Validate Module

---

(Note, the results of the validate module within our submitted paper is our very preliminary results. Here, we have extended the validating to all possible pairs related with the pathway. We also try to predict to newly found interacting pairs. )

Yeast Pheromone Pathway

To analyze the results for their utility in the design of new experiments, we compared the predictions by our method to their labels for one specific pathway, the yeast pheromone pathway. The yeast mating factors MATa/a bind to their cognate membrane receptors, Ste2/3, members of the G protein coupled receptor family. Subsequent binding and activation of the G protein induces a MAP kinase signaling pathway via the G protein bg subunit. We get this pathway’s interacting network graph (Figure 1.) from “KEGG PATHWAY Database” (within MAPK signaling pathway - Saccharomyces cerevisiae). The pathway “Pheromone” line relates with 25 proteins. They are listed in the following table 1.

Table1. “Pheromone” related 25 yeast proteins

No.	1	2	3	4	5	6	7	8	9
ORF Name	YPL187W	YGL089C	YFL026W	YDR461W	YNL145W	YKL178C	YHR005C	YOR212W	YJR086W
Gene Name	MF(ALPHA)1	MF(ALPHA)2	STE2	MFA1	MFA2	STE3	GPA1	STE4	STE18
No.	10	11	12	13	14	15	16	17	18
ORF Name	YLR229C	YAL041W	YBR200W	YHL007C	YNL271C	YLR362W	YDR103W	YDL159W	YBL016W
Gene Name	CDC42	CDC24	BEM1	STE20	BNI1	STE11	STE5	STE7	FUS3
No.	19	20	21	22	23	24	25
ORF Name	YNL053W	YPL049C	YDR480W	YHR084W	YJL157C	YMR043W	YCL027W
Gene Name	MSG5	DIG1	DIG2	STE12	FAR1	MCM1	FUS1

FigureS1. “Pheromone” Pathway Network Graph on KeGG

Prediction on all the potential 300 proteins

For these 25 proteins, there are 300 (25*24) potential protein pairs for us to test their interacting probability. We used our method to predict if each these potential pair interacts or not. We also compared these pairs with our small positive label set and the KeGG pair interacting relationships within Figure 1 above. Please see the Excel table containing all 300 pairs’ involved protein ORF name, gene name, our prediction, the small_POS set label, and the KeGG interacting label. Here we also present the performance statistics in Table 2. For these 300 pairs, 11 of them are in labeled as positive pairs by both our pos-small labeled set and the KeGG pathway. For these 11 pairs, we predicted 8 of them as positive correctly. For the other cases, the table2 has clear performance numbers. For example, from the following numbers, we can derive and get that within the 19 (11 + 8 ) KeGG labeled nteracting pairs, our prediction correctly found 12 ( 8+4 ) of them (63.16%).

Table 2. Some statistics about our prediction

	Test Set	Our Predicted Positive	Our Predicted Negative
Whole	300	44 / 300	256 / 300
Labeled POS by both pos_small & KeGG	11 /300	8 / 44	3 / 256
Labeled only by pos_small	31 /300	13 / 44	18 / 256
Labeled only by KeGG	8 / 300	4 / 44	4 / 256
Remaining	251 / 300	19 (need further analysis) / 44	231 / 300

(Note, within the table, “Pos” mean positive interacting pairs. For the protein pairs within 300 that are also in our current data set, we use the subset that does not contain these pairs to train the model and get the final prediction. The threshold is set by a validation set as described within the paper experimental section.)

Analysis for 19 newly predicted proteins

For the table 2, we know that there are 19 our predicted pairs that can not be found within both the small positive set and also the KeGG pathway. The following table 3 lists these remaining 19 pairs. We made further studies for all these pairs (Detailed Comment)

Table 3. The new found 19 proteins pairs: Only 2 predictions are clearly incorrect, but are understandable (lines 1+2). 6 predictions are clearly correct, as confirmed by scientific literature. There are 11 new predictions, all of which are reasonable and would functionally make sense. They form two clusters, one is the possible interaction between STE5 anchor protein with the receptors. The receptor would then make additional interactions due to STE5 anchoring function. The second cluster is a possible interaction between the most downstream component of the signaling cascade and BNI1/BEM1.

