Decoding heal h om NMR spec a:
machine lea ning models o me abolic heal h
Alain Ibáñez de Opakuaa,b, Rubén Gil-Redondob, Maide Bizka guenagab, Ángela de Diegob, Rica do Condeb, Tammo Die cksb,
Bea iz González-Valleb, Nie es Embadeb, José Ma ía Ma ob, Osca Mille a,b
In oduc ion
Me hods
Resul s
Me abolic age
Disease classi ica ion
Conclusions
Re .
Me abolism o e s a ich and dynamic window in o human heal h, e lec ing bo h physiological balance and pa hological dis up ions. Nuclea Magne ic Resonance (NMR) spec oscopy, wi h i s
ep oducibili y and non-des uc i e na u e, p o ides a powe ul pla o m o me abolic p o iling. In ecen yea s, he in eg a ion o machine lea ning echniques wi h NMR da a has opened
new a enues o deciphe ing complex biochemical signa u es associa ed wi h aging and disease. In his s udy, we in oduce a comp ehensi e compu a ional amewo k designed o ex ac clinically
ele an insigh s om NMR me abolomics da a. By quan i ying me aboli e concen a ions om J- esol ed spec a and de i ing clinical pa ame e s om 1D ¹H NOESY spec a, we
gene a e a mul i-laye ed ea u e se ha cap u es di e se aspec s o me abolic heal h. These ea u es a e subsequen ly used o ain p edic i e models aimed a es ima ing biological age and
classi ying disease s a es. Ou app oach achie es high accu acy in age p edic ion, enabling he de ec ion o indi iduals whose me abolic p o iles de ia e om no ma i e ajec o ies—po en ially
signaling accele a ed o decele a ed aging. Fu he mo e, we demons a e he u ili y o ou models in dis inguishing be ween mul iple heal h and disease condi ions, highligh ing he po en ial o
NMR-based machine lea ning as a scalable and in e p e able ool o p ecision heal h moni o ing.
1Measu ing Biological Age ia Me abonomics: The Me abolic Age Sco e. He el e al. J P o eome Res, 2016.
2Me abolic age based on he BBMRI-NL 1H-NMR me abolomics eposi o y as bioma ke o age- ela ed disease. Van den Akke e al. Ci c Genom P ecis Med, 2020.
3NMR me abolomic modelling o age and li espan: a mul i-coho analysis. Lau e al. medRxi , 2023.
¹H-NMR spec a we e acqui ed om o e 30,000
se um samples using bo h 1D NOESY and as -
acquisi ion 2D J- esol ed (JRES) expe imen s. The 2D
JRES spec a we e used o quan i y 49 me aboli es
h ough a dedica ed pipeline designed o minimize
signal o e lap. In pa allel, 1D NOESY spec a we e
used o es ima e 25 clinical pa ame e s ia
supe ised eg ession models, co e ing bo h di ec ly
obse able ma ke s (e.g., CRP, albumin) and in e ed
physiological indices (e.g., calcium, eGFR). While
models based on aw 1D spec a gene ally o e
highe p edic i e pe o mance, he use o de i ed
ea u es enhances explainabili y and suppo s
downs eam clinical in e p e a ion.
To es ima e ch onological age— e e ed o as me abolic age—, we
de eloped a s acking ensemble machine lea ning model ained on a
subse o ~8,000 indi iduals selec ed om an ini ial popula ion o
~28,000. This educ ion was pe o med o ensu e a uni o m
dis ibu ion ac oss he age ange and o mi iga e eg ession o he
mean e ec s du ing model aining.
The model based on 1D NOESY spec a shows a s ong co ela ion be ween
ch onological and me abolic age (R=0.92), ou pe o ming p e ious NMR-based
me abolic age models1,2,3, which didn' each 0.8. Consis en esul s we e ob ained
wi h o he NMR da ase s: CPMG (R=0.91), NOESY FID (R=0.87), and quan i ied
me aboli es plus clinical pa ame e s (R=0.88).
Addi ionally, dis ibu ions o me abolic dis o ion
( he di e ence be ween me abolic and
ch onological age) show signi ican al e a ions
ac oss disease g oups. Fo example, indi iduals
wi h p os a e cance exhibi a posi i e
dis o ion, sugges ing accele a ed me abolic
aging. In con as , pa ien s wi h li e diseases
display a b oade dis ibu ion, likely due o he
he e ogeneous na u e o li e condi ions, each
associa ed wi h dis inc me abolic consequences.
These indings unde sco e he u ili y o me abolic
age as a bioma ke capable o cap u ing disease-
ela ed physiological al e a ions.
Using quan i ied me aboli es and p edic ed clinical pa ame e s, we ained a mul iclass XGBoos classi ie o
di e en ia e be ween nine heal h ca ego ies. The con usion ma ix (colo s no malized pe ow) shows
excellen o e all pe o mance wi h an accu acy o 0.80. Misclassi ica ions occu p edominan ly be ween
expec ed neighbo ing classes, such as younge s old adul s (wi h he h eshold se a 50 yea s), olde
adul s s long CoVid pa ien s (due o mild esidual e ec s obse ed in long CoVid), and olde adul s s
me abolic synd ome (a condi ion closely associa ed wi h aging).
ROC cu es o each class
demons a e obus disc imina ion,
wi h AUC alues anging om 0.91
o 1.00, con i ming he model’s
high sensi i i y and speci ici y
ac oss condi ions.
