Multiparameter prediction of myeloid neoplasia risk

The myeloid neoplasms encompass acute myeloid leukemia, myelodysplastic syndromes and myeloproliferative neoplasms. Most cases arise from the shared ancestor of clonal hematopoiesis (CH). Here we analyze data from 454,340 UK Biobank participants, of whom 1,808 developed a myeloid neoplasm 0–15 years after recruitment. We describe the differences in CH mutational landscapes and hematology/biochemistry test parameters among individuals that later develop myeloid neoplasms (pre-MN) versus controls, finding that disease-specific changes are detectable years before diagnosis. By analyzing differences between ‘pre-MN’ and controls, we develop and validate Cox regression models quantifying the risk of progression to each myeloid neoplasm subtype. We construct ‘MN-predict’, a web application that generates time-dependent predictions with the input of basic blood tests and genetic data. Our study demonstrates that many individuals that develop myeloid neoplasms can be identified years in advance and provides a framework for disease-specific prognostication that will be of substantial use to researchers and physicians.

may have had a MN at the *me of sampling (rather than CH) and this MBN was simply undiagnosed.How did the authors address this?Are persons with abnormal blood parameters in MPN or MDS range excluded from subsequent analysis?Clinically many pa*ents have abnormal blood parameters da*ng back a number of years before diagnosis of MPN. 4. Again, specifically focusing on JAK2, the authors report that this occurs as a CH muta*on with a prevalence of 1.9% (Fig 1A).In Fig 3F, the MPN-free survival associated with JAK2 muta*on is between 80% and 50% aqer about 5 years follow up.This seems to predict an extremely high mortality and/or incidence of MPN and would predict many more MPN pa*ents than occur in reality (incidence approximately 1/100,000 per year).5. Related to this, it is unexpected that the incidence of MPN is so much higher than other myeloid neoplasms, almost double that of MDS or AML.What is the explana*on for this? 6. Did MPN include CML (which is associated with CH)? 7.More granularity s required for the 129 of 2045 cases where mul*ple MN were diagnosed contemporaneously.This is not a phenomenon I recognise clinically to occur with such frequency and does raise a concern about veracity of the data, an issue with these large biobanks.8. Given the recent evidence that MPN (and other MN) develop over many decades, with a modest yearon-year fitness advantage, is it not a surprise that CH muta*ons were only found in 32.7% pre-MN cases, typically <10 years before diagnosis).Can the authors model this somehow using data from the literature.The authors used UKBB data to iden*fy individuals with CHIP and then developed a model to predict risk of progression from CHIP to myeloid neoplasia.

Major comment:
The authors have tested their MN-predict model only in healthy volunteers and as the authors note their data is therefore suscep*ble to !healthy volunteer bias".In clinical prac*ce, MN-predict will be most relevant to pa*ents who are found to CHIP or CCUS and referred to a haematology or cardiology clinic for further evalua*on.Can the authors validate the MN-predict model in !real life" clinically relevant cohort of individuals?Figure-specific comments/ques*ons: Figure 1 Frequency of JAK2 = 1.9%,CALR = 0.6% indicates JAK2 muta*on is approx three *mes more frequent than CALR muta*on in healthy individuals.This finding does not align with previously published data of healthy individuals in the popula*on where JAK2 muta*on was found to be approx nineteen *mes more frequent than CALR muta*on, using droplet digital PCR for genotyping: hOps:// pubmed.ncbi.nlm.nih.gov/31217187/How do the authors explain these discordant findings?Do the authors think they are !missing"JAK2V617F muta*ons due to low JAK2 coverage (Supp Table 2)?If so can they postulate how this impacts their findings?Are the authors confident they are restric*ng CALR muta*ons to indel muta*ons in exon 9 that are associated with the development of MPN?
Also, what is the lower VAF limit the authors can detect and does it vary for different muta*ons (I could not find this stated in the manuscript) Figure 2C It is notable that the SRSF2 muta*on is strongly associated with AML, MDS and MPN risk?Can the authors decipher addi*onal factors that determine which MN subtype individuals with SRSF2 muta*ons develop?
Figure 3F The authors have previously reported that JAK2V617F-mutant CH clones have the lowest frac*on of clones growing at a constant rate, as compared to other CH-associated muta*ons (Fabre et al., Nature  2022).Yet in this paper they show that JAK2V617F is the muta*on most strongly associated with MN development and that higher JAK2V617F VAF is associated with higher MPN risk.How do the authors align these somewhat contradictory findings?Figure 3G -no rela*onship between MPL VAF and risk of developing MPN -can the authors explain?It is notable that the MPL muta*ons that the authors include in their analyses include many non MPNassociated MPL muta*ons, some of which are typically germline (Supp Table 1).Can the authors re-do the analysis with MPN-associated MPL muta*ons only?Since CALR is more common than MPL (Supp Figure 2), why is MPN risk for CALR muta*ons not shown?Figure 4 Can the authors show the actual blood values rather then normalized values?Is it possible that a subset of the individuals with elevated PLTs and HgB have undiagnosed MPN? Seeing the actual blood counts would be helpful in this regard.
Minor comments: Figure 1  The authors build a risk predic*on tool for myeloid neoplasms using informa*on on clonal hematopoiesis of indeterminate poten*al (CHIP) as well as several other clinical and demographic variables using data from the UKBiobank.The study was well-conducted and I was impressed with the predic*ve performance of their models.Generally, the story was easy to follow though some of the descrip*ons lacked sufficient detail.I have made several sugges*ons below for strengthening the paper.
Several parts of the Methods sec*on were too brief and require much greater detail, perhaps as addi*onal Supplementary Methods.
The models were run with a Cox PH model and predic*ve ability was measured via AUC.This is not so straighporward and the authors only men*on that !ROC curves were constructed by comparing the probability of developing MN by the final *me point with the real clinical outcomes, using the R package How was death as a compe*ng risk handled?I assume that the authors did not run a Fine-Grey model with a cumula*ve incidence func*on, as this was not men*oned.Did they use a cause-specific hazard func*on?Please provide more details.
The descrip*on of the forward selec*on procedure requires more detail.I could piece together what the authors meant, but it took a while for me to look at Supp Table 5 in order to get it.Please explain this more carefully and in greater detail.Also, what metric of !concordance" was used in the forward selec*on procedure.
The authors should consider using a random survival forest or other ML technique to compare with the Cox PH model.Such methods are known to increase predic*ve performance.

