Zama.ai Bounty for Season 8: Implement an FHE-based Biological Age and Aging Pace Estimation ML Model Using Zama Libraries

• CpG group := (Cytosine - phosphate - Guanine)
  • Notable for role in gene regulation through methylation processes

• Just polynomials of degree 1, very efficient representation in FHE applications

• Manhattan plot

• HumanMethylation450 BeadChip (released in 2008)
• Targets over 450k sites across the human genome

Human arrays

1. Models trained to predict chronological age of tissue based on biomarkers
2. Delta betwen chronological age and real age used as marker to predict
   • Mortality risk
   • Disease states
   • etc.

• Inputs (CgPs)
• Outputs (Predicted chronological age)

1. Start with simpler models (linear regression-based clocks) as easier to implement in FHE
2. Balance accuracy vs compute complexity - some models might use hundreds of CpG sites that would be expensive in FHE
3. Horvath clock well established but uses elastic net regression with many features
4. DunedinPACE measures aging pace rather than biological age, which might be interesting but more complex

1. Linear models most straightforward
2. Avoid non-linear activation functions, if possible
3. Consider feature count - Less CpG sites means faster FHE computation
4. Some biological clocks use relatively few CpG sites (~10-50) which would be ideal (NOTE: Need to validate this)

1. Start with Concrete ML for higher-level abstractions
2. Need to quantise model (using brevitas-nn / concrete-ml, etc.)
3. Benchmark acc between original + FHE - expect precision loss (though this depends on implementation / model, etc.)
4. Reduce multiplicative depth

• Use concrete's compiler to analyse circuit depth / bottlenecks
• Precision vs performance trade-off (this one is likely key)
• Heavily consider preprocessing strategies pre-encryption to offload computation

1. Client: Encrypts methylation data
2. Server: Processes encrypted data without decryption
3. Client: Receives and decrypts the predicted biological age

• Demo + sample data

27k_reference: probeAnnotation21kdatMethUsed
CBL_common: coefs
CBL_specific: coefs
Cortex_common: coefs
DunedinPACE: coefs gold_standard_means
HannumG2013: coefs
HorvathS2013: coefs
HorvathS2018: coefs
LevineM2018: coefs
LuA2019: coefs
McEwenL2019: coefs
ShirebyG2020: coefs
YangZ2016: epiTOCcpgs
ZhangQ2019: coefs
ZhangY2017: coefs
subGSE174422: betas info