Introduction 

For many years proteomics type mass spec experiments were all acquired using the same basic approach, Data Directed Analysis (DDA), but now a new acquisition methodology is becoming more popular Data Independant Analysis (DIA). In this post I’ll describe both processes, compare their performance and describe where we stand in the BioMS Facility on this topic. The principles of DIA vs DDA are the same for both peptides and small molecules/metabolites. This article is predominately focused on peptides but the comments are valid for both molecular classes, though the data analysis software is different. 

DDA: Data Directed Analysis 

Data directed analysis or DDA has been around for nearly 30 years, since Micromass (now Waters) introduced the first hybrid quadrupole-time of flight (QTOF) mass spectrometer. This system was able to get a high-resolution profile of the molecules (peptides) present in the sample at any time point in an LC-MS run and then automatically select them individually to fragment them for subsequent identification with the cutting edge Mascot software (if you were European) or Sequest software (if you were American). The systems were comparatively slow and therefore you needed to design your acquisition method to make the most of the time that was available. Method optimization included setting what mass range and charge states you were interested in, the intensity threshold that had to be exceeded before a peak was selected, when and for how long a mass would be ignored after it was fragmented and the number of different peptides that would be fragmented per analysis cycle. As equipment improved these settings became tighter and faster, allowing more and more fragmentation spectra to be acquired per experiment. 

Quantification in DDA is obtained from the precursor data acquired before fragmentation (MS1). This is done by summing the intensities of the signal across the peak. In a simplistic model the spectral intensity of a single ion or a small mass range are summed together. In a more ‘analytical’ approach an extracted ion chromatogram (EIC) is created and the peak area is integrated by drawing a baseline across the peak and looking at the area or maximum height above that line. The EIC route is considered the most accurate route as it allows for background signal contamination.

The challenge with DDA is that fragmentation data is required to identify each molecule to be quantified and if it wasn’t identified then it could not be quantified, resulting in many missing values. A couple of software approaches were developed to address this issue. The first was Non-Linear Dynamics’ Progenesis software, which was developed from their earlier 2D gel scanning software. In essence that software converted the MS1/pre-fragmentation data and warped each run so that it could determine with confidence that the signal seen in one run was the same as another seen in a different run. If the peptide was identified in one run then that identification was mapped across that signal in all runs. They called this process QI, ie Quantification before Identification. Most other quantification software performed Identification before Quantification (IQ) but the same principles were brought into Jurgen Cox’s MaxQuant software and termed Match Between Runs (MBR), a term now used in most similar software. In this process identification is performed across all runs and a master list of the m/z and elution time for each peptide is created. This list is then used to extract quantitative data from each run so that each peptide is looked for in each run, whether it was initially identified in that run or not. This process was controversial, with many claiming that there was not enough confidence that a MBR result was the equivalent of a non-MBR analysis. From a BioMS perspective, this process provided excellent results from standard samples like a mixed lysate and therefore we use MBR by default. 

The latest iterations of DDA analysis software uses machine learning based models to increase the confidence of identifications and allow the identification of multiple peptides from a mixed spectra. This works by using massive datasets of identified spectra to build a model of how a peptide will fragment and its retention time. There are two main groups driving this approach. The Nesvizhskii lab has used this approach on publicly available data and created MSFragger software, which is now part of the FragPipe collection of tools. The Kuster lab took a different approach and received major funding to synthesise 100,000’s of peptides covering most common sequence and collect high quality and confidence data for their models, which have now been commercialized by MSAID in the form of CHIMERYS software incorporated in Thermo’s Proteome Discoverer software (PD). CHIMERYS software and PD are particularly good at handling mixed spectra for increasing the quantified peptides from DDA data. 

BioMS uses both MSFragger and CHIMERYS software for their data. MSFragger is particularly regarded for its ability to handle post-translational modifications and CHIMERYS/PD software is by far the best option for analysing DDA data from Thermo mass spectrometers, providing datasets up to 50% larger than other software. 

DIA: Data Independent Analysis 

Data Independent Analysis or DIA has also been around for a long time, in this case 20 years, and was designed to address the sampling issue in DDA analysis. The concept is that you fragment all the ions entering the mass spectrometer from the sample and then use software to pull out quantitative data for individual molecules/peptides. In this manner there should be few or no missing values, with the data often classed as being a ‘digital fingerprint’ of a sample that can be reinterrogated as further analytes of interest are identified. 

Quantification in this case is from the fragmentation data (MS2). This is often classed as a more ‘analytical’ approach as it provides a much cleaner signal, in fact it is the basis of most commercial mass spec quantification. The reason for this is that it is not uncommon for multiple molecules to occupy the same space/overlap in MS1 but they rarely have the same ions when fragmented. The signals in MS2 data are spread over multiple ions and therefore the absolute signal strength is lower however system sensitivity is defined by the signal to noise ratio (S/N) and in MS2 the noise is reduced by more than the signal and therefore MS2 is considered to more sensitive as it has greater S/N. Although DIA mostly uses MS2 software like Spectronaut also allows MS1 quantification and this can be more applicable at very low sample loading levels.

