NMDP [002 SP_B]

Vidhi Lalchand, Ph.D.

IMU Biosciences

13th Feb, 2026

[203 donors]

Data Semantics & Processing

NMDP 002 SP_B Combined Ratios: Tv5 Panel

Columns dropped due to NaNs for significant cross-section of donors:

    clean_tv_num = clean_tv_num.loc[:, clean_tv_num.nunique(dropna=False) > 1] # Drop columns with only one unique value
    clean_tv_num = clean_tv_num.drop(columns=["aTreg_Tv5", "mTreg_Tv5", "rTreg_Tv5"])

Log1p + Standardization

Batch effect + Noise columns

Off-the-shelf algorithms: 10-CV sweep

Algorithm	% Acc	AUC	% Pos. class
Kernel Logistic Reg.	0.5821 +/- 0.133	~0.5224	0.451 +/ 0.2
Logistic Reg.	0.5664 +/- 0.214	~0.5219	0.343 +/- 0.2
RF Estimator	0.665 +/- 0.077	~0.517	< 0.20
MLP + Dropout	0.579 +/- 0.053	~0.58	0.455 +/ 0.2
Boosting + XTrees	0.670 +/- 0.042	~0.530	< 0.20
SVC	0.606 +/- 0.167	0.511	< 0.20
GradientBoosting	0.626 +/- 0.099	~0.49	< 0.20

Goal: Predicting the binary response variable of chronic relapse {0,1} from the megatables (ratios). Unseen metrics with std. error of the mean (metric) for stable config:

Key failure mode: The fraction of chronic relapse donors correctly picked up by the model.

The AUC and % Accuracy obscure positive recall - which is the key thing to look out for here.

MLP + Dropout / w. imbalanced class weighting

w_1 = \frac{N}{2N_1}, \qquad w_0 = \frac{N}{2N_0}.

SP_A - Validation donors

SP_B - Validation donors

Reserved if time