Comparative assessment of protein large language models for enzyme commission number prediction

Open Access
© The Au ho (s) 2025. Open Access This a icle is licensed unde a C ea i e Commons A ibu ion-NonComme cial-NoDe i a i es 4.0
In e na ional License, which pe mi s any non-comme cial use, sha ing, dis ibu ion and ep oduc ion in any medium o o ma , as long
as you gi e app op ia e c edi o he o iginal au ho (s) and he sou ce, p o ide a link o he C ea i e Commons licence, and indica e i you
modi ied he licensed ma e ial. You do no ha e pe mission unde his licence o sha e adap ed ma e ial de i ed om his a icle o pa s o
i . The images o o he hi d pa y ma e ial in his a icle a e included in he a icle’s C ea i e Commons licence, unless indica ed o he wise
in a c edi line o he ma e ial. I ma e ial is no included in he a icle’s C ea i e Commons licence and you in ended use is no pe mi ed
by s a u o y egula ion o exceeds he pe mi ed use, you will need o ob ain pe mission di ec ly om he copy igh holde . To iew a copy
o his licence, isi h p:// c ea i eco mmons. o g/ licen ses/ by- nc- nd/4. 0/.
RESEARCH
Capelae al. BMC Bioin o ma ics (2025) 26:68
h ps://doi.o g/10.1186/s12859-025-06081-9
BMC Bioin o ma ics
Compa a i e Assessmen o P o ein La ge
Language Models o Enzyme Commission
Numbe P edic ion
João Capela1*, Ma ia Zimme mann-Kogadee a2, Aal D. J. an Dijk3,4, Dick de Ridde 3, Osca Dias1,5 and
Miguel Rocha1,5
Abs ac
Backg ound: P o ein la ge language models (LLM) ha e been used o ex ac ep-
esen a ions o enzyme sequences o p edic hei unc ion, which is encoded
by enzyme commission (EC) numbe s. Howe e , a comp ehensi e compa ison o di -
e en LLMs o his ask is s ill lacking, lea ing ques ions abou hei ela i e pe o -
mance. Mo eo e , p o ein sequence alignmen s (e.g. BLASTp o DIAMOND) a e o en
combined wi h machine lea ning models o assign EC numbe s om homologous
enzymes, hus compensa ing o he sho comings o hese models’ p edic ions. In his
con ex , LLMs and sequence alignmen me hods ha e no been ex ensi ely compa ed
as indi idual p edic o s, aising unadd essed ques ions abou LLMs’ pe o mance
and limi a ions ela i e o he alignmen me hods. In his s udy, we se ou o assess
he pe o mance o ESM2, ESM1b, and P o BERT language models in hei abili y o p e-
dic EC numbe s, compa ing hem wi h BLASTp, agains each o he and agains mod-
els ha ely on one-ho encodings o amino acid sequences.
Resul s: Ou indings e eal ha combining hese LLMs wi h ully connec ed neu al
ne wo ks su passes he pe o mance o deep lea ning models ha ely on one-ho
encodings. Mo eo e , al hough BLASTp p o ided ma ginally be e esul s o e all, DL
models p o ide esul s ha complemen BLASTp’s, e ealing ha LLMs be e p e-
dic ce ain EC numbe s while BLASTp excels in p edic ing o he s. The ESM2 s ood
ou as he bes model among he LLMs es ed, p o iding mo e accu a e p edic ions
on di icul anno a ion asks and o enzymes wi hou homologs.
Conclusions: C ucially, his s udy demons a es ha LLMs s ill ha e o be imp o ed
o become he gold s anda d ool o e BLASTp in mains eam enzyme anno a ion
ou ines. On he o he hand, LLMs can p o ide good p edic ions o mo e di icul -
o-anno a e enzymes, pa icula ly when he iden i y be ween he que y sequence
and he e e ence da abase alls below 25%. Ou esul s ein o ce he claim ha BLASTp
and LLM models complemen each o he and can be mo e e ec i e when used
oge he .
Keywo ds: Enzyme anno a ion, Deep lea ning, La ge language models
*Co espondence:
[email protected]
1 Cen e o Biological
Enginee ing, Uni e si y o Minho,
B aga 4710-057, Po ugal
2 Genome Biology Uni ,
Eu opean Molecula Biology
Labo a o y, Heidelbe g, Ge many
3 Bioin o ma ics G oup,
Depa men o Plan Sciences,
Wageningen Uni e si y
and Resea ch, Wageningen, The
Ne he lands
4 Biosys ems Da a Analysis,
Uni e si y o Ams e dam,
Ams e dam, The Ne he lands
5 LABBELS - Associa e Labo a o y,
B aga/Guima ães, Po ugal
Page 2 o 21
Capelae al. BMC Bioin o ma ics (2025) 26:68
Backg ound
Enzymes a e na u al ca alys s o biochemical eac ions, accele a ing eac ion a es by
dec easing he ac i a ion ene gy equi ed o o m o b eak chemical bonds [1]. Enzymes
gene ally demons a e speci ici y o pa icula subs a es, al hough hey can be p omis-
cuous owa ds s uc u ally simila compounds [2]. Accu a ely and e icien ly p edic ing
hei unc ion is impo an o enhancing genome anno a ion wo k lows [3]. These p e-
dic ions a e c i ical o es ablishing connec ions be ween p o eins and eac ions, de is-
ing new pa hways, and gaining p ecise insigh s in o cellula me abolism.
Enzyme Commission (EC) numbe s o e a hie a chical amewo k, o ganising
enzyme unc ions. In he con ex o genome anno a ion p ocedu es, EC numbe s a e
assigned o p o ein sequences o succinc ly encapsula e pa e ns o chemical eac ions
[4]. These co espond o he chemical changes acili a ed by he enzyme linked o he
espec i e EC numbe .
Gi en he low p opo ion o enzymes assigned wi h EC numbe s [5], many ools ha e
eme ged o his anno a ion ask, le e aging simila i y and machine lea ning (ML) ech-
niques. The o me is based on he widely held idea ha enzymes sha ing high sequence
simila i y will likely ha e simila unc ions [6–8]. This anno a ion ans e echnique is
commonly adop ed using he gold s anda d ool BLASTp (Basic Local Alignmen Sea ch
Tool o p o ein sequences) [9]. Al hough i is he mos employed app oach in p ac ice,
i o e s no way o assigning a unc ion o p o eins wi h no homologous sequences.
Recen ly, ML algo i hms o he classi ica ion o enzymes ha lea n pa e ns and
make p edic ions om p o ein sequence da a gained a lo o a en ion [3, 8, 10–17]. In
many o hese s udies, he i s aim is o con e inpu samples, speci ically amino-acid
sequences, in o a nume ical o ma , e e ed o as ea u es. Va ious app oaches ha e
been employed o gene a e hese ea u es, including me hods based on homology [3, 8,
11, 12, 18], physicochemical p ope ies [8, 18], p esence o unc ional domains [8, 10],
amino-acid sequence p ope ies [12, 18], he aw amino-acid sequence [3, 8, 13], o he
usage o lea ned ep esen a ions om a p o ein language model [14, 16].
