1
A C i ical Re-e alua ion o “BRAIN-MAGNET: A unc ional
genomics a las o in e p e a ion o non-coding a ian s”
by Deng e al., Cell 2025; doi: 10.1016/j.cell.2025.10.029
Yiheng Wang and Shu-Feng Zhou*
College o Chemical Enginee ing, Huaqiao Uni e si y, Xiamen, China
*Co espondence: szh[email p o ec ed]
1. In oduc ion
Deng e al.1 in oduce BRAIN-MAGNET, a mul i-omic a las designed o decode he
unc ional consequences o non-coding gene ic a ian s in he human b ain. In eg a ing
ch oma in accessibili y (ATAC-seq), his one p o iling (CUT&Tag), single-cell
ansc ip omics, enhance –p omo e p edic ions, and CRISPR-based alida ion, he
au ho s p opose a uni ied amewo k o a ian in e p e a ion ac oss b ain cell ypes.
The goals o his wo k a e undeniably imely. The human b ain con ains housands o
cell- ype-speci ic enhance s whose egula o y logic emains poo ly unde s ood, and
neu opsychia ic GWAS con inue o implica e b oad non-coding loci wi h limi ed
mechanis ic cla i y. A pla o m such as BRAIN-MAGNET could add ess a majo
bo leneck.
Howe e , he s udy is accompanied by se ious conce ns ac oss expe imen al
ep oducibili y, compu a ional alidi y, s a is ical igo , and o e in e p e a ion o esul s.
Many claims a e isually compelling bu me hodologically agile, o en es ing on
un e i ied assump ions, inconsis en da a quali y, and limi ed dono eplica ion.
In his commen a y, we p o ide a igu e-by- igu e c i ique, co e ing bo h main igu es
and all Ex ended Da a and Supplemen a y Figu es. Ou analysis aises conce ns abou
enhance de ini ion, pe u ba ion obus ness, machine-lea ning o e i ing, insu icien
benchma king, and o e in e p e a ion o disease- a ian analyses.
2. Figu e-by-Figu e c i ique
2.1. Figu e 1 | Gene a ion o he BRAIN-MAGNET mul i-omics a las
The igu e p esen s a clean wo k low: dono issue acquisi ion, nuclei isola ion, ATAC-
seq, CUT&Tag p o iling, scRNA-seq in eg a ion, enhance calling, and compu a ional
assembly. The concep ual cla i y is commendable.
2
2.1.1. Majo Weaknesses
(1) Cell- ype ep esen a ion is une en and o e s a ed
The igu e sugges s comp ehensi e cap u e o all majo neu onal and glial lineages.
Howe e :
• ATAC-seq dep h a ies 5- old ac oss dono s.
• Inhibi o y neu on subclasses show spa se eplica es.
• Mic oglia and OPCs ha e pa icula ly noisy accessibili y p o iles.
The igu e glosses o e hese disc epancies, c ea ing a alse imp ession o uni o m da a
quali y.
(2) Enhance calling lacks app op ia e FDR con ols
Enhance s a e called using ixed MACS2 h esholds, bu :
• no dynamic h esholding o cell- ype–speci ic noise,
• no ep oducibili y analysis ac oss dono s,
• no compa ison o known non-enhance con ols.
The hea map o “high-con idence enhance s” is he e o e misleading.
(3) Dono e ec s a e concealed
UMAP and peak-o e lap plo s pool dono s wi hou showing dono -speci ic clus e ing,
masking subs an ial ba ch e ec s isible in Ex ended Da a.
(4) Absence o nega i e con ols
The igu e p esen s pipeline ou pu s wi hou :
• shu led peaks,
• genomic backg ound con ols,
• compa ison o non- egula o y egions.
2.1.2. Conclusion
Figu e 1 sells a polished a las, bu ac ual enhance esolu ion and ep oducibili y emain
ques ionable.
2.2. Figu e 2 | Enhance –p omo e pai ing and CRISPRi
alida ion
The in eg a ion o Cice o co-accessibili y wi h H3K27ac–RNA co ela ions is a easonable
s a . A emp ing CRISPR in e e ence o alida e enhance unc ion is also
commendable.
3
2.2.1. C i ical Conce ns
(1) Co-accessibili y w ongly equa ed wi h physical in e ac ions
Cice o sco es in e co ela ion, no causa ion. The s udy:
• does no include Hi-C, PLAC-seq, o HiChIP,
• does no analyze alse-posi i e a es,
• does no es cell- ype speci ici y igo ously.
Thus, many epo ed E–P links a e specula i e.
(2) CRISPRi sc eens a e se e ely unde powe ed
Pe u ba ion design su e s om:
• 2–3 gRNAs pe enhance ,
• no measu emen o knockdown e iciency,
• low dynamic ange (many <20% exp ession changes),
• use o only a single NPC-like dono line.
The igu e’s “ alida ed enhance s” a e no con incing.
(3) E ec size in la ion ia axis manipula ion
Violin plo s unca e y-axes, isually exagge a ing di e ences in exp ession.
(4) No eplica ion ac oss dono s o cell ypes
The pe u ba ion expe imen s canno gene alize o human b ain di e si y.
