2D-TPP experiments

2D-TPP experiments for profiling PCI-34051 and BRD-3811 were performed as described²: HL-60 (DSMZ, ACC-3) cells were grown in Iscove’s modified Dulbecco’s medium supplied with 10% fetal bovine serum (FBS). Cells were treated with a concentration range (0, 0.04, 0.29, 2, 10 μM) of PCI-34051 (Selleckchem) or BRD-3811 (synthesized in-house²⁸ with >99% purity as determined by HPLC-UV254 nm) for 90 min at 37 °C, 5% CO₂. The samples from each treatment concentration were split into 12 portions, which were then heated each at a different temperature (42–63.9 °C) for 3 min and then incubated at room temperature for 3 min. Next, 30 μl of ice-cold phosphate-buffered saline (PBS) (2.67 mM KCl, 1.5 mM KH₂PO₄, 137 mM NaCl, and 8.1 mM NaH₂PO₄ pH 7.4) were supplemented with 0.67% NP-40 and protease inhibitors were added to the samples. Subsequently, cells were frozen in liquid nitrogen for 1 min, briefly thawed in a metal block at 25 °C, and then placed on ice and resuspended by pipetting. Samples were then incubated with benzonase for 1 h at 4 °C, followed by centrifugation at 100,000 × g for 20 min at 4 °C. Then, 30 μl supernatant were transferred into a new tube and were subjected to gel electrophoresis and sample preparation for MS analysis.

The 2D-TPP experiment to assess GTP-binding proteins was performed using gel-filtered lysate as described². In short, Jurkat E6.1 cells (ATCC, TIB-152) were cultured in RPMI (GIBCO) medium supplemented with 10% heat-inactivated FBS. The cells were collected and washed with PBS. The cell pellet was resuspended in lysis buffer (PBS containing protease inhibitors and 1.5 mM MgCl₂) equal to ten times the volume of the cell pellet. The cell suspension was lysed by mechanical disruption using a Dounce homogenizer (20 strokes) and treated with benzonase (25 U/ml) for 60 min at 4 °C on a shaking platform. The lysate was ultracentrifuged at 100,000 × g, 4 °C for 30 min. The supernatant was collected and desalted using PD-10 column (GE Healthcare). The protein concentration of the eluted lysate was measured using Bradford assay. The protein concentration of the lysate was maintained at 2 mg/ml for the assay. The lysate was treated using a concentration range of GTP (0, 0.001, 0.01, 0.1, 0.5 mM) for 10 min at room temperature. The samples from each GTP concentration were split into 12 portions, which were then heated each at a temperature (42–63.9 °C) for 3 min. Post-heat treatment, the protein aggregates were removed using ultracentrifugation at 100,000 × g, 4 °C for 20 min. Subsequently, the supernatants were processed as described above.

Protein identification and quantification

Raw MS data were processed with Isobarquant¹¹ and searched with Mascot 2.4 (Matrix Science) against the human proteome (FASTA file downloaded from Uniprot, ProteomeID: UP000005640) extended by known contaminants and reversed protein sequences (search parameters: trypsin; missed cleavages 3; peptide tolerance 10 p.p.m.; MS/MS tolerance 0.02 Da; fixed modifications were carbamidomethyl on cysteines and TMT10-plex on lysine; variable modifications included acetylation on protein N terminus, oxidation of methionine, and TMT10-plex on peptide N termini). Protein FDR was calculated using the picked approach²⁹.

Reporter ion spectra of unique peptides were summarized to the protein level to obtain the quantification s_i,u for protein i measured in condition u = (j, k), i.e., at temperature j and concentration k. Isobarquant additionally computes robust estimates of fold change r_i,u for each protein i in condition u relative to control condition \(u^{\prime}\) using a bootstrap approach. We used these to obtain per-condition log2 signal intensities computed as \({y}_{i,u}={\mathrm{log}\,}_{2}(({r}_{i,u}/{\sum }_{l}{r}_{i,l}){\sum }_{l}{s}_{i,l})\). This is one particular choice of protein quantification; we expect that our method can be used equivalently with input from other quantification methods.

In the resulting abundance table Y = (y_i,u), entries for which the value r_i,u was obtained by not more than one peptide were marked as unreliable (i.e., set to not available (NA) in the software). For each protein i we computed the total number of non-NA measurements p_i and only retained proteins with p_i ≥ 20 for subsequent analysis. In other words, proteins had to be quantified at least at four different temperatures and five different ligand concentrations each to be included in our analysis.

The MS experiment comprising the temperatures 54 and 56.1 °C was excluded from the analysis of the PCI-34051 dataset as we noticed that it contained unexpectedly high noise levels. In particular the relative reporter ion intensities at 54 °C showed about ten times higher variance than all other temperatures, likely due to a drop in instrument performance during the time this sample was measured.

Moreover, in the PCI-34051 and BRD-3811 datasets, we noted that measured profiles of some proteins appeared to have been affected by carry-over from previous experiments. These profiles exhibited a characteristic pattern as depicted in Supplementary Fig. 4c in which apparent stabilization of these proteins was observed only in half of the TMT channels corresponding to every other temperature. These proteins were filtered out by manual inspection.

Data pre-processing of public datasets

The panobinostat and JQ1 datasets were downloaded from the publisher websites as spreadsheets provided as supplementary data together with the publications^2,23. Abundance tables Y were computed and filtered as described above.