O e he yea s, se e al ML models we e de ised o lea n om hese nume ic ep e-
sen a ions. Be ween 2007 and 2018, he p edominan models included k Nea es Neigh-
bo s (k-NN) [10, 18], Random Fo es s (RF) [12], and Suppo Vec o Machines (SVM)
[18]. F om 2018 onwa d, he ocus shi ed o Deep Lea ning (DL) me hods, bes sui ed
o asks in ol ing la ge da a olumes, aligning wi h he apid expansion o da abases
o cu a ed enzymes [8, 19]. DEEP e [8], one o he ea lies DL models, ackles he issue
o nonuni o m ea u e dimensionali y, which a ises when ea u es depend on sequence
leng h, by de eloping a hyb id DL a chi ec u e. This combines Con olu ional Neu al
Ne wo ks (CNN) wi h Long-Sho Te m Memo y (LSTM) ne wo ks o lea n la en p o-
ein ep esen a ions ha depend on sequence leng h, like one-ho encoding and Posi-
ion-Speci ic Sco ing Ma ix (PSSM), and inco po a es a ully connec ed Deep Neu al
Ne wo k (DNN) o lea n leng h-independen ea u es, such as he p esence o P am [20]
unc ional domains. Ul ima ely, he la en ep esen a ions a e conca ena ed and p o-
cessed by a DNN o p edic he EC numbe . D-SPACE [21], ano he app oach based on
CNNs, was de eloped as a p o ein anno a o capable o pe o ming a ious asks using
a single model, including classi ying EC numbe s, Gene On ology (GO) e ms and p o-
ein amilies. DeepEC [3] was de eloped la e , combining CNNs wi h simila i y-based
Page 3 o 21
Capelae al. BMC Bioin o ma ics (2025) 26:68
sea ches pe o med wi h DIAMOND [22]. Mo e ecen ly, P o eIn e [13] was de eloped,
le e aging deep dila ed CNNs and da a pa allelism o ain wi h ull-leng h sequences o
amino acids and o e ing an in e p e a ion o he p edic ions h ough class ac i a ion
mapping.
Al hough hese DL models show g ea p omise by hemsel es, app oaches combin-
ing he model p edic ions and hose o o he app oaches, such as pai -wise alignmen s
wi h BLASTp [13] and DIAMOND [3, 17], o mul iple sequence alignmen s [14] we e
explo ed. The sys ema ic usage o hese app oaches sugges s ha DL models p o ide
sub-op imal esul s ha a e la e compensa ed by alignmen s’ p edic ions. The au ho s
o P o eIn e p o ided an analysis o he esul s o BLASTp and hose o DL models,
eaching he gene al conclusion ha BLASTp pe o ms sligh ly be e han P o eIn e
deep dila ed CNN, and an ensemble o he wo deli e s pe o mance su passing ha
achie ed by he indi idual echniques [13]. On he o he hand, CLEAN showcased
highe p edic i e capabili ies han BLASTp o independen small da ase s o unde s ud-
ied enzymes [16] by mi iga ing he unbalanced dis ibu ion o EC numbe s wi hin he
aining da ase h ough he applica ion o con as i e lea ning.
Since La ge Language Models (LLMs) ha e gained signi ican ac ion, p o ein LLMs
ha e been used o EC numbe p edic ion [14–17]. These models a e based on he ans-
o me ne wo k, which employs sel -a en ion mechanisms o weigh he impo ance o
di e en pa s o a sequence, enabling o cap u e long- ange dependencies.
Two ypes o p e- ained LLMs ha e been used o EC numbe p edic ion so a :
E olu iona y Scale Modeling (ESM) [23, 24] and P o BERT [25]. ESM models a e
ans o me ne wo ks p e- ained wi h UniP o KB da a [19] and we e used o ea u e
ex ac ion, in ol ing aking he ou pu s om a laye wi hin hese models and using hem
as ea u es (embeddings) o p edic ing he EC numbe [14, 16]. ESM models ha e been
p o en o be highly compe i i e amongs LLMs o p o ein ep esen a ion and unc ion
p edic ion [26]. On he o he hand, P o BERT is a ans o me ne wo k p e- ained wi h
da a om UniP o KB and he BFD da abase [27] and was ine- uned o EC numbe p e-
dic ion [15, 17]. Indeed, al hough many s udies ha e in es iga ed ESM and P o BERT
models o EC numbe p edic ion [14–17], a ho ough compa ison o hese LLMs as
ea u e ex ac o s has no ye been conduc ed. This lack o compa a i e analysis lea es
unanswe ed ques ions abou he ela i e e ec i eness o di e en LLMs in his speci ic
applica ion.
Fu he mo e, i is impo an o ecognize ha he me hodologies discussed in hese
s udies ypically do no demons a e hei pe o mance independen ly wi hou he aid
o alignmen s. Mo eo e , none o he abo e in es iga ions ha e ho oughly compa ed
LLMs o BLASTp on la ge es se s. This lea es open ques ions abou LLMs’ pe o -
mance agains BLASTp, hei po en ial ad an ages, and whe e LLMs migh all sho o
equi e u he de elopmen o mee o exceed BLASTp’s capabili ies.
This s udy p esen s an in-dep h e alua ion o ad anced p o ein ep esen a ions and
DL a chi ec u es o p edic ing EC numbe s. By add essing he no iceable gap in exis -
ing esea ch, we aim o cla i y hei ela i e e ec i eness. To his end, we ha e designed
a obus expe imen al amewo k o e alua e DL models ailo ed o EC numbe p e-
dic ion, ained wi h embeddings o ESM2, ESM1b, and P o BERT. Mo eo e , we p o-
ide implemen a ions o DeepEC and D-SPACE o compa ison. Since hese me hods
Page 4 o 21
Capelae al. BMC Bioin o ma ics (2025) 26:68
use one-ho encodings as inpu a he han LLM embeddings, hey o e a use ul way o
compa e di e en inpu ypes. Th ough his comp ehensi e app oach, we aim o c i i-
cally assess he pe o mance o bo h BLASTp and DL models, he eby con ibu ing al-
uable insigh s in o he e ec i eness o DL models, enhanced by s a e-o - he-a LLM
embeddings o one-ho encodings, compa ed o a adi ional, alignmen -based p o ein
unc ion p edic ion me hodology.
Ma e ials andme hods
De ining EC numbe classi ica ion p oblem
EC numbe p edic ion was de ined as a mul i-label classi ica ion p oblem inco po a ing
p omiscuous and mul i- unc ional enzymes (wi h mo e han one EC numbe ). Le X be
he se o p o ein sequences and Y be he se o EC numbe s. Each p o ein sequence
xi
in
X has an associa ed bina y label ec o
yi
o leng h |Y|, whe e |Y| is he o al numbe o
unique EC numbe s. Each ec o elemen
yij
is ei he 0 o 1, indica ing whe he p o ein
sequence
xi
is associa ed wi h he EC numbe j in Y. I is wo h no ing ha all le els we e
included in he esul ing ma ix as p oposed by [10], so i an enzyme is assigned he EC
numbe s 1.1.1.1 and 4.1.1.1, ones a e placed in
yi
a he posi ions co esponding o 1, 1.1,
1.1.1, 1.1.1.1, 4, 4.1, 4.1.1 and 4.1.1.1. He ein, we ake a global app oach o hie a chical
mul i-label classi ica ion, whe e a single classi ie is challenged o p edic he en i e hie -
a chy o labels and hei ela ionships [28].
