained on your own process data, competes against the alternatives on runs it has not seen, and is approved by a reviewer before use.
01 — MODEL FAMILIES
Every modern method.
Ready for your process.
Gradient boosting to deep learning, chemometrics to state estimation. BioMedAna brings the methods data scientists trust to every soft sensor, with no code.
Skip card journey ↓Seven families of models01 / 07 ↓ SCROLL TO EXPLORE 01 02 03 04 05 06 07
Deep learning
Neural networks that read how each signal has been moving, not just where it is now.
- LSTM networks
- Temporal convolutional networks
Probabilistic & kernel
Smooth non-linear models that know how sure they are.
- Gaussian processes
- Support vector regression
- k-nearest neighbours
Chemometrics
The language of PAT: full spectra in, clear and explainable estimates out.
- PLS on full spectra
- Ridge
- Elastic net
- Lasso
Time series
Models that use the recent past, including the last lab result, to see what comes next.
- Autoregressive (ARX)
- Lagged-signal models
State estimation
Estimates that correct themselves every time a new assay arrives.
- Kalman filtering
- Moving-horizon estimation
Hybrid & mechanistic
Biology and physics first, with data closing the gap.
- Hybrid CHO kinetics
- Reaction kinetics
- Fermentation & vaccine twins
02 — FROM DATA TO DEPLOYED MODEL
Five steps.
No notebooks.
The whole model lifecycle runs inside BioMedAna, from the first data import to daily monitoring.
- 01
Connect
BioMedAna Hub brings in bioreactor, PAT, historian and lab data and lines up every signal with its assay.
- 02
Choose
Pick a soft sensor from the library. The app shows which ones your data can already support.
- 03
Compete
Every suitable model family trains on your runs and is scored on runs it has never seen.
- 04
Approve
The best model goes to an independent reviewer with its evidence attached.
- 05
Monitor
Deployed models are checked against new lab results every day and flag drift.
03 — WHAT YOU CAN PREDICT
From the first feed
to the final vial.
A sample of what the library's soft sensors estimate, across the process.
Upstream
- Viable cell density & viability
- Glucose, lactate, glutamine, ammonia
- Amino acids & osmolality
- Product titer & harvest timing
- Glycans & charge variants
- Oxygen uptake & limitation
Downstream
- Protein A breakthrough & yield
- HCP & DNA clearance
- Aggregates & pool purity
- Dynamic binding capacity
- TFF flux & UF/DF endpoint
- Viral clearance
Drug product & advanced modalities
- Formulated concentration
- Lyophilization moisture & endpoint
- LNP size & encapsulation
- mRNA integrity
- AAV vector & capsid titer
- Vaccine antigen & infectivity
Small-molecule API
- Reaction conversion & impurities
- Calorimetric conversion
- Unreacted reagent accumulation
- Reaction endpoint
- Polymorph & crystal size
- Drying endpoint
04 — ACROSS MODALITIES
139 soft sensors.
Five modalities.
Upstream, downstream and drug product, with sensors designed around each modality's critical questions.
Monoclonal antibodies & recombinant proteins
Titer, metabolites, amino acids, glycans, charge variants, aggregates, purification and fill-finish.
Microbial fermentation
Biomass, by-products, oxygen limitation, product titer and harvest.
Viral vaccines & mRNA
Infection progress, antigen, particle and LNP quality.
Cell & gene therapy
Vector titers, capsid quality and potency risk.
Small-molecule API
Conversion, impurities, calorimetry, crystallization and drying.
05 — WHY A LIBRARY
Stop starting
from scratch.
From scratch
BioMedAna Model Library
Starting point
A blank notebook
139 bioprocess soft sensors, ready to configure
Model choice
One algorithm, picked by habit
18 model families compete on unseen runs
Spectra & PAT
Custom preprocessing per project
Full spectra feed models directly
Lab results
Used once, for training
Used continuously to correct estimates
Governance
Files and emails
Review, approval, deployment and daily monitoring
06 — BEST MODEL WINS
Thousands of options.
One trusted answer.
See sooner
Continuous estimates between lab samples, so changes show up while there is still time to act.
Sample smarter
Spend analytical effort where it matters, with soft sensors filling the gaps in between.
Start from a library
Begin from a proven sensor definition instead of a blank notebook.
Stay in control
Versioned, fingerprinted models with review, rollback and a full audit trail.
Part of the BioMedAna platform: data from Hub, models in Twin, questions through Assist & Agents, governance in OS.
07 — QUESTIONS
Common questions about the model library
Short answers on what is included, how models are chosen and what you need to get started.
Ask us anything else ↗How many models does BioMedAna have?
More than 2,500 ready-to-train model configurations: 139 soft sensors combined with 18 model families, plus process twin and analysis models.
Which machine-learning models are included?
XGBoost, LightGBM, gradient boosting, random forests, extra trees, Gaussian processes, support vector regression, k-nearest neighbours, LSTM and temporal convolutional networks, PLS, ridge, elastic net and lasso, autoregressive models, Kalman filtering and moving-horizon estimation.
Can BioMedAna build soft sensors from Raman or NIR spectra?
Yes. A full Raman, NIR, FTIR or UV spectrum can feed a soft sensor directly, alongside process signals such as off-gas, pH and feeds.
Which bioprocess modalities are covered?
Monoclonal antibodies and recombinant proteins, microbial fermentation, viral vaccines and mRNA, cell and gene therapy, and small-molecule API, across upstream, downstream and drug product.
How does BioMedAna choose the best model?
Every suitable model family is trained and scored on runs it has never seen. The most accurate model is kept, a simpler one wins when it is nearly as good, and a reviewer approves it before deployment.
Do I need data scientists to use it?
No. Scientists pick a sensor, map their data and train it from the app. The model comparison, validation and approval steps are built in.
Do the models work without my data?
No. The library provides ready-to-train configurations; each model is trained and qualified on your own process and laboratory data for its intended use.
Configurations combine catalog soft sensors with the model families that suit them, plus process and analysis models. Each model is trained and qualified on your own process data before use. Illustrations are conceptual.