Results
A comparison of the fit sta*s*cs in the training vs test sets would help the reader understand whether the model is overfit and a different model might be considered.
Is there any other risk tool that this can be compared to?If not, I would like to see a comparison of the final risk model with a model that includes the non-CHIP informa*on, i.e., how much is the predic*ve performance enhanced by including informa*on on CHIP?I wasn#t able to use the MN-predict hyperlink so I could not try it out the tool.
Can the authors provide an explana*on/intui*on for why the AUCs are so much beOer for MN in 0-1yr versus 1-5 and >5 years?Supp Table 6 should display be HRs rather than the raw coefficients.
We thank our reviewers for their posi*ve assessments of our manuscript and for their helpful comments and sugges*ons.We have now responded to all comments in full by performing addi*onal analyses, valida*ng our models in two independent cohorts, genera*ng new figures, adding new methodological detail and making several changes/addi*ons to the text.The revised manuscript is significantly improved and its conclusions are very robust as a result.
Below is a point-by-point response to all reviewers' comments, with reference to the relevant changes/ addi*ons to the manuscript.The original reviewers' comments are in black font with our responses in blue.We use yellow highlight to point reviewers to changes in the manuscript text/figure/tables.In the manuscript text, we also use yellow highlight to indicate changes/addi*ons.

