deconvolveCa.m

function [c, s, options] = deconvolveCa(y, varargin)
%% infer the most likely discretized spike train underlying an fluorescence trace
%% Solves mutliple formulation of the problem
%  1) FOOPSI,
%       mininize_{c,s} 1/2 * norm(y-c,2)^2 + lambda * norm(s,1)
%       subject to  c>=0, s>=0, s=Gc
%  2) constrained FOOPSI
%       minimize_{c,s} norm(s, q)
%       subject to  norm(y-c,2) <= sn*sqrt(T), c>=0, s>=0, s=Gc
%       where q is either 1 or 0, rendering the problem convex or non-convex.
%  3) hard threshrinkage
%       minimize_{c,s} 1/2 * norm(y-c, 2)^2
%       subjec to c>=0, s=Gc, s=0 or s>=smin
%  4) Nonnegative least square problem (NNLS)
%       min_{s} norm(y - s*h, 2)^2 + lambda * norm(s,1)
%       subject to s>=0

%% inputs:
%   y: T x 1 vector, fluorescence trace
%   varargin: variable input arguments
%       type: string, defines the model of the deconvolution kernel. possible
%           options are:
%           'ar1':  auto-regressive model with order p=1
%           'ar2':  auto-regressive model with order p=2
%           'exp2': the convolution kernel is modeled as the difference of two
%               exponential functions -
%               h(t) = (exp(-t/tau_d) - exp(-t/tau_r)) / (tau_d-tau_r)
%           'kernel':   a vector of the convolution kernel
%       pars: parameters for the specified convolution kernel. it has
%           different shapes for differrent types of the convolution model:
%           'ar1':  scalar
%           'ar2':  2 x 1 vector, [r_1, r_2]
%           'exp2': 2 x 1 vector, [tau_r, tau_d]
%           'kernel': maxISI x 1 vector, the kernel.
%       sn:  scalar, standard deviation of the noise distribution. If no
%           values is give, then sn is estimated from the data based on power
%           spectual density method.
%       b:  fluorescence baseline vlaues. default is 0
%       optimize_pars:  estimate the parameters of the convolution kernel. default: 0
%       optimize_b:     estimate the baseline. default: 0
%       lambda:     penalty parameter
%       method: methods for running deconvolution. {'foopsi',
%       'constrained_foopsi' (default), 'thresholded'},

%% outputs:
%   c: T x 1 vector, denoised trace
%   s: T x 1 vector, deconvolved signal
%   b: fluorescence baseline
%   kernel: struct variable containing the parameters for the selected
%       convolution model
%   lambda: Optimal Lagrange multiplier for noise constraint under L1 penalty
%     """olves the noise constrained sparse nonnegat

%% Authors: Pengcheng Zhou, Carnegie Mellon University, 2016
% ported from the Python implementation from Johannes Friedrich

%% References
% Friedrich J et.al., NIPS 2016, Fast Active Set Method for Online Spike Inference from Calcium Imaging

%% input arguments
y = reshape(y, [], 1);  % reshape the trace as a vector
options = parseinputs(varargin{:});     % parse input arguments
if isempty(y)
    c = []; s = [];
    return;
end
win = options.window;   % length of the convolution kernel
% estimate the noise
if isempty(options.sn)
    options.sn = GetSn(y);
end
% estimate time constant
if isempty(options.pars)
    switch options.type
        case 'ar1'
            try
            options.pars = estimate_time_constant(y, 1, options.sn);
            catch 
                c = y*0;
                s = c; 
                fprintf('fail to deconvolve the trace\n'); 
                return; 
            end
            if length(options.pars)~=1
                c = zeros(size(y)); 
                s = zeros(size(y)); 
                options.pars = 0; 
                return; 
            end
        case 'ar2'
            options.pars = estimate_time_constant(y, 2, options.sn);
            if length(options.pars)~=2
                c = zeros(size(y)); 
                s = zeros(size(y)); 
                options.pars =[0,0]; 
                return; 
            end
        case 'exp2'
            g = estimate_time_constant(y, 2, options.sn);
            options.pars = ar2exp(g);
        case 'kernel'
            g = estimate_time_constant(y, 2, options.sn);
            taus = ar2exp(g);
            options.pars = exp2kernel(taus, options.win);  % convolution kernel
    end
end