UniP o KB da a ex ac ion andp ocessing
The SwissP o (manually anno a ed), T EMBL (au oma ically anno a ed) p o ein da a
and UniRe 90 clus e s we e downloaded om UniP o KB [19] in XML o ma in Feb-
ua y 2023. Then, an XML pa se was de eloped o ex ac he EC numbe s, p o ein
sequences, and iden i ie s o SwissP o and T EMBL. Only he UniRe 90 clus e ep-
esen a i es we e kep in he da ase s, as i s p oposed by [10]. By de ini ion, UniRe
clus e ep esen a i es a e chosen based on en y quali y, anno a ion sco e, o ganism
ele ance, and sequence leng h, enhancing biologically ele an in o ma ion e ie al.
This p ocedu e ensu es ha enzymes ha sha e mo e han 90% iden i y wi h hese ep-
esen a i es a e emo ed om he da ase , hus a oiding edundancy and p io i izing
en ies wi h be e anno a ion.
A e wa ds, o all EC numbe s wi h less han 100 sequences, we added da a om
T EMBL, i a ailable. The pipeline hen sys ema ically emo es he unde ep esen ed EC
numbe s, numbe s wi h less han 100 sequences assigned, and p o isional EC numbe s
wi h he le e “n” in he las le el. The inal da ase included 380,811 enzymes, ca ego-
ized unde EC numbe s as ollows: 7 o le el 1, 69 o le el 2, 230 o le el 3, and 2465
o le el 4.
Da a spli ing
Di iding da a in o aining, alida ion and es se s poses challenges wi h highly spa se
ma ices o labels like he ones in his wo k. I is c ucial o ensu e ha all labels a e well-
dis ibu ed ac oss he se s. To add ess his, we op ed o a spli o 60% o aining and
20% each o alida ion and es ing, i s ly using an i e a i e app oach o s a i ica ion
o mul i-labelled da a om he sciki -mul ilea n package . 0.2.0, desc ibed by [29].
Page 5 o 21
Capelae al. BMC Bioin o ma ics (2025) 26:68
Upon unning his ool, he s a i ica ion was unsuccess ul, as a ew EC classes had no
sequences assigned in he es o alida ion se s. We, hus, de eloped an algo i hm (Sup-
plemen a y Algo i hmS1) ha i e a i ely ans e s a calcula ed po ion o sequences
associa ed wi h EC numbe s om he aining se o he es o alida ion se . The algo-
i hm a ge s EC numbe s whe e he es -o e all da a p opo ion is lowe han 15%, a
alue de e mined as op imal a e some es s. I calcula es he numbe o samples o be
mo ed o compensa e o his gap. This ensu es ha unde ep esen ed EC numbe s a e
mo e e enly dis ibu ed ac oss he da ase s, as co obo a ed in Supplemen a y TableS1.
All he esul ing da ase in o ma ion can be e e ed o he Supplemen a y TablesS1
and S2. This p ocess o s a i ied spli ing may lead o ce ain EC numbe classes ha ing
ewe han 100 sequences each in he esul ing da ase s.
Complemen a ily, we de eloped a ain- es spli s a egy ha ocuses on iden i y
h esholds, while esampling he da a in o i e olds. We implemen ed six dis inc iden-
i y h esholds (90%, 80%, 70%, 60%, 50% and 40%) using CD-HIT [30] o ensu e ha no
sequences in he aining se sha e mo e han he speci ied iden i y h eshold wi h hose
in he es se . In o al, we go 30 ain- es combina ions (six iden i y h esholds imes
he i e olds). This app oach allowed us o e alua e how simila i y a ec s he pe o -
mance o each me hod.
Designing hee alua ion da ase s
In his wo k, we e alua ed he p edic i e capaci y o he models on i e di e en da a-
se s. The i s h ee a e p og essi ely e ined da ase s de i ed om he ini ial es se
based on UniP o KB da a and ob ained om he employed da a-spli ing s a egy. The
i s da ase included he en i e es se , while he second was e ined o p o eins wi h an
EC numbe co obo a ed by a leas one li e a u e e e ence, e med he e idence-le el
da ase . While his es se may con ain some au oma ically in e ed anno a ions, we
belie e ha i s eliance on manually cu a ed SwissP o en ies wi h li e a u e ci a ions
s ikes a good balance be ween anno a ion quali y and da ase size, especially since man-
ually anno a ing all en ies would be oo ime-consuming, mo e in o ma ion is included
in he Supplemen a y In o ma ion. The hi d da ase , aken om he e idence-le el
da ase , ocused on p omiscuous and mul i- unc ional enzymes. Two addi ional, ully
independen da ase s we e conside ed o e alua ion: one om [31] wi h 149 bac e ial
enzymes, known o i s expe imen ally alida ed ye challenging da a, and ano he com-
p ising 36 incomple ely anno a ed halogenases iden i ied om he UniP o KB da abase,
u he cu a ed by [16]. These wo da ase s ha e been used o pe o mance compa i-
sons [16, 17]. Supplemen a y TableS2 con ains he numbe o enzymes and EC numbe s
pe da ase .
Language models embeddings o enzyme sequences
In his s udy, we used h ee di e en p e- ained LLMs as ea u e ex ac o s o he
enzyme amino acid sequences. By ex ac ing he ou pu s o he las laye o hese models
and applying an elemen -wise mean o e i , we ob ained embeddings o he EC numbe
p edic ion. The models used as ea u e ex ac o s we e P o BERT [25], ESM1b [23], and
ESM2 [24], ained on p o ein sequences wi hou label supe ision. Fo de ails on he

Page 6 o 21
Capelae al. BMC Bioin o ma ics (2025) 26:68
p e-p ocessing o p o ein sequences, e e o Supplemen a y In o ma ion, Figu eS1 and
TableS3.
Le us deno e he ou pu o he ans o me ’s las laye as
M
, a ma ix o size
×n
,
whe e ep esen s he numbe o amino acids and n he dimensionali y o he ea u e
ep esen a ion o each oken. In p o ein LLMs, each oken ep esen s an indi idual
amino acid, se ing as he smalles da a uni p ocessed by he model. Mo eo e , a special
“CLS” oken is always added a he beginning o he sequences, and i does no ep esen
any amino acid. Fo his eason, we exclude i om ou calcula ion. The e o e, we con-
side he subma ix
M′
o
M
, excluding he i s ow, ep esen ed as:
whe e,
Mi,j
ep esen s he ea u e alue o he
j
- h dimension o he
i
- h oken. The
elemen -wise a e age o each dimension ac oss all okens is hen compu ed as ollows:
whe e
Ej
is he a e age o he
j
- h dimension ac oss all okens. The esul ing ec o o
hese a e ages,
E
, which is o dimension
n
, ep esen s he elemen -wise a e age o he
las laye ’s ou pu :
E
p o ides a single ep esen a ion ha cap u es he mean ea u es ac oss all okens in
he las laye o he ans o me . Se en dis inc embeddings we e calcula ed, employing
P o BERT, ESM1b, and i e a ia ions o ESM2. The esul ing dimensions o he used
LLM embeddings a e de ailed in Supplemen a y TableS4.