Remarks to the Author:
In this study the authors report a further analysis of UK Biobank data with a focus on clonal hematopoiesis (CH) and risk of development of myeloid malignancy.The novelty of the study as ar*culated by the authors is that (1) their analysis is focused on other myeloid neoplasm subtypes in addi*on to acute myeloid leukaemia, (2) they incorporate haematological and biochemistry blood parameters and (3) they develop a risk tool (MPN-Predict) for clinical implementa*on.
It is my view that the "proof of principle that individuals that develop any MN subtype can be iden*fied years in advance" is already rather well established and this does not represent a major step forward in the field.Previous studies have analysed haematological parameters alongside CH muta*ons in the same UKB cohort across all haematological malignancies (e.g.DOI: 10.1038/s41591-021-01521-4).
A personalised risk score for CH pa*ents is poten*ally quite impacpul and is novel, but as there are features of the UK Biobank that might introduce certain biases, this tool requires valida*on in large independent cohorts of pa*ents (rather than quasi-randomisa*on) before clinical implementa*on.
We thank reviewer #1 for their posi*ve overall assessment of our manuscript.We agree that previous studies have demonstrated that MN as a group can be iden*fied years in advance.We wanted to make the point that this had not been demonstrated for individual MN subtypes, as this is per*nent to our Author Rebu*al to Ini-al comments manuscript.However, to avoid misunderstanding, we have changed the relevant text in our abstract to remove the term "proof-of-principle".The relevant sentence now reads as follows: "Our study demonstrates that individuals that develop any MN subtype can be iden7fied years in advance…" (Abstract: line 47).
We also accept that UK Biobank (UKB) par*cipants are healthier than the general popula*on, something that can influence the findings of associa*on studies of clonal hematopoiesis (CH) with nonhaematological diseases.However, there is liOle evidence that the risk of MN is significantly influenced by lifestyle choices, with the possible excep*on of tobacco smoking, which is adequately represented in the UKB.Nevertheless, we en*rely accept that valida*on of our findings in external cohorts would strengthen our work.However, large well-annotated studies equivalent to the UKB are lacking and even if they were available could be subject to healthy volunteer biases.Instead, we have chosen to validate our models in two real-world clinical cohorts of pa*ents with clonal cytopenia of undetermined significance (CCUS) and show that our models perform very well in both (added Introduc*on: line 77-79, Results: line 214-231, Supplementary Tables 8 & 9 and Supplementary Figures 14 & 15).As these cohorts are quite different in composi*on to the UKB, this demonstrates the robustness of our models in different and clinically relevant contexts.
I have some specific comments.Indeed, the incidence of CHIP was 1% lower (in the same dataset) so the authors need to carefully address this.
We thank the reviewer for these excellent points.Applying a VAF cut-off is indeed an effec*ve way to remove genomic noise and sequencing artefacts, in order to improve the accuracy of variant calling.However, we used a slightly different approach to tackle the same problem.Instead of using an arbitrary VAF cut-off, our approach relied on sta*s*cal methods that take the minimum coverage, recurrence and distribu*on of VAFs of pre-filtered calls to remove puta*ve false posi*ves.We have now revised our approach to align it to that used in the aforemen*oned recent paper (Vlasschaert et al, Blood 2023) and adopted their soma*c driver defini*ons for the reasons discussed below.A cut-off of VAF>0.1 was only used for hotspot muta*ons (listed in Supplementary Table 3), as these were iden*fied using Samtools mpileup that does not apply sta*s*cal filters.For these, we used stricter inclusion criteria (VAF>0.1 and ≥3 reads).
To understand the differences between our calling approach and that by Vlasschaert et al, we compared our CH muta*on calling methods against theirs.Overall, both ours and Vlasschaert et al's methods addressed the same problems using similar approach with differences in details/cut-offs.Both methods used Mutect2 to call variants and Samtools to rescue the U2AF1 variants that are systema*cally missed due to an error in the human GRCh38 reference sequence.Also, we both applied a binomial test to the number of mutant reads, in order to remove more germline variants.However, Vlasschaert et al's method had two advantages over our original method, namely that they tested candidate driver muta*ons iden*fied in >20 people for: i) correla*on with age and ii) associa*on with a common gene*c variant in the TERT promoter (rs7705526) which is associated with a higher risk of CH, in order to improve the quality/veracity of driver muta*on calls.We therefore adopted their driver defini*on as outlined in their paper.(Methods: line 359-367) Se•ng a minimum number of mutant reads for calling CH muta*ons, can lead to varying sensi*vity/ specificity on detec*ng CH drivers dependent on sequencing coverage.In order to use the op*mal cutoff for our study, we next inves*gated which number of mutant reads in the Mutect2 output resulted in the best performance of our models.For this we, tested three thresholds, namely: ≥2, ≥3 and ≥5 mutant reads.Expectedly, the ≥3 cut-off gave the most similar data to Vlasschaert et al's paper, as they also used this cut-off (Supplementary Table 4).However, despite following their muta*on calling approach, we detected ~2,000 more TET2 muta*ons that Vlasschaert et al (muta*on in other genes agreed closely).Looking at TET2, we found that our number of muta*ons at each recurrent hotspot also agreed well with Vlasschaert et al.However, we also iden*fied ~2,000 low-recurrence TET2 muta*ons (each with ≤3 cases in the UKB), including 1437 nonsense muta*ons and 604 missense muta*ons in the TET2 func*onal domains (p.1104-1481 and 1843-2002, excluding known germline/artefact -see Supplementary Table 1), which meet their driver defini*ons, but were not listed in the supplementary data by Vlasschaert et al.Importantly, both missense and nonsense low-recurrence TET2 muta*ons collec*vely showed posi*ve correla*ons with age, mirroring the behavior of known drivers (e.g.TET2 I1873T) and in clear contrast with known germline mutants (e.g.TET2 Y867H) -see Figure R1A), indica*ng that most were likely to be drivers.Notably, inclusion of these low-recurrence muta*ons did not affect predic*ve model performance (Figure R1B-E).Based on these reasons, we believe it is appropriate to include these lowrecurrence TET2 muta*ons as CH drivers.(A) UKB par*cipants were grouped into 2-year age bins in order to examine the age distribu*on of low recurrence TET2 nonsense muta*ons and missense muta*ons in func*onal domains (as defined in Supplementary Table 1).In each age bin, the frac*on of individuals with the depicted muta*on(s) was calculated (as the propor*on of the total number of individuals with that muta*on in the UKB) and then normalized to the total number of individuals in that bin.A known driver (I1873T) and germline muta*on (Y867H) are included as posi*ve and nega*ve controls.We then built our models (including ones for stepwise regression for feature selec*on and the final models) on the muta*ons called using each of these cut-offs (≥2, ≥3 and ≥5 mutant reads), and found that all 3 cut-offs gave similar model performances, but "≥2 mutant reads" gave higher concordance scores for the AML model compared to the other 2 cut-offs (Supplementary Figure 18).Based on these observa*ons, we chose "≥2 mutant reads" as our final cut-off.This cut-off improved the AUC of ROC curves of AML predic*on from 0.73 to 0.78 for the >5y category (Figure 4A before vs aqer revision).The 0-1y AUC decreased from 0.93 to 0.88 but this is less of an impact on the overall predic*ve performance, due to the small number of samples in this *me range (n=11).
In the manuscript, we now: -refer to the use of Vlasschaert et al's method in Results: lines 84-86, -explain this in more detail in Methods (Line 356-364), -update Supplementary Table 1 to match their driver defini*on -add Supplementary Table 4 to show muta*on calls and VAFs -add Supplementary Figure 18 to show the effect of cut-offs on our model performance In consequence of changing the muta*on calling approach, we have now iden*fied slightly more CH muta*ons in the UKB, which leads to many changes of numbers throughout the manuscript highlighted in yellow, as well as updates of majority of the results: -Supplementary Table : 1, 4, 6, 7 2. Of the 648 pa*ents with a previous diagnosis of myeloid neoplasm (line 89), were these included in the 21362 individuals described with CH in the previous paragraph (line 83)?
The 648 individuals with a prior diagnosis of MN were included in the analysis of 454,340 UKB whole exome sequences searched for the presence of CH.Out of these 648 pa*ents, we detected CH muta*ons in 233 and these were included in the 21,362 individuals (now 22,735 aqer revisions outlined in our response to Point 1) with CH driver muta*ons.However, all 648 were excluded from subsequent analyses and model development/tes*ng.To make this clearer, we have now added this bracketed clause in the second paragraph of Results: "(of whom 233 had CH driver muta7ons)" Result: line 97.
3. The incidence of JAK2 CHIP is rather low compared to studies using targeted analysis for JAK2V617F e.g.DOI: 10.1182/blood.2019001113 in a Danish popula*on.This may relate to VAF sensi*vity?This Danish cohort is notable as many of the pa*ents had hematological parameters consistent with a diagnosis of MPN (without a formal MPN diagnosis).This raises a broader concern that pa*ents sampled may have had a MN at the *me of sampling (rather than CH) and this MBN was simply undiagnosed.How did the authors address this?Are persons with abnormal blood parameters in MPN or MDS range excluded from subsequent analysis?Clinically many pa*ents have abnormal blood parameters da*ng back a number of years before diagnosis of MPN.
As alluded to by the reviewer, it is highly likely that the lower prevalence of JAK2 V617F in our study, compared to the Danish popula*on study by Cordua et al, is due to differences in assay sensi*vity.In fact Cordua et al used digital droplet PCR, a highly sensi*ve method that detected JAK2 V617F down to a VAF 0.009%.In that study, 508 of 613 (83%) JAK2 V617F cases had VAF<1%, which is below the sensi*vity of exome sequencing.We now discuss the study by Cordua et al in Discussion: lines 296-300).
As regards the possibility of undiagnosed MPN, we agree that this is a valid considera*on and we thank the reviewer for poin*ng it out.Although a diagnosis of MPN cannot formally be reached without ruling out alterna*ve reasons for a raised hemoglobin or platelet count, we do concede that many individuals in the UKB had undiagnosed MPN.For this reason, we isolated all UK par*cipants that met the latest diagnos*c criteria 1 for Polycythemia Vera (PV) and Essen*al Thrombocythemia (ET), in par*cular we used the following criteria: PV: JAK2 muta*on + Hb >16.5g/dl (men) or Hb >16.0g/dl (women) These are the less stringent cut-offs that normally require a bone marrow biopsy confirma*on, but we apply them here to ensure that we exclude all individuals who had PV at the *me of blood sampling.
ET: JAK2 or CALR or MPL muta*on + Plts ≥450 x 10 9 /l The requirement for a BM biopsy was waved, in order to ensure that we exclude individuals who had ET at the *me of blood sampling A diagnosis of myelofibrosis (MFS) cannot be made on blood test results, but any individuals with MFS and high platelet counts will have been removed by the ET criteria above.
Like MFS, a diagnosis of MDS cannot be made on blood test results and requires a bone marrow biopsy.
In fact cytopenias like anemia are rela*vely common in the absence of MDS and, even in the presence of soma*c muta*ons the diagnosis can oqen be CCUS.So whilst it is possible that some of the UKB par*cipants may have had MDS, we could not iden*fy them with any degree of confidence, so we could not apply the same approach as we used for PV and ET.
We have now described the removal of likely undiagnosed MPN cases from training/valida*on of our models in Result: line 97-102 and Methods: 388-391.This leads to many changes in the numbers of cases throughout the manuscript highlighted in yellow, and updates of majority of analyses: We are grateful for this comment as we had similar considera*ons when analyzing our data.The reported incidence of MPN varies widely 2 and this may be because the registra*on of MPNs by cancer registries was very variable prior to the introduc*on of ICD-O-3.Before this, ICD10 was in use and the MPNs were classified as "D" codes (i.e.not considered as cancer and thus not rou*nely registered).So whilst many of the published studies es*mate an incidence of 1 per 100,000, more up-to-date and specialist studies significantly higher incidence rates.For example, in a large European study, MPN incidence was found to be 3.1/100,000 per year 3 and in a UK "real-world" study it was found to be 6/100,000 per year 4 .Also, the median age of UK biobank par*cipants was 58 years (i.e. higher than the median for popula*on studies) at the *me of blood sampling and the median follow up was 12.6 years (range 7.4-15.5 years).The annual incidence of MPN in this age range is significantly higher and a recent study found it to be 18.6/100,000 in those aged 70-80 years 5 .If we apply this number to the UKB, the expected number of cases of MPN over this period would be: Incidence* number of par<cipants * years (of follow-up) = 0.000186 x 454,340 x 12.6 = 1064 This number is close to the observed number of MPNs in our study (n=1000) and in a recent study inves*ga*ng germline risk of MPN amongst a very similar, albeit not iden*cal, UKB par*cipant popula*on (n=1086) 6 .
However, please note that the 1.9% in Figure 1A refers to the percentage of cases of CH that carried JAK2 V617F (rather than the % of UKB par*cipants carrying JAK2 V617F).We have now made this clearer in the legend to Figure 1A: line 113.
5. Related to this, it is unexpected that the incidence of MPN is so much higher than other myeloid neoplasms, almost double that of MDS or AML.What is the explana*on for this?As discussed above, the likely explana*on for this is that the incidence of MPN is variable and overall higher than previously realized.

