Therapeutic Product Immunogenicity Community

I wanted to post this here because as the paradigm for ADA assays shifts, something has been gnawing at me. I'm fully on board with moving away from the three tier approach, and nothing I'm raising here is "fixed" by keeping it. But if we're now using language like "ADA are a biomarker," then I think there are analytical implications to that shift that I haven't seen well explained. Thank you to Robert for the recent presentation, it pushed me to try to make my thoughts more coherent.

Before ADA appears, the only analytical variance we can characterize is assay noise because we have no relevant surrogate to measure analytical variability of an ADA response. The mixture is too heterogeneous to represent meaningfully until it exists in a study sample. But once ADA appears, the assay is measuring an analyte, a mixture yes (IgM, IgG, low affinity, high affinity), but that mixture is the analyte.

At that point the meaningful question becomes: how consistently can we measure that analyte?

I've heard the argument that because the analyte mixture may shift from day to day or timepoint to timepoint, that variability should be treated as biological. But that doesn't sit well with me if we are treating ADA assays like biomarker assays. With other biomarkers, we do not assume precision; we measure it using endogenous samples because they contain the analyte in the true matrix. ADA assays seem to be the only place where we go straight from "the signal is above assay noise" to "this might be biologically meaningful," without ever asking whether the measurement itself is stable enough to make that determination.

Consider a simple hypothetical. Take one ADA positive sample, split it into identical aliquots, freeze them, and run them on different days. You might get S:N values of 5, 7, 10, 6, 9. The biology is unchanged, same antibodies, same complexes, yet the measured ADA response swings twofold simply because the assay was run on a different day. To me, that level of variation is potentially meaningful. If modelers receive a data transfer with a S:N of 5 vs 10, they may arrive at a different conclusion as to what is clinically meaningful response. It is also important to point out that those divergent interpretations come entirely from analytical variation, repeated measures of the same sample.

So if ADA is a biomarker and S:N is the quantity we intend to report, the three-tier approach and cutpoints are gone, but analytical rigor doesn't go away, it shifts. It is worth asking whether we are now responsible for understanding the uncertainty around the S N measurement itself.  If a sample is above assay noise, do we need to re-analyze it multiple times to understand the analytical variation around that sample before drawing PK, PD, or clinical conclusions? That is exactly how other ligand binding biomarker assays are handled: precision is established against the real analyte (endogenous) in the real matrix. Why would ADA be the exception?   Because the analyte mixture changes from time point to time point?  Because it's harder?

If we don't want to go to the level of reporting a CV or a SD from a n=6 of every sample above the assay noise – I think the median of an n=3 would be a far more accurate S:N measurement, on average, than simply going with the S:N we received in the first test.

And just to be clear, the goal here is not to do extra work, but logically that seems to be the direction it is heading. Thoughts?

I wanted to post this here because as the paradigm for ADA assays shifts, something has been gnawing at me. I'm fully on board with moving away from the three tier approach, and nothing I'm raising here is "fixed" by keeping it. But if we're now using language like "ADA are a biomarker," then I think there are analytical implications to that shift that I haven't seen well explained. Thank you to Robert for the recent presentation, it pushed me to try to make my thoughts more coherent.

Before ADA appears, the only analytical variance we can characterize is assay noise because we have no relevant surrogate to measure analytical variability of an ADA response. The mixture is too heterogeneous to represent meaningfully until it exists in a study sample. But once ADA appears, the assay is measuring an analyte, a mixture yes (IgM, IgG, low affinity, high affinity), but that mixture is the analyte.

At that point the meaningful question becomes: how consistently can we measure that analyte?

I've heard the argument that because the analyte mixture may shift from day to day or timepoint to timepoint, that variability should be treated as biological. But that doesn't sit well with me if we are treating ADA assays like biomarker assays. With other biomarkers, we do not assume precision; we measure it using endogenous samples because they contain the analyte in the true matrix. ADA assays seem to be the only place where we go straight from "the signal is above assay noise" to "this might be biologically meaningful," without ever asking whether the measurement itself is stable enough to make that determination.

