Considerations for the inclusion of MS data in our studies (with sources)

[heleensev]

This issue is about mass spectrometry data (MS), we have not made a decision about whether to include the data or not. We want our affinity predictor to have no bias towards peptides that are processed in the cell (antigen processing or AP). This is because we want our predictor (DeepRank) to solely focus on modelling the interaction based on physio-chemical features. MS data has this bias because the pMHC (HLA bound to peptide) was eluted from human cells. Furthermore, the MS
technique has the limitation that highly hydrophobic peptides cannot be detected, and cysteine containing peptides are also harder to measure)

Here are some sources (and conclusions from these sources) that will help us make a decision about the inclusion of MS data in our experiments:

• MHCflurry 2.0 uses both BA (affinity) data and MS (mass spectrometry) for their affinity predictor in a 31/69 ratio in order not to bias their predictor too much to antigen processing (MS pMHC samples have been processed by the cell, BA pMHC are not necessarily processed by the cell but might be)
  • MHCflurry 2.0 only uses decoys as MS negatives for their BA predictor training data: "random sequences sampled according to the same amino acid distribution as the hits" The ratio of negative peptides to MS hits is not clear.
  • The MS hits are assigned a measurement value of "< 100nm" in order not to influence the training loss too much "..MS training data does not guide the relative ranking (exact IC50) of strong binders"
    • The section "Considerations of using MS data" also mentions the use of random negative peptides instead of real MS negatives will not bias their BA predictor to peptides that have been filtered by AP. Possible caveat: how do we know that random negative peptides are not actually binders?
  • The cysteine depletion in MS data is addressed when discussing the AP predictor: In one AP predictor experiment cysteine containing peptides were removed in both training and testing, it showed no significant decrease in AUC. Caveat: why not test cysteine-containing peptides and see if they are disproportionally classified as negative, their check does not seem robust
  • And in the discussion: "An important limitation of this work is that we apply datasets of MHC class I ligands detected by MS both to train and to benchmark our predictors. Assay biases, which we expect are modeled by the AP predictor, have the potential to erroneously inflate our accuracy scores. Although the main known bias, depletion of cysteine, does not seem to have a dramatic effect on AP predictive accuracy, we cannot rule out the contributions of other kinds of bias"
  • The BA training set is comprised of 493,473 MS entries and 219,596 affinity measurements. The inclusion of MS will greatly increase the amount of (training) data of our experiments.
• Citations in the paper also address biases or artefacts in MS data:
  • MHCflurry 2.0 refers to Abelin et al. 2017 This paper is also a source of 4,808 mono-allelic MS hits. In this paper a small experiment is done to show a lack a peptide distribution bias between BA (from IEDB) and MS data (from their paper). For this experiment a predictor (ESP) is used from Fusaro et al. 2019 which was trained to measure "highly responding" peptides for detection with MS. The density plot shows that BA samples are almost just as likely detected by MS as the MS samples.
• We should take a look at the protocols from MHCflurry's (consequently our) main MS data sources (papers) and maybe find a subselection of studies that we believe will not introduce bias to the peptide distribution:
  • Immune Epitope Database (IEDB)(Vita et al., 2019) on April 27, 2020. (estimate: 305,043 hits)
  • SysteMHC Atlas project(Shao et al., 2018) (downloaded on May 6, 2019) (46,880 MS hits)
  • Sarakizova et al(Sarkizova et al., 2020) ( 136,742 MS hits)
  • Abelin et al (Abelin et al., 2019) (4,808 MS hits)
• Lastly we should also consider the biases that BA data may contain. The MHCflurry 2.0 dataset contains samples from many different protocols, measurement types (competive/direct/..) from many different studies.
  • Abelin et al. 2017 also mentions affinity biases: "Meanwhile, other unintentional forms of bias, such as pre-existing notions of the length distribution or limitations on peptide synthesis and solubility, are difficult to avoid"

[gcroci2]

Great great job @heleensev :) I've read the paper and your notes, and I'll put here my thoughts as well. Let's further discuss this tomorrow at the weekly meeting with the others.

Ideas:

• Train our models with exactly the same training set of MHCFlurry 2.0 BA predictor (~ 700k samples, MS + BA data) and evaluate on the same data (held-out from training set data, monoallelic and multiallelic benchmarks)
  • Good starting point, independently from the quality of the data
  • Our models should of course outperform MHCFlurry 2.0 BA. Our claims would be that we're adding chemical-physical and structural information and that we're using more sophisticated architectures.
• Test our models on more difficult cases, see how they perform, and compare with SOTA. Maybe it's already much better than SOTA (in such case, there is no strong need to modify the training set).
• If needed, remove data according to our considerations, and develop more reliable and accurate models
  • Test all SOTA methods on the same testing data/benchmarks

Comments:

• MHCFlurry 2.0 BA predictor performances on rare alleles (such as HLA-C) are good (see figure S1). But the alleles tested are almost all binders, and this makes the evaluation not robust.
• HLA-C alleles are in the training set of MHCFlurry 2.0 BA (~ 80k samples). We want to exclude such alleles from the training set, and then test on them. But since only HLA-A, HLA-B, and HLA-C alleles are present in humans, why should we exclude HLA-C from the training set? If no different alleles will appear in clinical potential applications, it's not clear to me why we should do such exclusion. The others argue that we want to develop a predictor able to classify also potential new alleles.
• even if only HLA-A, HLA-B, and HLA-C are present in humans, the training is done also on other alleles (and other species). Indeed such mechanisms are similar across species and we can have more data this way. It would be good for us as well to include other species.

Specific questions on MHCFlurry training dataset (S3 in the supplemental material)

• Before testing the predictor on the benchmarks, they tested it on data held-out from the training set. Did they just exclude the alleles listed in figure S1, or did they remove only some samples for each allele (stratification)? This was not clear to me from the paper.
• How can I understand which ms data are hits and which ones are decoys?
• As both the MHCflurry binding affinity predictor and NetMHCpan 4.0 do not incorporate multiallelic MS in their training sets [...] Looking at the S3 data (should be their training set), considering only mass_spec measurement kind, I see several different alleles (A, B, C, E, G, ...).