Model description

Two nested models were fitted to the abundance values of each protein i at temperature j and ligand concentration k. The null model is:

Here, the base intensity level at temperature j is \({\beta }_{i,j}^{(0)}\), and \({\epsilon }_{i,j,k}^{(0)}\) is a residual noise term. The alternative model is:

Here, \({\beta }_{i,j}^{(1)}\) is again the base intensity level at temperature j, the parameter δ_i describes the maximal absolute stabilization across the temperature range, α_i,j ∈ [0, 1] indicates how much of the maximal stabilization occurs at temperature j and κ_i is a common slope factor fitted across all temperatures. Finally, ζ_i(T_j) is the concentration of the half-maximal stabilization (i.e., pEC50), with \({\zeta }_{i}({T}_{j})={\zeta }_{i}^{0}+{a}_{i}T\), where a_i is a slope representing a linear temperature-dependent decay or increase of the inflection point, and \({\zeta }_{i}^{0}\) is the intercept of the linear model. Again, \({\epsilon }_{i,j,k}^{(1)}\) is a residual noise term. Both models were fit by minimizing the sum of squared residuals \({{\rm{RSS}}}_{i}^{(0)}={\sum }_{j}{\sum }_{k}{({\epsilon }_{i,j,k}^{(0)})}^{2}\) and \({{\rm{RSS}}}_{i}^{(0)}={\sum }_{j}{\sum }_{k}{({\epsilon }_{i,j,k}^{(1)})}^{2}\) using the L-BFGS-B algorithm³⁰ through R’s optim function.

The start values for the parameter and \({\beta }_{i,j}^{(0)} \,\, {\mathrm{and}}\,\, {\beta }_{i,j}^{(1)}\) in the iterative fit of the respective models were initialized with the mean abundance \({\bar{y}}_{i,j}\) of protein i at temperature j; α_i,j was initialized as α_i,j = 0 for all i and j; δ_i was set to the maximal difference between abundance values within a temperature for protein i; κ_i was initialized as the slope estimated by a linear model across temperatures; \({\zeta }_{i}^{0}\) was set to the mean log₁₀ drug concentration used; and a_i was set to 0. The two fitted models can be compared using the F-statistic:

with the degrees of freedom d₁ = ν₁ − ν₀ and d₂ = p_i − ν_i, where p_i is the number of observations for protein i that were fitted, and ν₀ and ν₁ are the number of parameters of the null and alternative model, respectively.

For inference, we used an empirical Bayes moderated version of (3), as implemented in the squeezeVar function in the R/Bioconductor package limma³¹. squeezeVar uses the observed variances \({s}_{i}^{2}={{\rm{RSS}}}_{i}^{(1)}/{d}_{2}\) to identify a common value \({s}_{0}^{2}\) and shrinks each \({s}_{i}^{2}\) towards that value. The motivation for such moderation is to accept a small cost in increased bias for a large gain of increased precision. To do so, squeezeVar assumes that the true \({\sigma }_{i}^{2}\) are drawn from a scaled inverse χ² distribution with parameter \({s}_{0}^{2}\):

Using the assumption that the residuals follow a normal distribution, Bayes’ theorem, and the scaled inverse χ² prior, it can be shown²⁰ that the expected value of the posterior of \({\widetilde{s}}_{i}^{2}\) is

Here, the hyperparameters \({s}_{0}^{2}\) and d₀ are estimated by fitting a scaled F-distribution with \({s}^{2} \sim {s}_{0}^{2}{F}_{d,{d}_{0}}\). Details are described by Smyth et al.²⁰. Thus, we computed moderated F-statistics with

FDR estimation

To estimate the FDR associated with a given threshold θ for the F-statistic obtained for a protein i with m_in_i observations, we adapted the bootstrap approach of Storey et al.¹⁹ as follows. To generate a null distribution the following was repeated B times: (i) resample with replacement the residuals \({\epsilon }_{i,w}^{1}\) obtained from the alternative model fit for protein i in MS experiment w to obtain \({\epsilon }_{i,w}^{1* }\) and add them back to the corresponding fitted estimates of the null model to obtain \({y}_{i,w}^{* }={\mu }_{i,t}^{0}+{\epsilon }_{i,w}^{1* }\). (ii) Fit null and alternative models to \({y}_{i,w}^{* }\) and compute the moderated F-statistic \({\widetilde{F}}^{0b}\). An FDR was then computed by partitioning the set of proteins {1, . . . , P} into groups of proteins with similar number D(p) of measurements, e.g., \(\gamma (p)=\lfloor \frac{D(p)}{10}+\frac{1}{2}\rfloor\) and then

The proportion of true null events \({\hat{\pi }}_{0g}\) in the dataset of proteins in group g was estimated by:

In the case of the standard DLPTP approach, the same procedure as above was performed using non-moderated F-statistics.

Fluorometric aminopeptidase assay

LAP3 activity was determined using the Leucine Aminopeptidase Activity Assay Kit (Abcam, ab124627) and recombinant LAP3 (Origene, NM_015907). Recombinant LAP3 enzyme was dissolved in the kit assay buffer and incubated for 10 minutes at room temperature with vehicle (dimethyl sulfoxide) or 100 μM of either PCI-34051 or BRD-3811. All other assay steps were performed as described in the kit. Fluorescent signal (Ex/Em = 368/460 nm) was detected over 55 min.

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.