Model a chi ec u es andhype pa ame e s
Fi s , we ained baseline models wi h each o he embeddings o es ablish a ounda-
ion o e alua ing he e ec i eness and impac o mo e complex models. The baseline
linea models we e shallow ne wo ks wi h only inpu and ou pu laye s. The o me has
as many neu ons as he dimension o he ea u e ec o , and he la e as many neu ons
as labels (EC numbe s o p edic ) wi h a sigmoid ac i a ion unc ion. The loss unc ion
used was bina y c oss-en opy.
Then, we added hidden laye s wi h he Rec i ied Linea Uni (ReLU) ac i a ion unc-
ion, ollowed by ba ch no maliza ion, o he baseline models o assess he pe o mance
compa ed wi h he baselines. In his con ex , we pe o med an a chi ec u e sea ch o
ou ials o he se en embeddings. A chi ec u es 1 and 2 ea u ed a single hidden laye
wi h 640 and 2560 neu ons, espec i ely, while a chi ec u es 3 and 4 had wo hidden
laye s, wi h neu on coun s o 640 and 320 o he o me and 2560 and 5120 o he la -
e . We ained hese ne wo ks wi h he Adam op imize wi h a lea ning a e o 0.001,
as a e a ew p elimina y expe imen s on he aining se , we ound his alue o be
(1)
M
′=




m
1,0
m
1,1
··· m
1,n
m2,0 m2,1 ··· m2,n
.
.
..
.
.....
.
.
m ,0 m ,1 ··· m ,n




E
j=
1
( −1)

i=1
mi,j o j=0, 1, ...,
n
E=[E0,E1, ..., En]
Page 7 o 21
Capelae al. BMC Bioin o ma ics (2025) 26:68
he mos sui able alue o he lea ning a e. The aining p ocess was conduc ed o
30 epochs. Fo e minology, hese models will be e e ed o as “DNN”, ollowed by he
name o he language model ha p o ided he inpu embedding. Fo cla i y, models u i-
lizing LLM embeddings will hence o h be designa ed as DNN-LLM models, di e en i-
a ing hem om models ha do no employ such embeddings.
Among all he ained models, we chose hose wi h he highes pe o mance on he
alida ion se based on hei embeddings and a chi ec u e. Fo example, i a chi ec u e
4 deli e ed he bes pe o mance wi h ESM1b among all es ed a chi ec u es o DNN
ESM1b, we would selec i . As ESM2 has i e di e en embedding a ia ions, i he bes
a chi ec u e o DNN ESM2 3B ou pe o med DNN ESM2 650M, we p oceeded wi h he
bes a chi ec u e o DNN ESM2 3B. These op-pe o ming models we e hen e ained
using aining and alida ion da a. Complemen a ily, we c ea ed a model ensemble, a
o ing classi ie consis ing o ESM2 3B, ESM1b, and P o BERT bes models. The con-
sensus class p edic ion o his ensemble is based on ha d o ing, i.e. he inal p edic ion
is de e mined by a simple majo i y o e. Fo e minology, he ensemble o op-pe o m-
ing models is e e ed o as ’Models ensemble’.
Then, we eimplemen ed he D-SPACE and DeepEC app oaches, as hey encode
enzymes wi h one-ho encoding a he han LLM embeddings. Compa ing he pe o -
mance o hese models wi h ou s will p o ide aluable insigh s in o he ela i e e ec i e-
ness o LLM embeddings as encodings o enzyme sequences o EC numbe p edic ion.
D-SPACE [21] comp ises h ee epea ed CNN componen s and a ully connec ed ne -
wo k o gene a e a p edic ion ec o . We eimplemen ed he EC numbe s p edic ion
pa , which consis s o h ee CNNs wi h wo laye s o 1D con olu ions, ollowed by
ba ch no maliza ion in each, a max pooling laye a he end, and h ee ully connec ed
laye s and an ou pu laye wi h as many neu ons as he numbe o labels o p edic .
In DeepEC [3], he aining p ocess in ol es h ee CNNs: CNN-1, which assesses
whe he a p o ein is an enzyme; CNN-2, which p edic s EC numbe s up o he hi d
le el; and CNN-3, which ex ends he p edic ion o he ou h le el. The inal EC num-
be p edic ion is de i ed when he ou pu s o CNN-2 and CNN-3 a e consis en . How-
e e , i ei he CNN-2 o CNN-3 ail o p o ide a eliable p edic ion, DIAMOND [22]
is employed o make he EC p edic ion. In ou e ised implemen a ion o DeepEC, he
a chi ec u e exclusi ely inco po a es he CNN-3 model.
A h eshold o 0.5 was used o he ou pu laye o de e mine posi i e p edic ions
ac oss all models. By se ing he h eshold a 0.5, any ou pu p obabili y equal o o
g ea e han his alue is classi ied as posi i e, while p obabili ies below i a e classi ied
as nega i e. This app oach ensu es consis ency in e alua ion ac oss all models.
All models we e implemen ed and ained using he py o ch package . 2.1.2. Supple-
men a y Sec ion4 p esen s mo e de ails on he implemen a ion.
Pai ‑wise alignmen s
BLAST is designed o compa e a que y sequence agains a da abase o e e ence
sequences o ind egions o simila i y. BLAST uses he E- alue (Expec a ion alue) as a
measu e o he signi icance o a sequence alignmen . The E- alue is he expec ed num-
be o andom hi s wi h a simila o be e sco e ha one would encoun e in a da abase
sea ch o he gi en que y sequence pu ely by chance.
Page 8 o 21
Capelae al. BMC Bioin o ma ics (2025) 26:68
An expe imen was conduc ed using BLASTp o anno a e he es da ase s and compa e
he pe o mance wi h he DL models. We me ged he ain and alida ion se s o se e as
he e e ence da abase o BLASTp. The hi wi h he lowes E- alue was selec ed, and he
EC numbe was assigned di ec ly om he co esponding sequence. Al hough we used a
ully au oma ed p ocedu e, we acknowledge ha esul s om bo h BLAST and DL models
could be imp o ed wi h manual anno a ion. Howe e , his is beyond he scope o his wo k
and would be labo ious on a la ge scale. The BLAST E- alue cu o s we e se o
10−5
, as
sugges ed by [32] as a ypically used alue. NCBI BLAST e sion 2.12.0+ was used. A ull
lis o pa ame e s can be ound in Supplemen a y TableS5. We u he c ea ed a combined
ensemble o he op-pe o ming models and BLASTp, based on ha d o ing. The sys em
chooses he posi i e p edic ion when he e is a ie o p edic ions. In e ms o e minology,
he ensemble ha includes bo h he op-pe o ming models and BLASTp is called ’Models
+ BLASTp.’
