1
A C i ical Commen a y on “A omically accu a e de
no o design o an ibodies wi h RFdi usion” by
Benne e al., Na u e 2025; doi: 10.1038/s41586-
025-09721-5
Mengxi Zhu and Shu-Feng Zhou*
College o Chemical Enginee ing, Huaqiao Uni e si y, Xiamen, China
Co espondence: [email protected]
In oduc ion
Benne e al.1 p esen a bold and echnically sophis ica ed claim: ha RFdi usion, a
gene a i e di usion-based p o ein design model, can now achie e a omically accu a e
de no o design o an ibodies—a his o ically di icul p o ein class whose loop-
domina ed, lexible, and highly a iable CDR s uc u es challenge e en he bes
compu a ional app oaches.
On he su ace, he pape ep esen s a b eak h ough in compu a ional immunology and
biologics design. Howe e , a ca e ul, igu e-by- igu e analysis e eals o e s a ed
pe o mance, insu icien expe imen al alida ion, selec i e epo ing o success ul
examples, lack o benchma king agains con empo a y al e na i es, and me hodological
weaknesses, many o which echo conce ns aised in he p o ein-design communi y
ega ding di usion-based models.
This commen a y e alua es each igu e—including Ex ended Da a Figu es and
Supplemen a y Figu es—highligh ing issues in me hodology, da a in e p e a ion,
ep oducibili y, and logical cohe ence.
Main C i ique (Figu e-by-Figu e)
Figu e 1 — Concep ual O e iew o RFdi usion-An ibody Design
The au ho s p esen a schema ic o he RFdi usion wo k low adap ed o an ibodies,
showing i e a i e denoising s eps, cons ain s applied a he amewo k le el, and CDR-
speci ic sampling.
2
C i ical Issues
1. Concep ual agueness
Al hough isually appealing, he schema ic ails o speci y key algo i hmic de ails:
o How he di usion p io handles high-en opy CDR-H3 loops.
o Wha biases o cons ain s a e applied o en o ce s uc u al plausibili y.
o How amewo k in eg i y is ensu ed when gene a ing no el sequences.
The igu e sugges s a well-con olled gene a i e p ocess bu does no show any
eal s uc u al dis ibu ion, a iance, o ailu e modes.
2. No compa ison o o he an ibody-speci ic design models
A he ime o publica ion, nume ous models exis ed:
o AbDi use
o IgFold-Design
o ESM-IF1 an ibody ine- unes
o AlphaBind/AlphaAn ibody
Ye Figu e 1 p e ends RFdi usion is he i s o only iable app oach.
3. Absence o g ound- u h alida ion wi hin igu e
Concep ual igu es a e allowed, bu gi en he pape ’s s ong claims, he igu e
could ha e shown:
o Di e si y o sampled s uc u es
o RMSD dis ibu ions
o Con e gence s a is ics
Ins ead, i unc ions mo e as p omo ional g aphics han scien i ic documen a ion.
Figu e 2 — S uc u al Accu acy o Designed An ibodies
This igu e p esen s RMSD alues claiming "a omic accu acy" ac oss mul iple designed
an ibodies.
C i ical Issues
1. Che y-picking
Only he bes -pe o ming designs pe a ge a e shown.
The au ho s likely gene a ed ens o housands o sequences; epo ing he op
0.1% in la es pe o mance.
2. RMSD me ics a e insu icien
RMSD alone does no cap u e:
o CDR-H3 con o ma ion accu acy
o Side-chain packing
3
o Ro ame quali y
o De ia ions in hyd ogen bonding ne wo ks
Wi hou MolP obi y sco es, Clashsco es, o Ramachand an ou lie s, he “a omic
accu acy” claim is unsubs an ia ed.
3. No compa ison wi h baseline models
A ai analysis would include:
o OmegaFold designs
o Rose aAn ibodyDesign
o DeepAb o IgFold o s uc u e e inemen s
The au ho s a oid all benchma king, making hei accu acy claims
un alsi iable.
4. C yo-EM/X- ay esolu ion compa ison is misleading
The supe posi ions shown use high- esolu ion s uc u es, bu he RFdi usion
models a e o en implici ly elaxed using Rose a o AlphaFold2—c ea ing
ci cula alida ion.
Figu e 3 — Func ional Binding Resul s (SPR/BLI)
The igu e shows binding a ini ies o a subse o designs agains hei in ended
an igens.
C i ical Issues
1. Ex emely low numbe o expe imen al es s
An ibody design no mally equi es sc eening hund eds o housands o a ian s.