Did MPN include CML (which is associated with CH)?
We did not include CML in our studies as this disease is directly linked to the acquisi*on of the BCR-ABL fusion gene and is outside the scope of our study.
7.More granularity is required for the 129 of 2045 cases where mul*ple MN were diagnosed contemporaneously.This is not a phenomenon I recognise clinically to occur with such frequency and does raise a concern about veracity of the data, an issue with these large biobanks.
We thank the reviewer for this comment.We looked carefully at these 129 MN cases and decided that it was safer to remove them as we could not be certain which of the MN subtypes (AML, MDS, MPN or CMML) arose first or if an overlap syndrome was present.
For clarity, 71 of the 129 cases were diagnosed with more than one MN within 35 days (0-35 days, mean 5.2 days).In 60/71 cases, AML was one of the two diagnoses, the other being MDS (n=37) or CMML (n=16) or MPN (n=10).In the remaining 11/71 cases MDS was listed in 10, MPN in 9 and CMML in 5.In 6/71 cases, three diagnoses were listed within 35 days.We believe that, in several of these cases, the pa*ents had dysplasia with increased blasts at a % blasts that placed them at the MDS/AML or CMML/ AML cut-off (10-20%).Other cases may have had an MPN/MDS overlap syndrome.As we could not be certain of the nature of these MNs and as our aim was to generate models that predict individual MNs separately, we opted to not include them into any of our three models (AML, MDS or MPN).
In the remaining 58 of 129 cases, the second MN diagnosis was made >35 days aqer the first (36-1839 days, mean 497 days).AML was the first diagnosis in all 58 cases and followed by MDS in 28 cases, MPN in 8 and CMML in 28.In 6/58 cases, three diagnoses were listed (within 172-1386 days).It is probable that some of these individuals had AML and then a treatment-associated MN or a return to a pre-AML neoplasia (e.g.MDS, MPN or CMML) that was present subclinically prior to AML diagnosis.As we could not dis*nguish between these possibili*es, we did not use these cases to train or validate our MN predic*ve models.
Overall, as our aim was to develop different predic*ve models for each of the main MN subtypes, we decide that omi•ng these cases was the safest path, par*cularly as their numbers are rela*vely small to substan*al aid model performance.
As regards the veracity of MN diagnoses in the UKB, it is very reassuring that both the gene*c muta*ons (Figure 2B) and the blood count results (Figure 4D-F and Supplementary Figure 3B-D), reflect those of the downstream MN subtype.
We now summarize the above in Methods under "Data acquisi*on" in Methods: line 336-341 8. Given the recent evidence that MPN (and other MN) develop over many decades, with a modest yearon-year fitness advantage, is it not a surprise that CH muta*ons were only found in 32.7% pre-MN cases, typically <10 years before diagnosis).Can the authors model this somehow using data from the literature.
We thank the reviewer for this interes*ng sugges*on.We recently described the longitudinal behavior of smaller JAK2 V617F clones over 10-15 years and found these to expand less predictably than any other type of CH (Fabre et al, Nature 2022, hOps://doi.org/10.1038/s41586-022-04785-z) 7.The unpredictable behavior of JAK2 V617F clones can also be inferred from the Danish study by Cordua et al, where the prevalence of small JAK2 V617F clones is significantly higher than that of larger clones, whilst the same comparison for CALR clones shows that a larger % of the total CALR clones expand to a large size (also discussed in response to the second comment by reviewer 2).
Our findings here suggest that larger clones (detectable by exome sequencing) expand more rapidly and predictably than smaller ones, but we are not aware of any other large dataset in the literature that includes sufficient numbers of larger clones for model development.We previously performed a "lookback" study of 12 individuals that developed JAK2 V617F posi*ve MPN 4.  In supplementary Figure 3 we wanted to display all blood test parameters in a single plot, so had to use the same quan*fica*on scale for all variables and opted for ten quan*les (Q1-Q10).Color-coding shows enrichment/ deple*on (expressed as Odds Ra*os) in each of the ten quan*les for each of the parameters, allowing for a visualiza*on of trends for each MN subtype.For example, platelet counts (PLT) are enriched in Q10 (high) for MPN, but not for AML or MDS, for which they are enriched in Q1 (low).
To make comparisons between MN categories more visually apparent, we have now used the same rank order of variables across all four MN categories in Supplementary Figure 3.
11.I was surprised that certain chromosomal abnormali*es, strongly associated with myeloid malignancy such as 5q-or 9pLOH did not refine their model.What is the explana*on for this?Most of the We thank the reviewer for this fair point.We did observe significant associa*on between chromosomal copy number changes (mCAs) and MN risk.However our models' AUCs ROC curves did not improve significantly by incorpora*ng mCA informa*on most probably due to: a) a rela*vely low number of cases with mCAs and b) linear dependency between mCA and blood/biochemistry parameters especially for pre-MDS/MPN cases.In other words, the presence of mCA affected blood/biochemistry parameters, which therefore captured the increased risk associated with the mC A. In light of this and as mCAs are not rou*nely captured by standard diagnos*c assays, we have not included mCA into our final models.
Nevertheless, in order to avoid readers thinking that mCAs are not associated with MN risk, we have now added two panels to Supplementary Figure 7A & B, displaying the significant associa*ons between pre-AML cases and -5q, pre-MDS and -5q/4qLOH, and pre-MPN and 9pLOH/+9p/+9 in the UKB and have added this text to the Results: Lines 178-184.
12. Minor point, the panel in Fig 1 states that lasso regression is used to smooth curves in D & E (panel E does not show curves) Thank you for spo•ng this.LASSO regression was use to smoothen curves C and D (rather than D and E).
We have now corrected this in Figure 1 legend: line 115-119.