Me ics
We employed se e al me ics o e alua e he pe o mance o he models. These me -
ics p o ide a comp ehensi e unde s anding o he model’s p edic i e capabili ies, espe-
cially when dealing wi h mul i-labelled da ase s. The key me ics a e he F1 sco e mac o
(mF1), F1 sco e weigh ed (wF1), Recall mac o (mRecall), Recall weigh ed (wRecall), P e-
cision mac o (mP ecision), P ecision weigh ed (wP ecision) and accu acy.
The mF1 is a me ic ha calcula es each label’s ha monic mean o p ecision and ecall
and hen a e ages hem o ob ain a single sco e o he en i e da ase . The same p inciple
is applied o mRecall and mP ecision, whe e he a e age o p ecisions and ecalls ac oss
labels is compu ed. Ano he way o calcula ing hese me ics is using he weigh ed a e -
age e sion, whe e he sco es o each label a e mul iplied by he ela i e equency o ha
label in he da ase . These me ics we e calcula ed using he sciki -lea n package e sion
1.2.0. Thei de ini ions and o mulas can be ound in he Supplemen a y TableS6.
Finally, we e alua ed he consis ency o he hie a chical p edic ions o ou global
mul i-label classi ie . As a hie a chical sys em, he EC numbe s mus be p edic ed as he
ollowing o mula ion: i a model p edic s a mo e speci ic label (like 1.1.1.1), i also mus
p edic all he co esponding highe -le el labels in he hie a chy (like 1.1.1, 1.1, and 1 in
his case). I he p edic ion aligns wi h he s a ed o mula ion, i is conside ed consis -
en ; o he wise, i is conside ed inconsis en . Acco dingly, we o mula ed a me ic called
Hie a chical Consis ency E o (HCE), de ined as ollows:
whe e T is he numbe o enzymes in he da ase , and
I(si)
is an inconsis ency indica-
o unc ion o each enzyme
si
, e u ning one i he e is a leas one inconsis ency in
he hie a chical p edic ions o enzyme
si
, and 0 o he wise. Le
P(si)
be he se o labels
p edic ed o enzyme
si
. Fo each label l in
P(si)
, le H(l) be he se o hie a chical highe -
le el labels o l. Then, he inconsis ency indica o unc ion
I(si)
o enzyme
si
is:
(2)
HCE
=
T
i=1I(si
)
T
(3)
I
(si)=

1, i ∃l∈P(si)such ha H(l)�⊆ P(si
)
0, o he wise
Page 9 o 21
Capelae al. BMC Bioin o ma ics (2025) 26:68
As EC numbe s can be ep esen ed as a hie a chical on ology in a di ec ed acyclic g aph
(DAG), whe e each node ep esen s an EC, we applied me ics speci ic o biological
hie a chical on ologies. In his de ini ion, le el 1 nodes b anch in o mo e speci ic sub-
classes a le el 2, u he speci ied by sub-subclasses a le el 3, and inally de ailed a a
deepe le el (le el 4), wi h each le el’s de ini ion dependen on i s p edecesso , es ablish-
ing a cascading speci ici y om gene al o p ecise.
Acco dingly, me ics o e alua ing how con iden each p edic ion is wi hin an on ol-
ogy we e conside ed. To in e p e hese p edic ions accu a ely, i is essen ial o conside
he emaining unce ain y, misin o ma ion, and S-min (seman ic dis ance) in ela ion o
a decision h eshold. P edic ions abo e o equal o he decision h eshold a e classi ied
as posi i e, while hose below a e nega i e. These me ics we e i s desc ibed in [33] and
implemen ed in Py hon in CAFA-e alua o [34]. We adap ed his implemen a ion o
he EC sys em.
Remaining unce ain y e lec s he in o ma ion missing om a p o ein’s ue anno a-
ion ha is no cap u ed by he p edic ion DAG. Fo mally, he emaining unce ain y
is he o al in o ma ion con en o nodes in he on ology ha exis s in he g ound u h
DAG bu is absen in he p edic ion DAG.
Misin o ma ion e lec s he o al in o ma ion con en o nodes along inco ec pa hs
in he p edic ion DAG, quan i ying how misleading he p edic ed anno a ion is.
Seman ic dis ance o S-min helps by combining hese wo measu es and is de ined as
he minimum dis ance om he o igin o he “ emaining unce ain y s. misin o ma-
ion” cu e. I does no jus look a how much in o ma ion is missing o inco ec indi-
idually bu balances bo h o gi e a single pe o mance sco e. When seman ic dis ance
is low, i indica es ha he p edic ion is bo h comple e (no missing key in o ma ion)
and accu a e (no misleading wi h inco ec in o ma ion). This helps ank algo i hms in
e ms o bo h hei accu acy and comple eness.
S a is ical me hods
We u ilized he Wilcoxon Signed-Rank (WSR) es o assess he s a is ical signi icance
o he di e ences in me ic alues be ween he me hods ac oss he es se s. In his sce-
na io, he samples comp ise he me ic alues gene a ed by wo me hods o each EC
numbe wi hin he mul i-label classi ica ion amewo k. The goal is o assess he signi i-
cance o any di e ences in he median me ic alues ac oss all EC numbe s. In applying
he WSR es o e alua e di e ences in me ic alues ac oss a ious iden i y h esholds
and EC numbe s, we also inco po a ed a False Disco e y Ra e (FDR) co ec ion o mul-
iple hypo hesis es ing. We applied he Benjamini-Hochbe g (BH) FDR co ec ion o
he p- alues. We deemed a co ec ed p- alue o less han 0.001 su icien o ejec ing
he null hypo hesis. Fo mo e in o ma ion on how we applied he s a is ical me hods,
please e e o Supplemen a y In o ma ion.
Resul s
DNN‑LLM pe o m gene ally onpa wi hBLASTp
Ou wo k low included a benchma k o models ained wi h ESM and P o BERT embed-
dings compa ed o BLASTp. Ini ially, linea baseline models we e es ablished, se ing
a e e ence o e alua ing mo e complex models. These baselines we e hen enhanced
Page 16 o 21
Capelae al. BMC Bioin o ma ics (2025) 26:68
indings e eal ha , among he di e en ially anno a ed classes i.e.
F1i = 0
, he p o-
po ion o classes be e anno a ed by DL models is lowe o hese unde ep esen ed
classes (24.9%) han he en i e se o classes (38.7%). This sugges s ha while BLASTp
gene ally anno a es mo e classes e ec i ely, his is mo e p onounced o classes ep e-
sen ed less well in he aining se .
DNN‑LLMs ou pe o m BLASTp ondi icul asks
To challenge he EC p edic ion me hods wi h mo e di icul asks, we applied hem o
wo published da ase s: newly anno a ed bac e ial enzymes [31] and halogenases [16].