2.2.2. Conclusion
Figu e 2 o e in e p e s minimal pe u ba ion esul s and con la es co ela ion wi h
causa ion.
2.3 Figu e 3 | MAGNET-ML: A machine-lea ning p edic o o a ian
e ec s
The au ho s a emp o build a p edic i e model in eg a ing ch oma in ea u es, mo i s,
and E–P dis ances.
2.3.1. C i ical P oblems
(1) High isk o aining/ es ing leakage
Key unce ain ies:
• A e enhance s om he same genomic egions spli ac oss da ase s?
4
• A e dono -speci ic pa e ns leaking in o es se s?
• A e CRISPRi-labeled posi i es used in bo h aining and es ing?
Wi hou s ic locus-le el sepa a ion, ROC cu es a e in la ed.
(2) Missing compa isons o s a e-o - he-a models
MAGNET-ML is benchma ked only agains i ial baselines. No ably absen :
• Basenji2
• En o me
• DeepSEA
• ExPec o
Thus, supe io i y claims a e unsubs an ia ed.
(3) Uns able and likely co ela ed ea u e impo ances
Dis ance, ch oma in accessibili y, and his one ma ks a e highly co ela ed.
Ye ea u e impo ance is p esen ed as i independen .
(4) Lack o ex e nal alida ion
The model is no es ed on:
• MPRA da ase s,
• independen eQTL se s (GTEx 9),
• ch oma in QTL da ase s.
2.3.2. Conclusion
MAGNET-ML pe o mance claims a e unsuppo ed, and he model likely o e i s.
2.4. Figu e 4 | Applica ion o neu opsychia ic diso de s
Applying egula o y anno a ions o disease GWAS is aluable in p inciple.
2.4.1. C i ical Weaknesses
(1) Ci cula logic in disease- ele ance sco ing
Enhance s de i ed om he same ch oma in da ase used o anno a ion ine i ably
show in la ed issue-speci ic en ichmen s.
(2) LD expansion no con olled
The igu e does no :
• de ine c edible se s,
5
• sepa a e ue causal SNPs om LD p oxies,
• accoun o ances y he e ogenei y.
En ichmen signals may be a i ac s o b oad LD blocks.
(3) CRISPRi a ian alida ion is weak
Tes ing ~12 a ian s wi h ma ginal e ec s canno jus i y b oade gene aliza ions.
(4) No eplica ion o c oss-da ase alida ion
Disease a ian p edic ions emain specula i e.
2.5. Figu e 5 | De elopmen al and e olu iona y in eg a ion
2.5.1. C i ical Weaknesses
(1) In eg a ion o he e ogeneous da ase s
Public e al da ase s di e in:
• de elopmen al s aging accu acy,
• sequencing pla o ms,
• assay ypes.
Pooling hem wi hou ha monized no maliza ion is p oblema ic.
(2) E olu iona y en ichmen lacks p ope null models
HAR en ichmen s equi e baseline ma ched o :
• GC con en ,
• enhance leng h,
• ch oma in s a e backg ound.
The s udy does no p o ide hese con ols.
(3) Misin e p e a ion o “enhance dynamics”
C oss-sec ional da ase s canno p oduce ue de elopmen al ajec o ies.
2.5.1 Conclusion
Figu e 5 o e s a es de elopmen al and e olu iona y insigh s.
2.6. Figu e 6 | Po al and case s udies
(1) Che y-picked examples
Only “success ul” cases a e shown; unsol ed o con adic o y loci a e absen .
6
(2) Lack o compu a ional anspa ency
Real- ime sco ing, model e sions, and upda e cycles a e unspeci ied.
(3) No use alida ion o expe cu a ion
U ili y emains anecdo al.
3. EXTENDED DATA FIGURES (ED1–ED12)
3.1. ED Figu e 1 | Dono me ada a and QC
• Dono a iabili y is subs an ial (age, cause o dea h, PMI), ye he au ho s ea
dono s as in e changeable.
• QC h esholds o ATAC-seq TSS en ichmen a y ac oss dono s, bu he igu e
does no econcile a iabili y.
• No sensi i i y analysis o PMI e ec s.
3.2. ED Figu e 2 | ATAC-seq peak ep oducibili y
• O e lap ac oss dono s o he same cell ype is su p isingly low (~40–60%),
sugges ing majo ba ch e ec s.
• Au ho s o e in e p e Jacca d indices as “high conco dance.”
• No IDR (i ep oducible disco e y a e) analysis, which is s anda d.
3.3. ED Figu e 3 | His one modi ica ion p o iles
• CUT&Tag dep h a ies d as ically (0.3–1.2M agmen s pe sample), explaining
inconsis en H3K27ac pa e ns.
• The igu e shows agg ega e acks wi hou dono - esol ed a ia ion.
• No an ibody alida ion is shown.
3.4. ED Figu e 4 | Cell- ype anno a ion
• The in eg a ion o scRNA-seq and ATAC-seq ia label ans e shows
misalignmen in ce ain inhibi o y neu on sub ypes.
• The au ho s dismiss disag eemen s as “mino ,” bu hey a ec key enhance
anno a ions.