Reviewer #2:
Remarks to the Author: The authors used UKBB data to iden*fy individuals with CHIP and then developed a model to predict risk of progression from CHIP to myeloid neoplasia.

Major comment:
The authors have tested their MN-predict model only in healthy volunteers and as the authors note their data is therefore suscep*ble to "healthy volunteer bias".In clinical prac*ce, MN-predict will be most relevant to pa*ents who are found to CHIP or CCUS and referred to a haematology or cardiology clinic for further evalua*on.Can the authors validate the MN-predict model in "real life" clinically relevant cohort of individuals?
We thank the reviewer for this valid comment.The UKB does have a healthy volunteer bias, in the sense that people joining the UKB were more likely to be health conscious and display reduced cardiovascular and some other risks.However, as the risk of de novo myeloid cancers is not significantly influenced by lifestyle, with the possible excep*on of tobacco smoking, there is no reason to consider that the UKB cohort differs substan*ally from the general popula*on with regards to myeloid neoplasia risk.As such, we believe that our model will perform well when such individuals are referred to CHIP/CCUS clinics and agree that valida*on in real life cohorts is important and we now include valida*on of our models in two independent "real life" CCUS cohorts in the revised manuscript.
Specifically, we applied our MN-predict model to two independent CCUS cohorts: i) the "Leeds CCUS cohort" composed of 204 pa*ents with CCUS recruited from 2014-2016 and followed up for up to 5.5 years (mean ± sd = 3.0 ± 1.7) and ii) the "Pavia CCUS cohort" composed of 312 pa*ents with CCUS and followed up for up to 15.1 years (mean ± sd of diagnosis = 4.4 ± 3.6).We found that our models performed very well in predic*ng progression to AML or MDS in the Leeds cohort and MDS in the Pavia cohort (this cohort only had 2 cases of progression to AML, so we could not test our AML model here).
These finding provide robust support for the relevance of our models to real life cohorts/pa*ents.As the reviewer suspected, this shows that small JAK2 V617F clones are much more abundant than small CALR-mutant clones, but the difference is much less pronounced for large clones.Our study used exome sequencing, so was only able to iden*fy larger clones for which the JAK2-V617F:CALR ra*o was 3.03, which is close to the Cordua et al ra*o for large clones.
It is also worth no*ng that Cordua et al acknowledge that they may have underes*mated the CALR muta*on rate, as their approach was only able to detect the common CALR muta*ons (type 1 and type 2), but not the less common/recurrent CALR muta*ons (collec*vely ~15-20% of all CALR muta*ons).This would have overes*mated the JAK2-V617F:CALR ra*o, which should have been lower than 19.Our approach detected all types of CALR muta*ons.
With regards to how "missing" small JAK2-V617F clones may impact our findings, we believe that the impact is likely to be small as the expansion of small JAK2 clones is known to be less predictable (Fabre et al, Nature 2022 and also evident by the data of Cordua et al above).By contrast, clones that reach a large size are manifestly much more likely to progress to MPN.As the main aim of a model such as MN-predict is to iden*fy individuals at high risk, the detec*on of larger clones should adequately capture most highrisk individuals.Also, in clinical prac*ce, anyone with a small JAK2-V617F clone would be monitored and highlighted as high risk if/when their clone expands.
We have now discussed this in Discussion: line 296-300 Are the authors confident they are restric*ng CALR muta*ons to indel muta*ons in exon 9 that are associated with the development of MPN?
Yes, we only selected MPN-associated frameshiq muta*ons of CALR in Exon 9 as described in Supplementary Table 1.Out of the 134 CALR muta*ons, 63 were L367Tfs (Type 1, 52bp dele*on site) and 50 were K385Nfs (Type 2, 5bp inser*on site) and 21 were other types of less common CALR exon 9 muta*ons.These rela*ve frequencies are in keeping with the MPN literature.We have now added Supplementary Table 4 to make this and other muta*on calls explicit.
Also, what is the lower VAF limit the authors can detect and does it vary for different muta*ons (I could not find this stated in the manuscript) Our muta*on calling was based on a combina*on of two pipelines: Mutect2 (used for all muta*ons) and Samtools mpileup (used specifically to iden*fy lower VAF hotspot muta*ons).For the filtering of Mutect2 output, we examined the recurrence of the called variant and the distribu*on of VAFs across the cohort and retained calls that passed our one-sided exact binomial test (described in Methods: Line 361-364), to remove germline muta*ons.For the Samtools output of hotspot muta*ons, we used the rela*vely strict threshold of "≥3 reads and a VAF>0.1" as these calls did not pass Mutect2 filters.For both methods, the ability to iden*fy muta*ons depended on the sequencing coverage of different genes/ exons.So the lower limit of muta*on VAF varied between different genes/posi*ons.To make this clear, we now provide sta*s*cs of VAF distribu*ons for recurrent muta*ons in Supplementary Table 4, as well as our summary of read coverage by gene in Supplementary Table 3.