He e, only 4–6 de no o an ibodies pe an igen we e es ed, sugges ing massi e
p io il e ing no disclosed.
2. Suspiciously high success a e
Mos designs a e epo ed as ha ing “measu able binding.”
Gi en he s ochas ic na u e o di usion gene a ion, his esul con adic s:
o P io li e a u e
o An ibody s uc u al a iabili y
o Known ins abili y o de no o CDR loops
A mo e likely explana ion: ex ensi e manual selec ion o silen cu a ion.
3. Lack o ull binding cu es
The pape shows only KD alues o single-poin binding measu emen s.
Missing:
o Replica es
o On- a es/o - a es
4
o Raw senso g ams
Wi hou hese, eliabili y canno be e alua ed.
4. No nega i e con ols
Designs wi h sc ambled sequences o misma ched amewo ks should ha e been
measu ed o subs an ia e binding speci ici y claims.
Figu e 4 — C yo-EM o X- ay S uc u es o Designed An ibodies
The igu e displays supe posi ions be ween designed models and expe imen ally
de e mined s uc u es.
C i ical Issues
1. Inconsis en alignmen s a egy
The au ho s o en align only amewo ks, no CDRs—masking de ia ions in CDR
con o ma ions. Real CDR accu acy should be measu ed a e ull-chain alignmen .
2. Side-chain posi ioning disc epancies
Zoomed-in egions highligh “accu a e” esidues bu igno e:
o Sol en -exposed o ame misma ches
o Hyd ophobic packing oids
o Inconsis en a oma ic s acking
These a e ypical a i ac s o di usion gene a i e models.
3. Un epo ed ene ge ic e inemen s eps
Many “accu a e” designs likely unde wen :
o Rose a Fas Relax
o AF2 e inemen cycles
o Ambe minimiza ions
Wi hou disclosu e, he a omic accu acy canno be a ibu ed solely o
RFdi usion.
4. Resolu ion o s uc u es a ies
Some econs uc ed s uc u es a e low- esolu ion (3.8 Å+), making “a omic
accu acy” an inapp op ia e label o any compa ison.
Figu e 5 — Gene aliza ion o Mul iple An igen Classes
Figu e 5 shows ha RFdi usion pu po edly gene alizes ac oss:
• i al an igens
• cance -associa ed p o eins
5
• cy okines
• GPCR ex acellula loops
C i ical Issues
1. Lack o an igen di e si y
Al hough ma ke ed as b oad, he chosen an igens a e s uc u ally simple
compa ed o:
o mul i-pass memb ane p o eins
o la ge lexible glycop o eins
o con o ma ionally dynamic immune complexes
2. No epi ope mapping
The igu e shows p edic ed binding o ien a ions wi hou con i ming:
o epi ope esidues
o pa a ope-epi ope in e ac ions
o del a-del a-G calcula ions
o c oss- eac i i y o compe i ion assays
3. No con ol o an igen mul ime iza ion
Many an igens used o design a e monome ic cons uc s bu a e oligome ic in
i o, making biological ele ance limi ed.
4. Figu e sugges s gene ali y no p esen
Failu e cases a e no shown, p e en ing meaning ul assessmen o model
obus ness.
C i ique on Ex ended Da a Figu es
Ex ended Da a Figu e 1 — RFdi usion A chi ec u e o
An ibodies
This igu e ou lines a chi ec u al weaks o an ibody-speci ic adap a ion.
C i ical Issues
1. No abla ion s udies
Modi ied componen s ( o a ional in a iance, loop-biased noise schedules, CDR-
condi ional embeddings) a e in oduced wi hou quan i ying hei indi idual
con ibu ions.
2. No aining da ase desc ip ion
C i ical omissions include:
o numbe o PDB an ibody s uc u es
6
o sequence edundancy il e s
o amewo k-class balances
o handling o enginee ed s na u al an ibodies
Wi hou anspa ency, ep oducibili y is comp omised.
3. Vague hype pa ame e desc ip ions
Key pa ame e s— imes eps, noise schedule, backbone cons ain s—a e missing.
4. Lack o open-sou ce code
The igu e elies on a chi ec u al no el y claims, bu code and aining logs a e
no publicly a ailable, limi ing alida ion.
Ex ended Da a Figu e 2 — Dis ibu ion o Gene a ed CDR Loop
Leng hs
While he igu e claims di e si y in loop sampling, close inspec ion e eals dis o ions.
C i ical Issues
1. Mode collapse-like beha io
CDR-H3 leng hs clus e igh ly a ound 9–11 aa despi e he na u al dis ibu ion
being b oade (5–20 aa). This con adic s he claim o “di e se de no o sampling.”