SHAP alues e ealed ha a iables
associa ed wi h sys emic
in lamma ion and me abolic
s abili y a e among he mos
ele an p edic o s o me abolic age.
GlycA and e y h ocy e sedimen a ion a e showed
he s onges posi i e con ibu ions, consis en
wi h hei oles as ma ke s o in lamma ion.
In con as , GlycB and albumin con ibu ed
nega i ely, likely e lec ing lowe ch onic
in lamma ion and be e nu i ional s a us,
bo h associa ed wi h a younge me abolic p o ile.
49 quan i ied me aboli es:
25 p edic ed clinical pa eme e s:
Fo disease classi ica ion, we implemen ed a mul iclass model
based on eX eme G adien Boos ing (XGBoos ) o dis inguish nine
heal h ca ego ies: se en disease condi ions and wo age-de ined
heal hy g oups (young and olde ). The heal hy e e ence popula ion
was selec ed o ma ch he age and sex dis ibu ion o he combined
disease coho s, enabling pa ial class balance and imp o ing
gene alizabili y.
1,5-Anhyd oso bi ol, 2-Aminobu y ic acid, 2-Hyd oxybu y ic acid, 2-Oxoglu a ic acid, 3-Hyd oxybu y ic
acid, 3-Hyd oxyisobu y ic acid, Ace ic acid, Ace oace ic acid, Ace one, Alanine, A ginine, Aspa agine,
Aspa a e, Be aine, Choline, Ci ic acid, C ea ine, C ea inine, Cys ine, D-Galac ose, Dime hylamine,
Dime hylsul one, E hanol, Fo mic acid, Glucose, Glu amic acid, Glu amine, Glyce ol, Glycine, His idine,
Isoleucine, Lac ic acid, Leucine, Lysine, Me hanol, Me hionine, Myo-inosi ol, N,N-Dime hylglycine,
O ni hine, Phenylalanine, P oline, Py u ic acid, Sa cosine, Se ine, Succinic acid, Th eonine,
T ime hylamine-N-oxide, Ty osine, Valine
Albumin, Apolipop o ein B, Bili ubin, Calcium, C eac i e p o ein, E y h ocy e sedimen a ion a e,
E y h ocy es, Es ima ed Glome ula Fil a ion Ra e, F uc osamine, Glyc A, Glyc B, HDL choles e ol,
Hemoglobin, I on, LDL choles e ol, Leukocy es, Lipop o ein(a), Pla ele s, SPC, To al choles e ol, To al
p o ein, T ans e ine, T iglyce ides, U a e, U ea
Le e aging NMR-based me abolomics and in eg a ed machine lea ning pipelines enables obus es ima ion o me abolic age and accu a e classi ica ion o mul iple disease s a es. The combina ion o
aw spec al da a wi h de i ed me aboli e and clinical ea u es enhances bo h model pe o mance and in e p e abili y. Dis inc shi s in me abolic age and ea u e pa e ns ac oss disease g oups
demons a e he amewo k’s abili y o cap u e physiological he e ogenei y and ea ly signs o me abolic dys egula ion. These ad ances suppo he de elopmen o scalable, non-in asi e ools o
p ecision heal h moni o ing and unde sco e he ele ance o me abolic p o iling in pe sonalized medicine.
aATLAS Molecula Pha ma, Bizkaia Science and Technology Pa k, 48160 De io, Spain
bP ecision Medicine and Me abolism Labo a o y, CIC bioGUNE, Bizkaia Science and Technology Pa k, 48160 De io, Spain
[email p o ec ed]
Machine Lea ning
ML
1H (ppm)
1D-NOESY 1H
Fas 2D J-Resol ed 1H
Se um
samples
Me abolic
age
classi ica ion
Disease
a ge s
inpu
ou pu
Me aboli e
quan i ica ion
Es ima ed clinical
pa ame e s
Age Disease
Clinical
pa ame e s
Me ada a
Me aboli e
quan i ica ion
om J-RES
In ensi y (a. u.)
J-RES 1H (se ine) Gaussian i
Clinical pa ame e
p edic ion wi h ML
Measu ed alues Measu ed alues Measu ed alues
P edic ed alues
HDL choles e ol (mg/dl)
R=0.97
Albumin (g/dl)
R=0.94
ESR (mm/h)
R=0.78
R=0.92
RMSE=7.2 yea s
Ch onological age
Me abolic age
0.0±6.8
p- alue=
3.2E-30
14.5±11.1
0.0±8.2
p- alue=
1.0E-19
4.9±9.2
Li e
diseases
P os a e
cance
Indi idual epo
(p ecision medicine)
To showcase he po en ial o pe sonalized
diagnos ics, we p esen an indi idual-
le el epo . This includes a spide plo
displaying he p obabilis ic class
p edic ions, modula ed using a
empe a u e scaling ac o o 10 o
p o ide a smoo he ep esen a ion o
classi ica ion con idence (e.g., p obabili y
o 1.00 o p os a e cance in he
example). Below, a co esponding SHAP
alue plo iden i ies he mos in luen ial
ea u es d i ing he model’s decision.
Toge he , hese ou pu s illus a e he
capaci y o he amewo k o suppo
p ecision medicine applica ions based on
indi idualized me abolic pheno yping.
COVID-19
long COVID
Obse ed labels
P edic ed labels
COVID-19
long COVID
COVID-19
long COVID
COVID-19
long
COVID
