how sarah nexus predicts - lhasa limited
TRANSCRIPT
Yes
YesDisplay
prediction Find
hypotheses
For each hypothesis find kNN over example set k=10
Check if this is an exact
match
Querycompound
No
Fragment structure
Yes
Display prediction
Fragment structure
Find hypotheses
Are all fragments adequately represented
in the training set?
No
No
Yes
Find kNN over whole training
set k=8
Calculate confidence level
confidence x = | Sx |
Display prediction
Calculate confidence level for eachindividual hypothesis based on the kNN
confidence h,x = | Sh,x |
Out ofdomain
Out ofdomain
Are all fragments adequately represented in the training set?
Calculate overall confidence level confidence x = | Sx |
Apply reasoning
No
Assign 100%confidence
Sx overall signal for the query compound x
signal strength of each individual hypothesis given the query xSh,x
How Sarah Nexus Predicts
* Process correct for Sarah Nexus v2.0.1
Yes
YesDisplay
prediction Find
hypotheses
For each hypothesis find kNN over example set k=10
Check if this is an exact
match
Querycompound
No
Fragment structure
Yes
Display prediction
Fragment structure
Find hypotheses
Are all fragments adequately represented
in the training set?
No
No
Yes
Find kNN over whole training
set k=8
1. Calculate the weighted signal
Wi,x = similarity (x, ei)
2. Further moderate the signal to accountfor the average distance of the
examples from the query structure
3. Calculate confidence levelconfidence x = | Sx |
Display prediction
1. Calculate the weighted signal for each hypothesis based on the kNN
Si = -1 if ei is a negative exampleSi = +1 if ei is a positive example
2. Further moderate the signal to accountfor the average distance of the
examples from the query structure
3. Calculate confidence level for each individual hypothesis
confidence h,x = | Sh,x |
Out ofdomain
Out ofdomain
Are all fragments adequately represented in the training set?
Apply reasoning
No
Assign 100%confidence
WSx = ∑k Wi,x × Si i=1
∑k Wi,x i=1
Sx = WSx x∑k Wi,xi=1
k
Wi,x = similarity (x, ei)
WSh,x = ∑k Wi,x × Si i=1
∑k Wi,x i=1
Sh,x = WSh,x x ∑k Wi,xi=1
k
Calculate overall confidence level confidence x = | Sx |
Sx = ∑m Sh,x × confidenceh,xh=1
signal of the nearest ith neighbour
WSh,x weighted signal of each individual hypothesis given the query x
weighting factor Wi,x
signal strength of each individual hypothesis given the query xSh,x
Si
overall signal for the query compound x Sx
number of relevant hypotheses m
k number of nearest neighbours examples in the training setei
weighted signal from nearest neighbours given the query xWSx
How Sarah Nexus Predicts – Detailed
* Process correct for Sarah Nexus v2.0.1
∑m confidenceh,xh=1