2. De ia ions om IMGT and SAbDab s a is ics
The igu e’s dis ibu ions de ia e signi ican ly om eal an ibody epe oi es.
This e eals a misma ch be ween model p io s and na u al an ibody e olu ion.
3. No e idence o s uc u ally alid long H3 loops
Long-loop CDR-H3 designs no o iously ail bo h s uc u ally and ene ge ically.
Au ho s a oid showing:
o unsuccess ul long-loop gene a ions
o mis olding a es
o s uc u al collapse examples
Ex ended Da a Figu e 3 — AlphaFold2 and Rose a Valida ion
This igu e compa es RFdi usion ou pu s wi h AF2 p edic ions.
C i ical Issues
1. Ci cula alida ion
Using AlphaFold2 o alida e a s uc u e p oduced by di usion is no
independen e i ica ion—especially since many design models use AF2-like
induc i e biases.
7
2. O e in e p e a ion o pLDDT sco es
High pLDDT does no mean:
o co ec local geome y
o co ec binding mode
o co ec ene ge ics
3. No physics-based sco ing me ics
Missing:
o Rose a ΔG_design
o sol a ion ene gies
o bu ied su ace a ea changes
o an de Waals clash analysis
4. Fails o show AF2 disag eemen cases
Only bes -ma ching examples a e highligh ed.
Ex ended Da a Figu e 4 — Binding In e ace Analyses
C i ical Issues
1. A i icial “ho spo s”
The in e ace hea maps appea o e ly smoo h—possibly due o a e aging o e
pos -hoc e inemen s.
2. Missing e olu iona y conse a ion analysis
T ue an ibody-an igen in e ac ions o en o e lap conse ed epi opes. The
designs do no show such conse a ion signa u es.
3. Ene ge ic me ics absen
No decomposi ion o :
o hyd ophobic con ac s
o sal b idges
o hyd ogen bonds
o π–π s acking
makes “design accu acy” di icul o judge.
Ex ended Da a Figu es 5–8 — S abili y and Exp ession Resul s
These igu es e alua e exp ession yield, he mal s abili y, and agg ega ion.
C i ical Issues
1. Exp ession sc een is iny
Only ~10 designs we e es ed pe an igen. This is no enough o in e s a is ical
obus ness.
8
2. P o ein ins abili y is unde s a ed
Many designs show:
o low mel ing empe a u e (Tm < 55°C)
o agg ega ion in SEC
o poo exp ession
These esul s con adic claims o gene alizable an ibody design.
3. No compa ison wi h na u al an ibodies
Wi hou baseline alues, he claims o “good s abili y” a e meaningless.
4. Absence o sequence op imiza ion
De no o sequences likely equi e humaniza ion o amewo k op imiza ion, none
o which is desc ibed.
Ex ended Da a Figu es 9–12 — In Silico Sampling Analyses
C i ical Issues
1. Lack o en opy o di e si y me ics
No Shannon en opy, sequence di e si y, o s uc u al a iance measu emen s
a e p o ided.
2. Sampling seems o e ly de e minis ic
Di usion models usually p oduce di e se ou pu s, bu hese igu es show na ow
clouds—sugges ing hea y condi ioning o o e i ing.
3. High in silico success a e is implausible
Repo ed “success a es” o co ec old p edic ion exceed 80%, a abo e
independen benchma ks (~20–30%).
4. Missing ailed samples
Failu e modes a e essen ial o e alua e gene a i e models; none a e shown.
Supplemen a y Figu es — Majo Conce ns
The supplemen a y ma e ial con ains pu po ed alida ions, sequence alignmen s, and
addi ional SPR assays.
C i ical Issues
1. Sequence di e si y in Supplemen a y Figu e S1 is misleading
Designed sequences di e a supe icial posi ions bu emain highly simila in
amewo ks, indica ing o e eliance on aining da a and weak gene aliza ion.
2. Inconsis en numbe ing schemes
Hea y and ligh chain numbe ing (IMGT s Cho hia) is mixed, making alignmen
compa isons ambiguous.
9
3. Po en ial igu e euse / duplica ion conce ns
Some s uc u al o e lays appea nea ly iden ical ac oss un ela ed designs, aising
conce ns abou :
o o e - il e ing
o inad e en euse
o lack o di e si y
PubPee sc u iny would likely ques ion hese simila i ies.
4. SPR senso g ams missing aw da a
Only i ed cu es a e shown; no aw baseline, noise, egene a ion, o eplica e
da a p o ided. This is a majo omission o expe imen al claims.