Since hese da ase s a e independen and con ain enzymes no ound in he aining se s
o any o he ollowing s a e-o - he-a app oaches, we compa ed ou me hods wi h wo
leading app oaches: P o eIn e , a CNN-based me hod, and CLEAN, a con as i e lea n-
ing model buil using ESM1b embeddings.
Fig. 4 Compa ison o DNN-LLM models’ and BLAST’s F1 sco es. a,c,e Dis ibu ion o DNN-LLM models’ and
BLASTp’s F1 sco e di e ences pe numbe o sequences in he aining se (
log10
), whe e each do ep esen s
an EC class. b,d, Numbe o classes wi h an
F1i
abo e a gi en h eshold

Page 17 o 21
Capelae al. BMC Bioin o ma ics (2025) 26:68
The o me comp ises enzymes whose unc ions we e ecen ly de e mined h ough a
high- h oughpu expe imen al gene ic s udy [31]. The au ho s p o ided new EC num-
be s o enzymes whe e p io anno a ions in he SEED o Kyo o Encyclopedia o Genes
and Genomes (KEGG) da abases we e ei he inco ec o inconsis en . Gi en his con-
ex , his da ase is likely en iched wi h p o eins whose unc ions pose a signi ican
challenge o compu a ional assessmen , making i in aluable o es ing and e ining
compu a ional anno a ion me hods. Fo e alua ion pu poses, wRecall, wP ecision, and
wF1 sco es a le el 4 we e exclusi ely employed o add ess he signi ican class imbal-
ance and o align wi h he me ics used by [16], acili a ing compa isons wi h o he ools.
BLASTp pe o ms wo se han all me hods excep DNN P o BERT (Table3). Mo eo e ,
P o eIn e pe o ms wo se han all me hods, which ein o ces he idea ha LLM-based
models ou pe o m CNN-based models. The cus om models de eloped o his wo k,
DNN ESM2 3B and DNN ESM1b, exhibi s ong pe o mance. The Models ensemble
shows balanced pe o mance ac oss all me ics bu does no ou pe o m he indi idual
DNN ESM2 3B model. The Models + BLASTp ensemble ou pe o ms o he me hods
in wP ecision, including CLEAN. On he o he hand, CLEAN deli e s be e wF1 and
wRecall han any o he app oach. This shows ha models ained wi h LLM embeddings
a e p omising o ( e-)anno a e enzymes ha a e di icul o p edic o ha e inconsis en
anno a ions ac oss biochemical da abases.
The second challenging da ase comp ises 36 halogenases, a g oup o incomple ely
anno a ed enzymes iden i ied om he UniP o KB da abase. The da ase p esen s a sig-
ni ican challenge since he halogenase amily is ela i ely unde s udied, and he exis ing
anno a ions in da abases a e o en incomple e o con lic ing, complica ing he ask o
accu a e classi ica ion. Despi e hese challenges, expe cu a ion was conduc ed by [16],
leading o he con iden assignmen o speci ic EC numbe s o all 36 halogenases in he
da ase . To assess me hod pe o mance, accu acy ac oss all le els was compu ed, gi en
he lesse deg ee o imbalance, ensu ing alignmen wi h he e alua ion c i e ia ou lined
in [16] and enabling compa a i e analysis wi h al e na i e ools. Fu he mo e, gi en he
small sample size in his da ase , binomial es s we e conduc ed o compa e he a es
o co ec ly and inco ec ly anno a ed p o eins o each EC numbe a each le el. The
expec ed p opo ions o co ec ly anno a ed p o eins we e es ima ed based on he o e -
all numbe o co ec ly anno a ed p o eins wi hin each le el.
Table 3 Model wRecall, wP ecision and wF1 sco es o he [31] da ase . The highes -pe o ming
esul o each me ic is indica ed in bold
Me hod wF1 wRecall wP ecision
CLEAN 0.495 0.584 0.467
P o eIn e 0.166 0.138 0.243
DNN ESM2 3B 0.433 0.408 0.535
DNN ESM1b 0.421 0.382 0.587
DNN P o BERT 0.357 0.316 0.526
Models Ensemble 0.392 0.362 0.525
BLASTp 0.372 0.329 0.499
Models + BLASTp 0.444 0.408 0.600
Page 18 o 21
Capelae al. BMC Bioin o ma ics (2025) 26:68
A halogenase e e s o an enzyme ha inco po a es halogen a oms (such as chlo-
ine, luo ine, o b omine) in o o ganic compounds, o en o ca alyzing halogena-
ion eac ions ha add bioac i e p ope ies o compounds [36]. In he con ex o EC
e minology, hese enzymes all unde speci ic EC classi ica ions based on hei eac-
ion ypes, such as halope oxidases (classi ied as oxido educ ases, EC 1.11.1.10) o
la in-dependen (1.14.19.x, depending on he subs a e),
α
-ke oglu a a e-dependen
(1.14.20.x and 1.14.11.x), and S-adenosyl-L-me hionine (SAM)-dependen haloge-
nases (which can be classi ied as ans e ases, EC 2.5.1.x, depending on hei exac
subs a e, o hyd olases, EC 3.13.1.8) [16].
BLASTp exhibi ed lowe accu acy han mos me hods, as p esen ed in Table4. As
BLASTp was no able o ind homologous sequences o ou que y sequences om
his da ase , his migh ha e a ec ed he esul s compa ed o DL models, since DNN-
LLMs could gene a e co ec p edic ions o hese sequences. A he same ime, DNN
ESM2 3B, DNN ESM1b, and DNN P o BERT a ied in hei highes accu acies ac oss
le els 1 and 2, bu all showed dec eased pe o mance a le els 3 and 4. The Models
ensemble ma ched he op pe o mances o DNN ESM2 3B and DNN ESM1b a he
i s wo le els bu also d opped in accu acy a he lowe le els. Models + BLASTp
achie ed he highes o e all ac oss all le els wi hou conside ing CLEAN’s pe o -
mance. CLEAN ob ained he bes accu acies ac oss he boa d, and P o eIn e ell
sho a all le els compa ed o all he o he me hods. Binomial es s e ealed ha , o
DNN ESM2 3B, he con idence in e al o he p edic ions o halope oxidases (EC
1.11.1.10) was sligh ly abo e he expec ed alue o baseline pe o mance a le el 4
(see Supplemen a y Ma e ial). Simila ly, DNN ESM1b and DNN P o BERT demon-
s a ed con idence in e als exceeding he baseline expec a ion o he p edic ion o
α
-ke oglu a a e-dependen halogenases (speci ically hose wi hin EC 1.14.11.x). These
esul s sugges ha hese models exhibi supe io anno a ion pe o mance o hese
enzyme classes compa ed o o he enzymes a he same le el.