Consider a simple hypothetical. Take one ADA positive sample, split it into identical aliquots, freeze them, and run them on different days. You might get S:N values of 5, 7, 10, 6, 9. The biology is unchanged, same antibodies, same complexes, yet the measured ADA response swings twofold simply because the assay was run on a different day. To me, that level of variation is potentially meaningful. If modelers receive a data transfer with a S:N of 5 vs 10, they may arrive at a different conclusion as to what is clinically meaningful response. It is also important to point out that those divergent interpretations come entirely from analytical variation, repeated measures of the same sample.

So if ADA is a biomarker and S:N is the quantity we intend to report, the three-tier approach and cutpoints are gone, but analytical rigor doesn't go away, it shifts. It is worth asking whether we are now responsible for understanding the uncertainty around the S N measurement itself.  If a sample is above assay noise, do we need to re-analyze it multiple times to understand the analytical variation around that sample before drawing PK, PD, or clinical conclusions? That is exactly how other ligand binding biomarker assays are handled: precision is established against the real analyte (endogenous) in the real matrix. Why would ADA be the exception?   Because the analyte mixture changes from time point to time point?  Because it's harder?

If we don't want to go to the level of reporting a CV or a SD from a n=6 of every sample above the assay noise – I think the median of an n=3 would be a far more accurate S:N measurement, on average, than simply going with the S:N we received in the first test.

And just to be clear, the goal here is not to do extra work, but logically that seems to be the direction it is heading. Thoughts?

I wanted to post this here because as the paradigm for ADA assays shifts, something has been gnawing at me. I'm fully on board with moving away from the three tier approach, and nothing I'm raising here is "fixed" by keeping it. But if we're now using language like "ADA are a biomarker," then I think there are analytical implications to that shift that I haven't seen well explained. Thank you to Robert for the recent presentation, it pushed me to try to make my thoughts more coherent.

Before ADA appears, the only analytical variance we can characterize is assay noise because we have no relevant surrogate to measure analytical variability of an ADA response. The mixture is too heterogeneous to represent meaningfully until it exists in a study sample. But once ADA appears, the assay is measuring an analyte, a mixture yes (IgM, IgG, low affinity, high affinity), but that mixture is the analyte.

At that point the meaningful question becomes: how consistently can we measure that analyte?

I've heard the argument that because the analyte mixture may shift from day to day or timepoint to timepoint, that variability should be treated as biological. But that doesn't sit well with me if we are treating ADA assays like biomarker assays. With other biomarkers, we do not assume precision; we measure it using endogenous samples because they contain the analyte in the true matrix. ADA assays seem to be the only place where we go straight from "the signal is above assay noise" to "this might be biologically meaningful," without ever asking whether the measurement itself is stable enough to make that determination.

Consider a simple hypothetical. Take one ADA positive sample, split it into identical aliquots, freeze them, and run them on different days. You might get S:N values of 5, 7, 10, 6, 9. The biology is unchanged, same antibodies, same complexes, yet the measured ADA response swings twofold simply because the assay was run on a different day. To me, that level of variation is potentially meaningful. If modelers receive a data transfer with a S:N of 5 vs 10, they may arrive at a different conclusion as to what is clinically meaningful response. It is also important to point out that those divergent interpretations come entirely from analytical variation, repeated measures of the same sample.

So if ADA is a biomarker and S:N is the quantity we intend to report, the three-tier approach and cutpoints are gone, but analytical rigor doesn't go away, it shifts. It is worth asking whether we are now responsible for understanding the uncertainty around the S N measurement itself.  If a sample is above assay noise, do we need to re-analyze it multiple times to understand the analytical variation around that sample before drawing PK, PD, or clinical conclusions? That is exactly how other ligand binding biomarker assays are handled: precision is established against the real analyte (endogenous) in the real matrix. Why would ADA be the exception?   Because the analyte mixture changes from time point to time point?  Because it's harder?

If we don't want to go to the level of reporting a CV or a SD from a n=6 of every sample above the assay noise – I think the median of an n=3 would be a far more accurate S:N measurement, on average, than simply going with the S:N we received in the first test.

And just to be clear, the goal here is not to do extra work, but logically that seems to be the direction it is heading. Thoughts?