Figure 2C
It is notable that the SRSF2 muta*on is strongly associated with AML, MDS and MPN risk?Can the authors decipher addi*onal factors that determine which MN subtype individuals with SRSF2 muta*ons develop?
This is an excellent observa*on by the reviewer as SRSF2 is the gene represented in appreciable numbers in all three pre-MN groups (Figure 2B).With regards to MPN, we believe that cases of SRSF2 CH that progressed to MPN are likely to have developed myelofibrosis and we previously showed that SRSF2 muta*ons can be acquired before JAK2 V617F in such cases 8 .With regards to AML vs MDS, our data suggests that SRSF2/TET2 co-mutated cases were more likely to develop MDS (9/12) than AML (3/12) (see Fig 2B & Supplementary Figure 3), whilst SRSF2/IDH2 co-mutated cases were more likely to develop AML (6/9) than MDS (3/9).These differences are captured by our predic*ve models, as are those that vary by blood count results (e.g. raised MCV or reduced platelets are more common amongst individuals that develop MDS than AML (Fig 4D -E).So whilst no individual predictor is specific to AML or MDS, a number of indicators make one or other more likely.We now allude to this in Discussion: line 283-285 The authors have previously reported that JAK2V617F-mutant CH clones have the lowest frac*on of clones growing at a constant rate, as compared to other CH-associated muta*ons (Fabre et al., Nature  2022).Yet in this paper they show that JAK2V617F is the muta*on most strongly associated with MN development and that higher JAK2V617F VAF is associated with higher MPN risk.How do the authors align these somewhat contradictory findings?
This is an interes*ng point that we also discuss in response to Reviewer #1's comment 8.In short it is clear that small JAK2 V617F clones behave less predictably than larger clones.Our current study iden*fied muta*ons through exome sequencing, so could not iden*fy small clones.For our longitudinal manuscript (Fabre et al, Nature 2022), we used deep sequencing and most of the clones iden*fied were small (VAF<1%) at study entry, with only 3 of 12 clones exceeding a VAF of 10% during the 15 year follow-up period.By contrast, most clones iden*fied in the UKB had a VAF >5% (Supplementary Table 4).This supports the premise that the behaviour of large clones is much more predictable/determinis*c than large clones.A similar conclusion can be reached from the data by Cordua et al (see response to the second comment of this reviewer), with a very large number of small clones and a much smaller number of small clones.We have now discussed this in Discussion: line 296-300 Figure 3G -no rela*onship between MPL VAF and risk of developing MPN -can the authors explain?It is notable that the MPL muta*ons that the authors include in their analyses include many non MPNassociated MPL muta*ons, some of which are typically germline (Supp Table 1).Can the authors re-do the analysis with MPN-associated MPL muta*ons only?
Since CALR is more common than MPL (Supp Figure 2), why is MPN risk for CALR muta*ons not shown?
We thank the reviewer for this comment that pertains to Figure 3G, which shows that small and large MPL clones have similar MPN-free survival.We apologize as this was the result of an error on our part, as we used the wrong VAF sizes in this par*cular plot.We have now replaced this panel with the equivalent plot for CALR muta*ons as per our reviewer's sugges*on (Figure 3G).In any case, aqer removing JAK2, MPL or CALR cases with blood test results compa*ble with possible undiagnosed MPN (in response to reviewer #1's comment 3), the number of MPL-mutant cases was reduced to only 3, making a Kaplan-Meier curve inappropriate.The first parameter (VAF) in the upper panel of Figure 3A is a measure of clonal size irrespec*ve of the mutated gene.We have now clarified this by changing "VAF" to "VAF of largest clone" in Figure 3A, and also modified Supplementary Figure 6 legend and Supplementary Figure 8 legend.

Reviewer #3:
Remarks to the Author: The authors build a risk predic*on tool for myeloid neoplasms using informa*on on clonal hematopoiesis of indeterminate poten*al (CHIP) as well as several other clinical and demographic variables using data from the UKBiobank.The study was well-conducted and I was impressed with the predic*ve performance of their models.Generally, the story was easy to follow though some of the descrip*ons lacked sufficient detail.I have made several sugges*ons below for strengthening the paper.
Several parts of the Methods sec*on were too brief and require much greater detail, perhaps as addi*onal Supplementary Methods.We thank our reviewer for these comments and for the sugges*ons below that have certainly helped us improve our manuscript.We have now provided methodological details in Methods, conducted new analyses and added several new supplementary figures.
The models were run with a Cox PH model and predic*ve ability was measured via AUC.This is not so straighporward and the authors only men*on that "ROC curves were constructed by comparing the probability of developing MN by the final *me point with the real clinical outcomes, using the R package "pROC"."Please expand on this descrip*on as I did not quite understand exactly how this was done.This is an excellent sugges*on, as CHIP muta*ons are responsible for many of the blood test changes.
We have now analyzed how model performance is influenced by the inclusion of informa*on on CHIP (soma*c muta*ons).Overall, models lacking muta*on informa*on performed reasonably well, but their performance (ROC) improved by the addi*on of this informa*on (Figure R2, A vs B, C vs D and E vs F).This improvement was most significant for AML, a disease in which blood count changes appear late in its evolu*on (vs MPN and MDS).We also tested how well blood test data predicted MN risk in cases who lacked muta*ons.Expectedly, blood test results were less good at predic*ng future MN in this context.Taken together these observa*ons suggest that muta*ons cause change in blood tests and these capture some of the risk associated with these muta*ons.However, the type and VAF of muta*ons are of addi*onal predic*ve value and it is important to retain these in our models.I wasn't able to use the MN-predict hyperlink so I could not try it out the tool.
We apologize for this.We have now provided the URL instead of aOaching a hyperlink in the manuscript text.Results: line 256 and Data availability: line 481.You can also follow the URL here: hOps:// muxingu.shinyapps.io/webappCan the authors provide an explana*on/intui*on for why the AUCs are so much beOer for MN in 0-1yr versus 1-5 and >5 years?
This is an excellent comment that probably alludes to the biology of how CH develops towards a MN: As CH expands in size to represent a larger propor*on of circula*ng blood cells, it begins to impose its own characteris*cs on the blood count.These characteris*cs are muta*on-dependent, for example SRSF2 and SF3B1 muta*ons may lead to lower HGB, raised MCV and reduced platelet counts (reflected in the pre-MDS changes in Figure 4E), whilst JAK2, CALR and MPL clones lead to a raised platelet and/or hemoglobin values, reflected in the pre-MPN blood counts (Figure 4F).By contrast, most common pre-AML muta*ons (e.g.DNMT3A, TET2 and ASXL1) have a lesser impact on blood count results (reflected in the pre-AML counts in Figure 4D).As we get nearer to the MN diagnosis, clones become bigger and so does their impact on blood count parameters.In other words, both the VAF size and the FBC changes become more pronounced closer to the diagnosis.
We believe that the reason for the improvement in AUC as we move closer to the diagnosis, is that larger clones have a more determinis*c behavior, both in terms of their clonal expansion trajectories and the likelihood of acquisi*on of addi*onal muta*ons (that are usually required to engender diagnoses such as AML and MDS).The determinis*c behavior is best illustrated by JAK2 V617F clones that behave unpredictably when they are small, but become a lot more predictable when they increase in size (see responses to reviewers #1 & #2 pertaining to this).Evidence for increased muta*on acquisi*on is reflected in the fact that many cases of pre-MN have mul*ple driver muta*ons, whilst the majority of CH cases in UKB have a single muta*on.It is also probable that having mul*ple muta*ons has an exacerbated impact on blood counts, increasing its effect on the AUC.
Another factor behind the improved AUC nearer to diagnosis may be that normal varia*on may mask subtle changes in blood counts that occur when CH clones are s*ll small (a very large number of samples would overcome this).
We now comment on this in Discussion: line 296-300.
Supp Table 6 should display be HRs rather than the raw coefficients.
Thank you -we have corrected this in the Supplementary Table 7.
16th Jun 2023 Dear Dr. Vassiliou, Thank you for submi•ng your revised manuscript "Mul*parameter predic*on of myeloid neoplasia risk" (NG-A61755R).It has now been seen by the original referees and their comments are below.The reviewers find that the paper has improved in revision, and therefore we'll be happy in principle to publish it in Nature Gene*cs, pending minor revisions to sa*sfy the referees' final requests and to comply with our editorial and forma•ng guidelines.
If the current version of your manuscript is in a PDF format, please email us a copy of the file in an editable format (Microsoq Word or LaTex)--we can not proceed with PDFs at this stage.
We are now performing detailed checks on your paper and will send you a checklist detailing our editorial and forma•ng requirements soon.Please do not upload the final materials and make any revisions un*l you receive this addi*onal informa*on from us.
Thank you again for your interest in Nature Gene*cs Please do not hesitate to contact me if you have any ques*ons.