No	ORF1	Gene1	ORF2	Gene2	Pair	Verdict	Detailed comment
1	YFL026W	STE2	YNL145W	MFA2*	Description	Wrong, but understandable	Ste2 is the GPCR binding to MATalpha, while Mfa2 is the ligand for Ste3, the MATa receptor*, so this is wrong, but should be very difficult to distinguish automatically.
2	YDR461W	MFA1*	YNL145W	MFA2	Description	Wrong, but understandable	Mfa1 and Mfa2 are two ligands binding to the two GPCRs. They probably do not physically interact with each other, but everything about them should be very similar. Again, all the indirect evidence should point at their interaction.
3	YNL145W	MFA2	YHR005C	GPA1	Description	correct	GPA1 is the G protein. MFA2 is bound to the G protein coupled receptor, and it is the complex between the ligand and the receptor that interactions with the G protein. This prediction is correct.
4	YNL145W	MFA2	YDR103W	STE5**	Description	New prediction	STE5 is a protein that interacts with many other proteins and appears to form an anchor role. Any interactions involving STE5 should be difficult to dissect automatically (see discussion in paper). Further experiments are needed to test if this prediction is correct. It would make sense to anchor members of the signaling pathway via STE5 to the membrane via the receptors (and MFA as a ligand is part of the receptor complex, see below line 6).
5	YKL178C	STE3	YJR086W	STE18	Description	Correct	STE18 is a subunit of the G protein. This prediction is correct.
6	YKL178C	STE3	YDR103W	STE5	Description	New prediction	See line 4. The interaction between the STE5 anchor on the receptors would make sense.
7	YHR005C	GPA1	YDR103W	STE5	Description	New prediction	See lines 4,6,7. If STE5 is anchored to the receptor, then it would also interact with the G protein.
8	YLR229C	CDC42	YDL159W	STE7	Description	New prediction	CDC42 is a small G protein of Rho superfamily. STE7 is a kinase downstream of CDC42 and interacts with STE5. It is possible that these interact, maybe as part of the Ste5 complex.
9	YLR229C	CDC42	YPL049C	DIG1	Description	New prediction	See line 8, DIG1 is further downstream.
10	YAL041W	CDC24	YHL007C	STE20	Description	Correct	Cdc24 is the GDP-GTP exchange factor for Cdc42. Cdc42 interacts with Ste20. These proteins can form a ternary complex.
11	YAL041W	CDC24	YLR362W	STE11	Description	New prediction	Ste11 is part of STE5 complex, and CDC42 is predicted to interact with its direct partners. See line 8 above.
12	YAL041W	CDC24	YDR103W	STE5	Description	New prediction	Directly supports line 8.
13	YAL041W	CDC24	YDL159W	STE7	Description	New prediction	See lines 8-12 above.
14	YBR200W	BEM1	YNL271C	BNI1	Description	Correct	BEM1 is SH3-domain protein that binds CDC42/CDC24 and BNI1 binds CDC42.
15	YBR200W	BEM1	YCL027W	FUS1	Description	New prediction	FUS1 is a cell surface protein that mediates cell fusion. It is one of the most downstream factors in the pathway.
16	YHL007C	STE20	YDL159W	STE7	Description	New prediction	See lines 8-13.
17	YNL271C	BNI1	YCL027W	FUS1	Description	New prediction	See line 15, above. Consistent with line 14: BEM1&BNI1 interact, so if line 15 is correct, then line 17 is correct also.
18	YPL049C	DIG1	YDR480W	DIG2	Description	Correct	see e.g. Kusari et al (2004) J. Cell Biol. 164, 267-277.
19	YDR461W	MFA1	YHR005C	GPA1	Description	Correct	Same argument as line 3.

*MATa = MFA1 (precursor) and MFA2 (lipopeptide hormone) è Ste3 is the receptor

MATalpha = MF(alpha)1 and MF(alpha)2 è Ste2 is the receptor

**STE5 is the anchor protein, described in the discussion of the paper

So from the table 2 and table 3, we got to know that:

¡ For the 19 newly predicted pairs we got

l 2 of them found wrongly predicted

l 6 of them further verified correct predicted by Judith

l The remaining 11 new predictions are reasonable and would functionally make sense. We are currently trying to validate the above new predictions experimentally.

¡ So for correctness, among the 44 interacting protein pairs we predicted

l 31 out of 44 are correct (70.45%)

l 2 out of 44 are incorrect (4.55%)

l 11 out of 44 form new hypothesis that need further lab experimental verification (25%)

¡ So for completeness,

l We do not know the complete interact pair information

l We missed 25 pairs (3 + 18 + 4)

l If assume 42+25 is the total true pair (62.7% recall)

l If assume 31+25 is the total true pair ( 55.4% recall)

Online service for all possible yeast protein pairs

We have used our method on all possible yeast protein pairs to predict if they interact or not. For 6270 yeast proteins, there are 19653315 possible protein pairs overall. Use a default threshold (-0.2), we predict 26301 pairs of them to interact.

The web service would is just published online at: http://www.medstory.net:8081/ppiws/ . People can either search a certain yeast protein pair’ interacting confidence or search a protein’s interacting partners by setting some confidence threshold.