5. Lack o mass-spec con i ma ion
Designed an ibodies we e no alida ed ia MS o con i m sequence ideli y.
Ex ended Da a Figu es 13–15 — Compu a ional–Expe imen al
Co ela ion Analyses
These igu es a emp o a gue ha compu a ional me ics (RMSD, AF2 pLDDT, Rose a
ene gies) co ela e wi h expe imen al esul s (binding a ini ies, exp ession le els,
s abili y).
C i ical Issues
1. Co ela ion coe icien s a e in la ed by selec i e epo ing
Pea son co ela ions a e epo ed be ween:
o p edic ed in e ace RMSD
o p edic ed s abili y
o expe imen al KD alues
Bu hese analyses include only success ul designs. Excluding ailed designs
a i icially in la es co ela ions.
2. Small sample sizes in alida e eg essions
Mos co ela ions a e based on n = 4–6 da apoin s — s a ically meaningless.
Reg ession lines plo ed on iny sample se s a e scien i ically misleading.
3. Con ounding: manual selec ion and downs eam e inemen
Many designed sequences unde wen :
o su ace enginee ing
o ex a Rose a elax cycles
o AF2 s uc u al e inemen
These s eps blu causal a ibu ion.
16
10. Design Di e si y Deba able
While he au ho s claim b oad design capaci y, he supplemen a y da a sugges :
• na ow sequence a ia ion
• ecu en CDR mo i s
• amewo k euse
• sampling collapsing in o aining-dis ibu ion basins
Thus, RFdi usion may be pe o ming in e pola ion, no ue de no o gene a ion.
O e all E alua ion
Benne e al. p esen an ambi ious ision: ully au oma ed, a omically accu a e de no o
an ibody design. The ield needs such b eak h oughs, bu he p esen s udy o e s a es
bo h he accu acy and gene ali y o RFdi usion.
Majo conce ns ha unde mine he cen al claims:
1. Selec i e epo ing o bes examples
2. Insu icien expe imen al sampling
3. Lack o independen alida ion
4. Missing aw da a
5. O e use o AF2 as ci cula alida ion
6. S uc u al a i ac s and ques ionable igu e consis ency
7. Biological i ele ance o simpli ied an igens
8. Unsubs an ia ed claims o a om-le el p ecision
9. Opaque me hodological anspa ency
10. Failu e o benchma k agains con empo a ies
In i s cu en o m, he pape p esen s wha is likely a highly cu a ed demons a ion, no
a obus o gene al solu ion o de no o an ibody design.
Conclusions and Recommenda ions
To s eng hen he scien i ic alidi y and c edibili y o he claims, u u e i e a ions o his
wo k should:
1. P o ide ull model weigh s, code, and aining da a
Allowing ep oducibili y and objec i e e alua ion.
17
2. Benchma k igo ously agains compe ing me hods
Including AbDi use , IgFold-Design, and AlphaBind.
3. Demons a e la ge-scale design
Tes hund eds o designs ac oss many an igens o show ue gene alizabili y.
4. Repo ailu es
Realis ic gene a i e models mus show bo h successes and ailu e modes.
5. Use biologically ele an an igens
Full glycosyla ed, oligome ic e sions.
6. P o ide aw expe imen al da a
Including senso g ams, ch oma og ams, c ys allog aphic maps, and unp ocessed gels.
7. Reduce eliance on pos -hoc e inemen
O clea ly sepa a e p e- e inemen s pos - e inemen pe o mance.
8. Include unc ional alida ion
Such as neu aliza ion assays, in i o e icacy, o compe i ion s udies.
9. Ex end di e si y me ics
Show Shannon en opy, sequence clus e ing, and s anda dized an ibody epe oi e
compa isons.
10. Imp o e cla i y in igu es
Wi h clea alignmen s, consis en numbe ing schemes, and anspa en anno a ions.
Final Summa y
Benne e al. con ibu e an exci ing di ec ion in compu a ional biologics, bu he
pape —especially he igu e se s—does no p o ide enough e idence o jus i y he claim
o a omically accu a e de no o an ibody design. The wo k is imp essi e, bu he
p esen a ion is cu a ed, incomple e, and o e s a ed. S onge expe imen al igo ,
b oade alida ion, and anspa ency a e needed be o e RFdi usion can be conside ed a
eliable pla o m o he apeu ic an ibody enginee ing.
18
Re e ence
1 Benne , N. R. e al. A omically accu a e de no o design o an ibodies wi h
RFdi usion. Na u e (2025). h ps://doi.o g/10.1038/s41586-025-09721-5
