Download presentation
Presentation is loading. Please wait.
Published byAudra Allison Modified over 9 years ago
1
FCM WIZARD IN ACTION SCENARIO: HIV DRUG RESISTANCE PREDICTION Main developers: Gonzalo NÁPOLES and Isel GRAU Supervisors: Elpiniki PAPAGEORGIOU and Koen VANHOOF
2
Scenario description This FCM was conceived for studying the drug resistance of the protease protein when single or multiple mutations take place. The scenario is a classification problem where the instances are protein mutations described by the amino acid on each position and a binary class denoting “Resistance” or “Susceptibility” of the mutation to an HIV drug. Concepts (graph nodes) represent the attributes (descriptors) in the prediction problem: the amino acid contact energy of each protease position associated to drug resistance. Decision concept represents the Resistance class. Dataset source: HIV protease resistance to Indinavir (IDV) See http://hivdb.stanford.edu for further information.http://hivdb.stanford.edu G. NÁPOLES, I. GRAU, R. BELLO, R. GRAU, TWO-STEPS LEARNING OF FUZZY COGNITIVE MAPS FOR PREDICTION AND KNOWLEDGE DISCOVERY ON THE HIV-1 DRUG RESISTANCE, EXPERT SYSTEMS WITH APPLICATIONS 41 (2014) 821–830.
3
Create a new map by using the menu File |New Map. The canvas is immediately cleaned and ready to be used in a new map. Designing the FCM model G. NÁPOLES, I. GRAU, R. BELLO, R. GRAU, TWO-STEPS LEARNING OF FUZZY COGNITIVE MAPS FOR PREDICTION AND KNOWLEDGE DISCOVERY ON THE HIV-1 DRUG RESISTANCE, EXPERT SYSTEMS WITH APPLICATIONS 41 (2014) 821–830.
4
Design a FCM Add concepts and causal links between them by drag and dropping from the source concept to the target one. Concepts represent protein positions and R denotes the resistance class. For this scenario experts do not know the causal influence, therefore we keep a random value for now. G. NÁPOLES, I. GRAU, R. BELLO, R. GRAU, TWO-STEPS LEARNING OF FUZZY COGNITIVE MAPS FOR PREDICTION AND KNOWLEDGE DISCOVERY ON THE HIV-1 DRUG RESISTANCE, EXPERT SYSTEMS WITH APPLICATIONS 41 (2014) 821–830.
5
Take into account the cut-off for the decision concept R in order to perform a classification. In this scenario the cut-off for resistance is taken from literature and normalized respect to the maximum resistance value in historical data (e.g. for IDV the cut-off is 0.006). Design a FCM Customize concepts and define the ranges for the classes using the cut-off for the decision concept. G. NÁPOLES, I. GRAU, R. BELLO, R. GRAU, TWO-STEPS LEARNING OF FUZZY COGNITIVE MAPS FOR PREDICTION AND KNOWLEDGE DISCOVERY ON THE HIV-1 DRUG RESISTANCE, EXPERT SYSTEMS WITH APPLICATIONS 41 (2014) 821–830.
6
In this scenario the model is computed from a dataset, and therefore we do not know about the causal relations among protein positions. In such cases a learning algorithm for learning the matrix may be applied (Run | Learning Methods). Learning of causal weights G. NÁPOLES, I. GRAU, R. BELLO, R. GRAU, TWO-STEPS LEARNING OF FUZZY COGNITIVE MAPS FOR PREDICTION AND KNOWLEDGE DISCOVERY ON THE HIV-1 DRUG RESISTANCE, EXPERT SYSTEMS WITH APPLICATIONS 41 (2014) 821–830.
7
The tool includes unsupervised and supervised learning algorithms. Set up the parameters for the chosen algorithm and browse for the historical dataset source (in ARFF format). Learning of causal weights G. NÁPOLES, I. GRAU, R. BELLO, R. GRAU, TWO-STEPS LEARNING OF FUZZY COGNITIVE MAPS FOR PREDICTION AND KNOWLEDGE DISCOVERY ON THE HIV-1 DRUG RESISTANCE, EXPERT SYSTEMS WITH APPLICATIONS 41 (2014) 821–830.
8
During the learning progress the statists (e.g. the error curve) are updated. The right panel shows additional statistics related to the knowledge base and the learning process. Learning of causal weights It can be noticed that 98% of instances were correctly classified in the evaluation so far! Now we are able to perform simulation on the FCM and classify new instances of the problem! G. NÁPOLES, I. GRAU, R. BELLO, R. GRAU, TWO-STEPS LEARNING OF FUZZY COGNITIVE MAPS FOR PREDICTION AND KNOWLEDGE DISCOVERY ON THE HIV-1 DRUG RESISTANCE, EXPERT SYSTEMS WITH APPLICATIONS 41 (2014) 821–830.
9
Set initial values for concepts and trigger the inference Run | Run Inference Process. Next a plotter will be showed illustrating the inference process in detail. FCM Inference G. NÁPOLES, I. GRAU, R. BELLO, R. GRAU, TWO-STEPS LEARNING OF FUZZY COGNITIVE MAPS FOR PREDICTION AND KNOWLEDGE DISCOVERY ON THE HIV-1 DRUG RESISTANCE, EXPERT SYSTEMS WITH APPLICATIONS 41 (2014) 821–830.