%% run deconvolution
c = y;
s = y;
switch lower(options.method)
    case 'foopsi'  %% use FOOPSI
        if strcmpi(options.type, 'ar1')  % AR 1
            [c, s, options.b, options.pars] = foopsi_oasisAR1(y-options.b, options.pars, options.lambda, ...
                options.smin, options.optimize_b, options.optimize_pars, [], options.maxIter);
        elseif strcmpi(options.type, 'ar2') % AR 2
            [c, s, options.b, options.pars] = foopsi_oasisAR2(y-options.b, options.pars, options.lambda, ...
                options.smin);
        elseif strcmpi(options.type, 'exp2')   % difference of two exponential functions
            kernel = exp2kernel(options.pars, options.window);
            [c, s] = onnls(y-options.b, kernel, options.lambda, ...
                options.shift, options.window);
        elseif strcmpi(options.type, 'kernel') % convolution kernel itself
            [c, s] = onnls(y-options.b, options.pars, options.lambda, ...
                options.shift, options.window);
        else
            disp('to be done');
        end
    case 'constrained'
        if strcmpi(options.type, 'ar1')  % AR1
            [c, s, options.b, options.pars, options.lambda] = constrained_oasisAR1(y,...
                options.pars, options.sn, options.optimize_b, options.optimize_pars, ...
                [], options.maxIter);
        else
            [cc, options.b, c1, options.pars, options.sn, s] = constrained_foopsi(y,[],[],options.pars,options.sn, ...
                options.extra_params);
            gd = max(roots([1,-options.pars']));  % decay time constant for initial concentration
            gd_vec = gd.^((0:length(y)-1));
            c = cc(:) + c1*gd_vec';
            options.cin = c1;
        end
    case 'thresholded'  %% Use hard-shrinkage method
        if strcmpi(options.type, 'ar1')
            [c, s, options.b, options.pars, options.smin] = thresholded_oasisAR1(y,...
                options.pars, options.sn, options.optimize_b, options.optimize_pars, ...
                [], options.maxIter, options.thresh_factor, options.p_noise);
            %             if and(options.smin==0, options.optimize_smin) % smin is given
            %                 [c, s, options.b, options.pars, options.smin] = thresholded_oasisAR1(y,...
            %                     options.pars, options.sn, options.optimize_b, options.optimize_pars, ...
            %                     [], options.maxIter, options.thresh_factor);
            %             else
            %                 [c, s] = oasisAR1(y-options.b, options.pars, options.lambda, ...
            %                     options.smin);
            %             end
        elseif strcmpi(options.type, 'ar2')
            [c, s, options.b, options.pars, options.smin] = thresholded_oasisAR2(y,...
                options.pars, options.sn, options.smin, options.optimize_b, options.optimize_pars, ...
                [], options.maxIter, options.thresh_factor);
            %             if and(options.smin==0, options.optimize_smin) % smin is given
            %                 [c, s, options.b, options.pars, options.smin] = thresholded_oasisAR2(y,...
            %                     options.pars, options.sn, options.optimize_b, options.optimize_pars, ...
            %                     [], options.maxIter, options.thresh_factor);
            %             else
            %                 [c, s] = oasisAR2(y-options.b, options.pars, options.lambda, ...
            %                     options.smin);
            %             end
        elseif strcmpi(options.type, 'exp2')   % difference of two exponential functions
            d = options.pars(1);
            r = options.pars(2);
            options.pars = (exp(log(d)*(1:win)) - exp(log(r)*(1:win))) / (d-r); % convolution kernel
            [c, s] = onnls(y-options.b, options.pars, options.lambda, ...
                options.shift, options.window, [], [], [], options.smin);
        elseif strcmpi(options.type, 'kernel') % convolution kernel itself
            [c, s] = onnls(y-options.b, options.pars, options.lambda, ...
                options.shift, options.window, [], [], [], options.smin);
        else
            disp('to be done');
        end
    case 'mcmc'
        SAMP = cont_ca_sampler(y,options.extra_params);
        options.extra_params = SAMP;
        options.mcmc_results = SAMP;
        plot_continuous_samples(SAMP,y);
end