In his case, he challenges ac oss di e en EC le els a ise om he need o dis-
inguish b oad unc ional ca ego ies (le el 1), eac ion- ela ed componen s like sol-
en s, co ac o s, and elec on accep o s (le els 2 and 3, especially in oxido educ ases),
and speci ic subs a es (le el 4). CLEAN e ec i ely iden i ies sub le pa e ns in he
Table 4 Model accu acies o he halogenases da ase . The highes -pe o ming esul o each le el
is indica ed in bold
a
This assessmen was pe o med based on he g ound u h de i ed only om he manual cu a ion in [16]
Me hod Accu acy
le el 1 le el 2 le el 3 le el 4
CLEAN 1.000 0.972 0.944 0.944
P o eIn e 0.611 0.388 0.083 0.694
DNN ESM2 3B 0.917 0.917 0.500 0.805
DNN ESM1b 0.917 0.917 0.500 0.833
DNN P o BERT 0.805 0.917 0.528 0.750
Models Ensemble 0.917 0.917 0.500 0.833
BLASTp 0.750 0.639 0.194 0.583
Models + BLASTp 0.917 0.917 0.555 0.833
Page 19 o 21
Capelae al. BMC Bioin o ma ics (2025) 26:68
ESM1b embeddings ha indica e subs a e-speci ic in o ma ion, which helps explain
i s s ong pe o mance a he de ailed, subs a e-speci ic le el 4 EC classi ica ion.
Discussion
This s udy compa ed DL models ained wi h embeddings de i ed om LLMs and he
gold s anda d me hod BLASTp. This s udy deli e s a ully ep oducible end- o-end pipe-
line o benchma king models and p o ein ep esen a ions o EC numbe p edic ion.
We show ha using LLMs as s aigh o wa d ea u e ex ac o s leads o be e esul s
han models elying on he one-ho encoding o sequences (e.g. D-SPACE, DeepEC and
P o eIn e ). Gene ally, DNN-LLM models we e ma ginally ou pe o med by BLASTp
o la ge da ase s wi h homologous sequences a ailable om da abases. While BLASTp
demons a ed a sligh pe o mance imp o emen o e DL models, i is impo an o
no e ha i has an inhe en ad an age in his con ex . Since i is one o he mos widely
used anno a ion me hods, i is used o anno a e enzymes in UniP o KB, which we use
o es ing he me hods, hus a ou ing BLASTp. Conside ing ha , DL models ained
wi h LLMs achie e imp essi e esul s, wi h a minimal di e ence o 0.004 in he mF1
sco e compa ed o he one om BLASTp. Despi e his sligh pe o mance gap, in sce-
na ios whe e he aining se lacked homologous sequences, LLM models ou pe o med
BLASTp, indica ing hei po en ial u ili y in mo e challenging anno a ion asks. Pa icu-
la ly, when BLASTp ails o ind an enzyme wi h mo e han 25% iden i y in he aining
se , i s pe o mance declines, and DL models ou pe o m i . Hence, we claim ha DL
models should be employed below his ma k, ins ead o BLASTp. Addi ionally, com-
bining BLASTp wi h DL models can enhance esul s ac oss nea ly all iden i y in e als.
We also con i m ha DL models a e sligh ly mo e p ecise and ha e lowe ecall han
BLASTp, as poin ed ou by [13].
As u u e wo k on in eg a ing BLASTp and DL models, a mo e nuanced app oach
could be adop ed a he han simply a ou ing DL models o e BLASTp based solely
on sequence iden i y. Speci ically, an al e na i e me hod could be implemen ed o
adjus p edic ions acco ding o he numbe , que y co e age, sequence iden i y and
o he me ics o iden i ied BLASTp hi s. This app oach could in ol e using a lin-
ea eg ession model o a simila ML echnique o op imize he in eg a ion o p e-
dic ions om bo h BLASTp and DL models. By le e aging he s eng hs o each
me hod unde di e en condi ions, his s a egy has he po en ial o enhance o e -
all p edic i e pe o mance.
I is impo an o emphasize ha his amewo k p edic s EC numbe s o a gi en
enzyme bu does no de e mine whe he he inpu molecule is an enzyme. Fu u e
e o s should ocus on benchma king p o ein LLM models o his speci ic ask.
C ucially, we show ha p o ein LLM models should s ill be imp o ed o u he
na ow o e en su pass he pe o mance gap wi h BLASTp and po en ially become
he gold s anda d in enzyme anno a ion. O e all, we s eng hen he claim ha
BLASTp and LLM models a e complemen a y ools ha combined can be ad an a-
geous when employed in mains eam p o ein anno a ion ou ines and ha d anno a-
ion asks.
Page 20 o 21
Capelae al. BMC Bioin o ma ics (2025) 26:68
Supplemen a y In o ma ion
The online e sion con ains supplemen a y ma e ial a ailable a h ps:// doi. o g/ 10. 1186/ s12859- 025- 06081-9.
Supplemen a y ile 1.
Supplemen a y ile 2.
Acknowledgemen s
EMBL IT Suppo is acknowledged o he p o ision o compu e and da a s o age se e s.
Au ho Con ibu ions
Capela J. de eloped he me hodology and he analysis. Capela J., Zimme mann-Kogadee a M., an Dijk A., de Ridde D.,
Dias O. and Rocha M. w o e he manusc ip and concep ualized he s udy. Zimme mann-Kogadee a M., an Dijk A., de
Ridde D., Dias O. and Rocha M. supe ised he s udy. All au ho s edi ed and app o ed he inal manusc ip .
Funding
FCT suppo ed his s udy unde he s a egic unding o he UIDB/04469/2020 uni and by LABBELS (LA/P/0029/2020).
We also hank FCT o he PhD ellowship o J. Capela (DFA/BD/08789/2021) and unding O. Dias (DOI 10.54499/CEEC-
IND/03425/2018/CP1581/CT0020). J. Capela also acknowledge he Eu opean Molecula Biology Labo a o y (EMBL)
Co po a e Pa ne ship P og amme o unding a sho - e m isi o ellowship. M. Zimme mann-Kogadee a is suppo ed
by he EMBL.
A ailibili y o da a and ma e ials
Code o he model implemen a ion and analysis is a ailable a : h ps:// gi hub. com/ jcape ls/ ec_ numbe s_ p edi c ion.
P- alues o all s a is ical es s can be ound in he Supplemen a y Ma e ial.
Decla a ions
E hics app o al and consen o pa icipa e
No applicable.
Consen o publica ion
No applicable.
Compe ing in e es s
The au ho s decla e ha hey ha e no Con lic o in e es .
Recei ed: 2 Sep embe 2024 Accep ed: 11 Feb ua y 2025
Re e ences
1. Kelle MA, Pied a i a G, Ralse M. The widesp ead ole o non-enzyma ic eac ions in cellula me abolism. Cu Opin
Bio echnol. 2015;8(34):153–61.
2. Le eson-Gowe RB, Maye C, Roel es G. The impo ance o ca aly ic p omiscui y o enzyme design and e olu ion.
Na Re Chem. 2019;3(12):687–705.
3. Ryu JY, Kim HU, Lee SY. Deep lea ning enables high-quali y and high- h oughpu p edic ion o enzyme commission
numbe s. PNAS. 2019;116(28):13996–4001.