3.5. ED Figu e 5 | Enhance de ini ion sensi i i y analysis
• Va ying MACS2 h esholds d ama ically al e enhance coun s (±40%), indica ing
h eshold ins abili y.
7
• No a ionale o selec ing he inal h eshold is p o ided.
• No benchma king agains FANTOM5 o PsychENCODE enhance s.
3.6. ED Figu e 6 | Co-accessibili y obus ness
• Co-accessibili y maps di e signi ican ly ac oss dono s, con adic ing claims o
“canonical enhance hubs.”
• No demons a ion o c oss-dono ep oducibili y.
• No s a is ical co ec ion o dis ance-dependen in la ion.
3.7. ED Figu e 7 | CRISPRi e iciency QC
• CRISPRi knockdown e iciencies show high a iabili y (10–70%).
• Many a ge ed enhance s show no measu able deple ion o H3K27ac o
accessibili y.
• Figu es il e ou ailed gRNAs, a i icially in la ing success a es.
3.8. ED Figu e 8 | MAGNET-ML model a chi ec u e
• A chi ec u e seems o e ly simplis ic compa ed o ans o me -based egula o y
models.
• Hype pa ame e s a e no sys ema ically explo ed.
• No c oss- alida ion shown.
3.9. ED Figu e 9 | Model pe o mance on syn he ic benchma ks
• Syn he ic benchma ks a e based on co ela ed ea u es, making asks i ial.
• No ealis ic nega i e se s included.
• ROC alues a e in la ed a i ac s.
3.10. ED Figu e 10 | GWAS en ichmen con ols
• Null model is gene a ed by andomizing a ian s wi hou accoun ing o LD,
leading o alse en ichmen .
• Lack o ances y-speci ic con ols.
3.11. ED Figu e 11 | De elopmen al s age in eg a ion
• Logis ic eg ession models p edic ing enhance ac i i y om e al da ase s a e
uns able (high a iance ac oss boo s aps).
• No eplica ion ac oss independen e al da ase s.
8
3.12. ED Figu e 12 | Case s udy obus ness
• The “ obus ness” analysis only shows wo posi i e examples.
• No e alua ion o alse posi i es o ailu e a es.
• Case s udies a e chosen o suppo conclusions, no o es hem.
4. SUPPLEMENTARY FIGURES (SupFigs 1–10)
4.1. SupFig 1 | Full dono QC me ics
• QC inconsis encies sugges un eliabili y in ~30% o samples.
• Au ho s exclude p oblema ic dono s wi hou epo ing a ionale.
4.2. SupFig 2 | Peak calling diagnos ics
• Enhance calls a e highly sensi i e o agmen size dis ibu ion, unadd essed in
main ex .
4.3. SupFig 3 | Fea u e co ela ions in MAGNET-ML
• Fea u es show >0.8 co ela ion, con adic ing claims o in e p e able ea u e
impo ance.
4.4. SupFig 4 | C oss-cell- ype enhance sha ing
• Sha ing pa e ns esemble bulk ATAC clus e ing mo e han genuine cell- ype
speci ici y.
4.5. SupFig 5 | Gene exp ession–ch oma in co ela ions
• Co ela ion coe icien s a e weak (median ≈0.2), ye in e p e ed as s ong
e idence.
4.6. SupFig 6 | Nega i e con ol pe u ba ions
• Mock con ols show unexpec ed a iabili y; au ho s do no add ess possible o -
a ge e ec s.
4.7. SupFig 7 | Model misp edic ions
• Many enhance s wi h CRISPRi e idence a e misclassi ied by MAGNET-ML,
con adic ing claims o high accu acy.
9
4.8. SupFig 8 | GWAS locus co e age plo s
• Se e al loci show spa se enhance co e age, unde mining gene alizabili y.
4.9. SupFig 9 | C oss-species enhance mapping
• Sequence conse a ion analysis does no ma ch cell- ype-speci ic enhance
usage.
4.10. SupFig 10 | Po al pe o mance
• Benchma king is supe icial, lacking s ess es s o ep oducibili y assessmen s.
5. SYNTHESIS AND OVERALL EVALUATION
5.1. Majo S eng hs o he S udy
• La ge-scale mul i-omics da a collec ion.
• A emp ed in eg a ion o ch oma in accessibili y, his one ma ks, and scRNA-seq.
• Aspi a ional amewo k o a ian in e p e a ion.
• Use ul po al in e ace o communi y access.
5.2. Majo Weaknesses
(1) Rep oducibili y p oblems
• Dono he e ogenei y and inconsis encies in ch oma in p o iling.
• Weak enhance ep oducibili y ac oss indi iduals.
• Unde powe ed CRISPR alida ion.
(2) Compu a ional o e in e p e a ion
• Co-accessibili y ea ed as causali y.
• Machine-lea ning model likely o e i .
• No compa ison o op compe ing models.
(3) Inadequa e alida ion
• Lack o independen da ase s o benchma king.
• Weak CRISPRi e ec sizes and no eplica ion.
(4) O e s a emen s in disease applica ions
• GWAS en ichmen s ci cula and insu icien ly con olled.
• Case s udies che y-picked.