10
Moreover, it is possible to exploit the learned model by classifying new instances (i.e. mutations) by using the predefined decision classes (i.e. Susceptible and Resistant). Classify a new instance G. NÁPOLES, I. GRAU, R. BELLO, R. GRAU, TWO-STEPS LEARNING OF FUZZY COGNITIVE MAPS FOR PREDICTION AND KNOWLEDGE DISCOVERY ON THE HIV-1 DRUG RESISTANCE, EXPERT SYSTEMS WITH APPLICATIONS 41 (2014) 821–830.
11
Set the values for each attribute concept (@data) of the new instance. You can see in the graphic the convergence behavior of the values for each concept during the simulation. Classify a new instance However, since the map was designed from historical data, some concepts can be still superfluous or contradictory. These are the values of concepts after inference process. The label-value ‘0’ denotes that this mutation’s class value is “Susceptible”. G. NÁPOLES, I. GRAU, R. BELLO, R. GRAU, TWO-STEPS LEARNING OF FUZZY COGNITIVE MAPS FOR PREDICTION AND KNOWLEDGE DISCOVERY ON THE HIV-1 DRUG RESISTANCE, EXPERT SYSTEMS WITH APPLICATIONS 41 (2014) 821–830.
12
The algorithms to optimize the map topology reduce unnecessary concepts and improve the global interpretability. They are based on population-based metaheuristics. Topology optimization G. NÁPOLES, I. GRAU, R. BELLO, R. GRAU, TWO-STEPS LEARNING OF FUZZY COGNITIVE MAPS FOR PREDICTION AND KNOWLEDGE DISCOVERY ON THE HIV-1 DRUG RESISTANCE, EXPERT SYSTEMS WITH APPLICATIONS 41 (2014) 821–830.
13
First the expert needs to select the proper search method (optimizer) and next configure the specific parameters. Default parameters have been carefully studied in the literature. Topology optimization G. NÁPOLES, I. GRAU, R. BELLO, R. GRAU, TWO-STEPS LEARNING OF FUZZY COGNITIVE MAPS FOR PREDICTION AND KNOWLEDGE DISCOVERY ON THE HIV-1 DRUG RESISTANCE, EXPERT SYSTEMS WITH APPLICATIONS 41 (2014) 821–830.
14
The runtime plotter shows the error evaluation in blue. Meanwhile the topology is updated on the canvas showing the reduced FCM-based system, which normally is more interpretable. Topology optimization After optimization we obtain 7 concepts from the original 18 (41%), representing the same scenario without loosing accuracy in the classification! G. NÁPOLES, I. GRAU, R. BELLO, R. GRAU, TWO-STEPS LEARNING OF FUZZY COGNITIVE MAPS FOR PREDICTION AND KNOWLEDGE DISCOVERY ON THE HIV-1 DRUG RESISTANCE, EXPERT SYSTEMS WITH APPLICATIONS 41 (2014) 821–830.
15
Analyzing the convergence of the resistance concept for each instance in the dataset we observe that the inference of the map is not stable, and the results could be compromised. Visualize system convergence This option Run | Convergence Plotter allows to examine the stability. We can improve the map convergence by learning the correct parameter for the transformation function on each concept, without affecting the causal weights and the classification ability of the FCM learned before. G. NÁPOLES, I. GRAU, R. BELLO, R. GRAU, TWO-STEPS LEARNING OF FUZZY COGNITIVE MAPS FOR PREDICTION AND KNOWLEDGE DISCOVERY ON THE HIV-1 DRUG RESISTANCE, EXPERT SYSTEMS WITH APPLICATIONS 41 (2014) 821–830.
16
It should be remarked that this algorithm is only useful for sigmoid-based FCM since it adjusts the slope of each transfer function to improve the convergence. Improve system convergence G. NÁPOLES, I. GRAU, R. BELLO, R. GRAU, TWO-STEPS LEARNING OF FUZZY COGNITIVE MAPS FOR PREDICTION AND KNOWLEDGE DISCOVERY ON THE HIV-1 DRUG RESISTANCE, EXPERT SYSTEMS WITH APPLICATIONS 41 (2014) 821–830.
17
First the expert needs to select the proper search method (Swarm Intelligence based algorithms) and next configure the parameters, although default values are provided. Improve system convergence G. NÁPOLES, I. GRAU, R. BELLO, R. GRAU, TWO-STEPS LEARNING OF FUZZY COGNITIVE MAPS FOR PREDICTION AND KNOWLEDGE DISCOVERY ON THE HIV-1 DRUG RESISTANCE, EXPERT SYSTEMS WITH APPLICATIONS 41 (2014) 821–830.
18
After system convergence have been improved, you can check again the stability of the modeled system, for each instance stored in the training dataset. Improve system convergence G. NÁPOLES, I. GRAU, R. BELLO, R. GRAU, TWO-STEPS LEARNING OF FUZZY COGNITIVE MAPS FOR PREDICTION AND KNOWLEDGE DISCOVERY ON THE HIV-1 DRUG RESISTANCE, EXPERT SYSTEMS WITH APPLICATIONS 41 (2014) 821–830.
19
FCM WIZARD IN ACTION SCENARIO: HIV DRUG RESISTANCE Main developers: Gonzalo NÁPOLES and Isel GRAU Supervisors: Elpiniki PAPAGEORGIOU and Koen VANHOOF
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.