4. Ko e a M, Okuno Y, Ha o i M, e al. Compu a ional assignmen o he EC numbe s o genomic-scale analysis o
enzyma ic eac ions. JACS. 2004;126(50):16487–98.
5. Baldazzi D, Sa oja do C, Ma elli PL, e al. BENZ WS: he bologna ENZyme web se e o ou -le el EC numbe anno-
a ion. Nucleic Acids Res. 2021;49(W1):W60–6.
6. Capela J, Lagoa D, Rod igues R, e al. An imp o ed amewo k o he econs uc ion o high-quali y genome-scale
me abolic models. Nucleic Acids Res. 2022;50(11):6052–66.
7. Ques e S, Schombu g D. EnzymeDe ec o : an in eg a ed enzyme unc ion p edic ion ool and da abase. BMC
Bioin . 2011;12(1):1–13.
8. Li Y, Wang S, Uma o R, Xie B, e al. DEEP e: sequence-based enzyme EC numbe p edic ion by deep lea ning. Bioin-
o ma ics. 2018;34(5):760–9.
9. Al schul SF, Gish W, Mille W, e al. Basic local alignmen sea ch ool. J Mol Biol. 1990;215(3):403–10.
10. Fe a i LD, Ai ken S, an Heme J, Go yanin I. EnzML: mul i-label p edic ion o enzyme classes using In e P o signa-
u es. BMC Bioin . 2012;12(13):61.
11. Kuma N, Skolnick J. EFICAz2.5: applica ion o a high-p ecision enzyme unc ion p edic o o 396 p o eomes. Bioin-
o ma ics. 2012;28:2687–8.
12. Nagao C, Nagano N, Mizuguchi K. P edic ion o de ailed enzyme unc ions and iden i ica ion o speci ici y de e -
mining esidues by andom o es s. PLoS One. 2014;9(1): e84623.
13. Sande son T, Bileschi ML, Belange D, e al. P o eIn e deep neu al ne wo ks o p o ein unc ional in e ence. eLi e.
2023;12: e80942.
14. Shi Z, Deng R, Yuan Q, e al. Enzyme commission numbe p edic ion and benchma king wi h hie a chical dual-co e
mul i ask lea ning amewo k. Resea ch. 2023;1:6.
Page 21 o 21
Capelae al. BMC Bioin o ma ics (2025) 26:68
15. Bu on N, Cos e F, Cun YL. P edic ing enzyma ic unc ion o p o ein sequences wi h a en ion. Bioin o ma ics.
2023;10:39.
16. Yu T, Cui H, Li JC, e al. Enzyme unc ion p edic ion using con as i e lea ning. Science. 2023;379:1358–63.
17. Kim GB, Kim JY, Lee JA, e al. Func ional anno a ion o enzyme-encoding genes using deep lea ning wi h ans-
o me laye s. Na Commun. 2023;11(14):7370.
18. Dalki an A, Ri aioglu A, Ma in MJ, e al. ECP ed: a ool o he p edic ion o he enzyma ic unc ions o p o ein
sequences based on he EC nomencla u e. BMC Bioin . 2018;19:1–13.
19. Ba eman A, Ma in MJ, O cha d S, e al. UniP o : he uni e sal p o ein knowledgebase in 2023. Nucleic Acids Res.
2023;1(51):D523–31.
20. Mis y J, Chugu ansky S, Williams L, e al. P am: he p o ein amilies da abase in 2021. Nucleic Acids Res.
2021;49(D1):D412–9.
21. Schwa z AS, Hannum GJ, Dwiel ZR, e al. Deep seman ic p o ein ep esen a ion o anno a ion, disco e y, and
enginee ing. bioRxi . 2018;.
22. Buch ink B, Xie C, Huson DH. Fas and sensi i e p o ein alignmen using DIAMOND. Na Me hods. 2015;12(1):59–60.
23. Ri es A, Meie J, Se cu T, e al. Biological s uc u e and unc ion eme ge om scaling unsupe ised lea ning o 250
million p o ein sequences. P oc Na l Acad Sci. 2021;4:118.
24. Lin Z, Akin H, Rao R, e al. E olu iona y-scale p edic ion o a omic-le el p o ein s uc u e wi h a language model.
Science. 2023;379(6637):1123–30.
25. Elnagga A, Heinzinge M, Dallago C, e al. P o T ans: owa d unde s anding he language o li e h ough sel -supe -
ised lea ning. IEEE T ans Pa e n Anal Mach In ell. 2022;44(10):7112–27.
26. Fenoy E, Ede a AA, S egmaye G. T ans e lea ning in p o eins: e alua ing no el p o ein lea ned ep esen a ions o
bioin o ma ics asks. B ie Bioin . 2022;23(4):bbac232.
27. S einegge M, Mi di a M, Söding J. P o ein-le el assembly inc eases p o ein sequence eco e y om me agenomic
samples many old. Na u e me hods. 2019;16(7):603–6.
28. Weh mann J, Ce i R, Ba os R. Hie a chical mul i-label classi ica ion ne wo ks. In: Dy J, K ause A, edi o s. P oc. 35 h
ICML. ol. 80 o P oc. Mach. Lea n. Res.; 2018. p. 5075–5084.
29. Sechidis K, Tsoumakas G, Vlaha as I. On he s a i ica ion o mul i-label da a. In: Gunopulos D, Ho mann T, Male ba D,
Vazi giannis M, edi o s. Machine lea ning and knowledge disco e y in da abases. Be lin Heidelbe g: Sp inge ; 2011.
p. 145–58.
30. Fu L, Niu B, Zhu Z, Wu S, Li W. CD-HIT: accele a ed o clus e ing he nex -gene a ion sequencing da a. Bioin o ma -
ics. 2012;28(23):3150–2.
31. P ice MN, We mo e KM, Wa e s RJ, e al. Mu an pheno ypes o housands o bac e ial genes o unknown unc ion.
Na u e. 2018;557(7706):503–9.
32. Choudhu i S. In: sequence alignmen and simila i y sea ching in genomic da abases. Else ie ; 2014. p. 133–155.
33. Cla k WT, Radi ojac P. In o ma ion- heo e ic e alua ion o p edic ed on ological anno a ions. Bioin o ma ics.
2013;29(13):i53–61.
34. Pio esan D, Zago D, Joshi P, De Paolis Kaluza MC, Mehdiabadi M, Ramola R, e al. CAFA-e alua o : a py hon ool o
benchma king on ological classi ica ion me hods. Bioin Ad . 2024;4(1): bae043.
35. Bileschi M, Belange D, B yan D, e al. Using deep lea ning o anno a e he p o ein uni e se. Na Bio ech.
2022;40(6):932–7.
36. C owe C, Molyneux S, Sha ma SV, Zhang Y, Gko si DS, Conna is H, e al. Halogenases: a pale e o eme ging oppo u-
ni ies o syn he ic biology-syn he ic chemis y and C-H unc ionalisa ion. Chem Soc Re . 2021;50(17):9443–81.
Publishe ’s No e
Sp inge Na u e emains neu al wi h ega d o ju isdic ional claims in published maps and ins i u ional a ilia ions.