There is some debate about who started the concept of ‘fragment all ions’ but the best known early example is Micromass’s (now Waters) MSE (“MS to the E”) in 2003, where their Q-TOF systems fragmented all the ions introduced to the MS without any selection. This produced a complex dataset that needed substantial computing power to deconvolute via their ProteinLynx Global Server (PLGS) software. This took a long time to perform and needed dedicated and expensive workstations. The approach was further developed by Waters to include new ion mobility options from their Synapt systems (HD-MSE) and using quadrupole scanning (SONAR). However, none of those methodologies were widely adopted by the proteomics field, probably due to the time requirements for analysis and sensitivity limitations of the MS equipment used for data generation. 

DIA was successfully commercialized by Sciex with the introduction of their TripleTOF Q-TOF type mass spectrometers and Serial Window Acquisition of all THeoretical fragment ion spectra (SWATH). The mass spectrometers were fast (10’s of spectra per second) and so the mass range could be divided into a number of overlapping regions (‘windows’) where all ions in those regions are fragmented. It was seen that the performance of this process was inversely related to the size of the window and therefore a process of variable sized windows was introduced whereby smaller widows were used in busier regions of the mass range (generally lower m/z’s) and larger windows in quieter regions (generally larger m/z’s). This process was needed as the number of windows in each cycle determined the ‘cycle time’ ie how long it took to acquire data over all the windows. As cycle times increased the number of points per peak reduced to a level where there was not enough for quantification and the risk of missing low signals increased. 

At this time the software used for data analysis required a library of peptide spectra and their retention times. To obtain this library the samples were often pooled and then fractionated using high pH reversed phase chromatography or ion exchange chromatography offline. Each fraction was then analysed using DDA methodologies to identify the peptides and create the libraries. This was a very time consuming process and the limits of DDA analysis meant that a reasonable level of each peptide was required and sensitivity was not as good as DIA data was capable of. 

A revolution came to DIA analysis via similar machine learning based models that aided DDA. In this case the two pieces of software that drove these developments was the academic software DIA-NN (Data Independent Analysis using Neural Networks) by Vadim Demichev from the Ralser Lab and Spectronaut from the contract analysis company Biognosys (now part of Bruker). Both of these software developed a ‘Library Free’ capability, where the peptide libraries were created by prediction models instead of experimentally. Note that both Library and Library-Free DIA approaches use libraries, just one is produced experimentally and the other by prediction. BioMS uses both DIA-NN and Spectronaut and both perform well, though Spectronaut is commercial and thereby much better supported, easier to get comparative data from and is generally the most reliable tool. 

Recent leaps forward in DIA analysis have been predominately driven by hardware, although the analysis software is still developing rapidly. The Sciex TripleTOF 7600 has had new Zeno trapping capability added, where ions are collected before separating in the analyser, boosting sensitivity, however the current battle is between Bruker and Thermo. Bruker introduced the TimsTOF system that was based on Q-TOF technologies but also included a Trapped Ion Mobility MS cell that provided rapid ion mobility separations which are then scanned in the TOF analyser providing extremely high data rates and complex spectra that combine separations by both mass and ion mobility. This data did not work well for DDA analysis as the quality of the MS1 data was poor but the introduction of PASEF-DIA acquisition was massively successful and became the dominant form of DIA. 

Thermo mass spectrometers are the dominant systems for DDA but were challenged with DIA as their orbitrap analysers were too slow to perform DIA analysis with smaller window sizes and the faster they were ran the lower the resolution they produced. DIA data on their later orbitrap centric systems such as the Exploris 480, was relatively competitive but did not compete well with the fastest TOF based systems. This issue was corrected with the release of the new Orbitrap Astral system, where the orbitrap analyser was coupled with a new asymmetric multi-reflecting TOF analyser and new single ion detectors. The new system was capable of collecting DIA data at 200 spectra a second with extremely high sensitivity. This speed meant that variable windows were no longer required and small windows of 2Da can be used over the entire acquisition range. 

DIA vs DDA 

The latest DIA datasets using TimsTOF or Astral mass spectrometers are unprecedented in scale and sensitivity quantifying 10,000+ proteins from a single abundant sample or even multiple thousands of proteins from a single cell. This performance isn’t without controversy though. As the software gets better at predicting data, less and less conventional data is required for a confident match. Whereas previously, multiple ions were required to be seen from the same peptide and the quantitative data must demonstrate a classical peak shape, the new algorithms confidently identify and quantify peptides using data that would not normally be accepted. The challenge thereby is to determine the level of confidence you require in your data. In early phase experiments you may accept more false positives in order to acquire the largest dataset possible, whereas later you may require more confident data. DIA data also currently has issues predicting the performance of post-translationally modified peptides, though new tools are improving this. 

Overall, the general shift is to move over to DIA acquisition on the new generation of mass spectrometers (the TimsTOF Pro 2 and Ultra from Bruker and the Orbitrap Astral from Thermo at the time of writing, Dec 2024) except for specific circumstances when looking at post-translational modifications. This does not mean that DDA has no value of that previous data cannot be mined further. 

Looking forward, it is likely that DIA will remain dominant but there may be further leaps forward in data analysis where holistic AI driven pattern analysis is used on the raw data before any peptide identification is performed, conceptually an evolution of the earlier Progenesis QI approach in DDA. BioMS will continue to monitor developments in hardware and software and assessing its capability to robustly support your research. 

Related