Decision Le*er, first revision:
In the abstract I would suggest to add the word some*mes (or similar): "…..our study demonstrates that individuals that develop any MN subtype can SOMETIMES be iden*fied years…."Otherwise the implica*on is that MN can always be iden*fied years in advance which I do not believe is the case.
Overall I congratulate the authors on an excellent study, I tried using the portal they have developed and it is very informa*ve and will be helpful to apply clinically.
Reviewer #2 (Remarks to the Author): Thank you for your responses.No further ques*ons.
One comment/sugges*on: In the *me since this manuscript was submiOed, the clonal hematopoiesis risk score (CHRS) has been published: hOps://evidence.nejm.org/doi/full/10.1056/EVIDoa2200310 Since this used the same UKBB data and apparently the same Pavia CCUS valida*on cohort, it would be appropriate to reference this publica*on in the discussion.
Reviewer #3 (Remarks to the Author): The authors have responded to all of my comments.I have no further edits or recommenda*ons.
11th Jul 2023 Dear George, I am delighted to say that your manuscript "Mul*parameter predic*on of myeloid neoplasia risk" has been accepted for publica*on in an upcoming issue of Nature Gene*cs.
Over the next few weeks, your paper will be copyedited to ensure that it conforms to Nature Gene*cs style.Once your paper is typeset, you will receive an email with a link to choose the appropriate publishing op*ons for your paper and our Author Services team will be in touch regarding any addi*onal informa*on that may be required.
Aqer the grant of rights is completed, you will receive a link to your electronic proof via email with a request to make any correc*ons within 48 hours.If, when you receive your proof, you cannot meet this deadline, please inform us at rjsproduc*on@springernature.com immediately.

Final Decision Le*er:
You will not receive your proofs un*l the publishing agreement has been received through our system.
Due to the importance of these deadlines, we ask that you please let us know now whether you will be difficult to contact over the next month.If this is the case, we ask you provide us with the contact informa*on (email, phone and fax) of someone who will be able to check the proofs on your behalf, and who will be available to address any last-minute problems.
Your paper will be published online aqer we receive your correc*ons and will appear in print in the next available issue.You can find out your date of online publica*on by contac*ng the Nature Press Office (press@nature.com)aqer sending your e-proof correc*ons.Now is the *me to inform your Public Rela*ons or Press Office about your paper, as they might be interested in promo*ng its publica*on.This will allow them *me to prepare an accurate and sa*sfactory press release.Include your manuscript tracking number (NG-A61755R1) and the name of the journal, which they will need when they contact our Press Office.
Before your paper is published online, we shall be distribu*ng a press release to news organiza*ons worldwide, which may very well include details of your work.We are happy for your ins*tu*on or funding agency to prepare its own press release, but it must men*on the embargo date and Nature Gene*cs.Our Press Office may contact you closer to the *me of publica*on, but if you or your Press Office have any enquiries in the mean*me, please contact press@nature.com.
Acceptance is condi*onal on the data in the manuscript not being published elsewhere, or announced in the print or electronic media, un*l the embargo/publica*on date.These restric*ons are not intended to deter you from presen*ng your data at academic mee*ngs and conferences, but any enquiries from the media about papers not yet scheduled for publica*on should be referred to us.
Please note that <i>Nature Gene*cs</i> is a Transforma*ve Journal (TJ).Authors may publish their research with us through the tradi*onal subscrip*on access route or make their paper immediately open access through payment of an ar*cle-processing charge (APC).Authors will not be required to make a final decision about access to their ar*cle un*l it has been accepted.<a href="hOps:// www.For authors selec*ng the subscrip*on publica*on route, the journal's standard licensing terms will need to be accepted, including <a href="hOps://www.nature.com/nature-porpolio/editorial-policies/self-archiving-and-license-to-publish.Those licensing terms will supersede any other terms that the author or any third party may assert apply to any version of the manuscript.
Please note that Nature Porpolio offers an immediate open access op*on only for papers that were first submiOed aqer 1 January, 2021.
If you have any ques*ons about our publishing op*ons, costs, Open Access requirements, or our legal forms, please contact ASJournals@springernature.com If you have posted a preprint on any preprint server, please ensure that the preprint details are updated with a publica*on reference, including the DOI and a URL to the published version of the ar*cle on the journal website.
To assist our authors in dissemina*ng their research to the broader community, our SharedIt ini*a*ve provides you with a unique shareable link that will allow anyone (with or without a subscrip*on) to read the published ar*cle.Recipients of the link with a subscrip*on will also be able to download and print the PDF.
As soon as your ar*cle is published, you will receive an automated email with your shareable link.
You can now use a single sign-on for all your accounts, view the status of all your manuscript submissions and reviews, access usage sta*s*cs for your published ar*cles and download a record of your refereeing ac*vity for the Nature journals.
An online order form for reprints of your paper is available at <a href="hOps://www.

9 .
Do the authors have any data on the muta*ons and VAF at *me of diagnosis on MN? 10.It is very difficult to see trends in blood cell parameters in different MN subgroups in Supp Fig 3 due to the normalisa*ons used.It would be easier if blood parameters were shown according to usual units.11.I was surprised that certain chromosomal abnormali*es, strongly associated with myeloid malignancy such as 5q-or 9pLOH did not refine their model.What is the explana*on for this? 12. Minor point, the panel in Fig 1 states that lasso regression is used to smooth curves in D & E (panel E does not show curves) Reviewer #2: Remarks to the Author: Figure 3A -the first gene

!
pROC"."Please expand on this descrip*on as I did not quite understand exactly how this was done.Did the authors use one of the defini*ons from Heagerty and Zheng (2005)?[Heagerty PJ, Zheng Y. Survival model predic*ve accuracy and ROC curves.Biometrics.2005;61(1):92-105].If so, which one?Perhaps an illustra*ve example would be helpful.

1 .
The methods state that VAF had to be >0.1 for predic*ve models yet many of the figures show pts with VAF <0.1.Oqen an arbitrary cut off of VAF 2% is used to define CHIP.Can the authors carefully clarify exactly what VAF cut-off was used and how this relates to the VAF shown in the figures.Perhaps in some cases they refer to VAF as a % and in other cases as a propor*on?The variant calling should be carefully benchmarked against published approaches.Both VAF and variant calling for UK biobank are d e s c r i b e d i n a r e c e n t p a p e r h O p s : / / d o i .o r g / 1 0 . 1 1 8 2 / b l o o d . 2 0 2 2 0 1 8 8 2 5 [eur03.safelinks.protec*on.outlook.com]where the authors highlight poten*al sequencing ar*facts.
(B) Hazard ra*os (HR) for different types of MN obtained from models trained without low-recurrence TET2 muta*ons.HRs show no significant difference from the original models.(C-E) Exclusion of low-recurrence TET2 muta*ons from our MN predic*on models did not affect model performance as assessed by area under curve (AUC) of ROC curves for (C) AML, (D) MDS and (E) MPN.

10 .
It is very difficult to see trends in blood cell parameters in different MN subgroups in Supp Fig 3 due to the normalisa*ons used.It would be easier if blood parameters were shown according to usual units.

Figure 4
Figure 4Can the authors show the actual blood values rather then normalized values?Is it possible that a subset of the individuals with elevated PLTs and HgB have undiagnosed MPN? Seeing the actual blood counts would be helpful in this regard.This is a valid point and we have now replaced normalized values with actual blood counts/units in Figure4 D-F.Also, we have now removed 108 cases with probable undiagnosed MPN (see reviewer #1, comment 3).

Figure 11 )
Figure 11) indica*ng that overfi•ng or underfi•ng is unlikely.We also discussed this in the manuscript Results: line 208-210.

Figure R2 :
Figure R2: Comparison of models using blood and biochemistry parameters only (excluding muta-onal parameters) with original models that includes muta-onal parameters ROC curves from Cox propor*onal hazard models for predic*on of progression to MNs, computed from predicted 15-year probability of MN-free survival and the diagnosis within the 15-year follow-up period.Individuals having at least 1 CH muta*ons (Mut) and individuals having no CH muta*on (NoMut) were ploOed separately.AUC=area under curve.(A) AML model using genotype and blood/biochemistry parameters.(B) AML model using blood/ biochemistry parameters only.(C) MDS model using genotype and blood/biochemistry parameters.(D) MDS model using blood/biochemistry parameters only.(E) MPN model using genotype and blood/biochemistry parameters.(F) MPN model using blood/biochemistry parameters only.
The new cohorts and analyses are now included in the revised manuscript: Introduc*on: line 77-79, Result: line 214-231, Discussion: line 305-308 and Methods: line 446-464.We also added Supplementary Figures 13 & 14 and Supplementary Table 8 & 9 to show these results.
springernature.com/gp/open-research/transforma*ve-journals">Find out more about Transforma*ve Journals</a> Authors may need to take specific ac*ons to achieve <a href="hOps://www.springernature.com/gp/open-research/funding/policy-compliance-faqs"> compliance</a> with funder and ins*tu*onal open access mandates.If your research is supported by a funder that requires immediate open access (e.g. according to <a href="hOps://www.springernature.com/gp/open-research/plan-s-compliance">PlanS principles</a>) then you should select the gold OA route, and we will direct you to the compliant route where possible.
nature.com/ reprints/author-reprints.html">hOps://www.nature.com/reprints/author-reprints.html</a>.Please let your coauthors and your ins*tu*ons' public affairs office know that they are also welcome to order reprints by this method.If you have not already done so, we invite you to upload the step-by-step protocols used in this manuscript to the Protocols Exchange, part of our on-line web resource, natureprotocols.com.If you complete the upload by the *me you receive your manuscript proofs, we can insert links in your ar*cle that lead directly to the protocol details.Your protocol will be made freely available upon publica*on of your paper.By par*cipa*ng in natureprotocols.com,you are enabling researchers to more readily reproduce or adapt the methodology you use.Natureprotocols.com is fully searchable, providing your protocols and paper with increased u*lity and visibility.Please submit your protocol to hOps:// protocolexchange.researchsquare.com/.Aqer entering your nature.comusername and password you will need to enter your manuscript number (NG-A61755R1).Further informa*on can be found at hOps://www.nature.com/nature-porpolio/editorial-policies/repor*ng-standards#protocols