function options=parseinputs(varargin)
%% parse input variables

%% default options
options.type = 'ar1';
options.pars = [];
options.sn = [];
options.b = 0;
options.lambda = 0;
options.optimize_b = false;
options.optimize_pars = false;
options.optimize_smin = false;
options.method = 'constrained';
options.window = 200;
options.shift = 100;
options.smin = 0;
options.maxIter = 10;
options.thresh_factor = 1.0;
options.extra_params = [];
options.p_noise = 0.9999; 

if isempty(varargin)
    return;
elseif isstruct(varargin{1}) && ~isempty(varargin{1})
    tmp_options = varargin{1};
    field_nams = fieldnames(tmp_options);
    for m=1:length(field_nams)
        eval(sprintf('options.%s=tmp_options.%s;', field_nams{m}, field_nams{m}));
    end
    k = 2;
else
    k = 1;
end
%% parse all input arguments
while k<=nargin
    
    switch lower(varargin{k})
        case {'ar1', 'ar2', 'exp2', 'kernel'}
            % convolution kernel type
            options.type = lower(varargin{k});
            if (k<nargin) && (isnumeric(varargin{k+1}))
                options.pars = varargin{k+1};
                k = k + 1;
            end
            k = k + 1;
        case 'pars'
            % parameters for the kernel
            options.pars = varargin{k+1};
            k = k+2;
        case 'sn'
            % noise
            options.sn = varargin{k+1};
            k = k+2;
        case 'b'
            % baseline
            options.b = varargin{k+1};
            k = k+2;
        case 'optimize_b'
            % optimize the baseline
            options.optimize_b = true;
            if (k<nargin) && (islogical(varargin{k+1}))
                options.optimize_b = varargin{k+1};
                k = k + 1;
            end
            k = k+1;
        case 'optimize_pars'
            % optimize the parameters of the convolution kernel
            options.optimize_pars = true;
            if (k<nargin) && (islogical(varargin{k+1}))
                options.optimize_pars = varargin{k+1};
                k = k+1;
            end
            k = k + 1;
            
        case 'optimize_smin'
            % optimize the parameters of the convolution kernel
            options.optimize_smin = true;
            if (k<nargin) && (islogical(varargin{k+1}))
                options.optimize_smin = varargin{k+1};
                k = k+1;
            end
            k = k+1;
        case 'lambda'
            % penalty
            options.lambda = varargin{k+1};
            k = k+2;
        case {'foopsi', 'constrained', 'thresholded', 'mcmc'}
            % method for running the deconvolution
            options.method = lower(varargin{k});
            k = k+1;
            if strcmpi(options.method, 'mcmc') && (k<=length(varargin)) && (~ischar(varargin{k}))
                options.extra_params = varargin{k};
                k = k+1;
            end
        case 'window'
            % maximum length of the kernel
            options.window = varargin{k+1};
            k = k+2;
        case 'shift'
            % number of frames by which to shift window from on run of NNLS
            % to the next
            options.shift = varargin{k+1};
            k = k+2;
        case 'smin'
            % number of frames by which to shift window from on run of NNLS
            % to the next
            options.smin = varargin{k+1};
            k = k+2;
        case 'maxiter'
            % number of frames by which to shift window from on run of NNLS
            % to the next
            options.maxIter = varargin{k+1};
            k = k+2;
        case 'thresh_factor'
            % number of frames by which to shift window from on run of NNLS
            % to the next
            options.thresh_factor = varargin{k+1};
            k = k+2;       
        case 'p_noise'
            % number of frames by which to shift window from on run of NNLS
            % to the next
            options.p_noise = varargin{k+1};
            k = k+2;
        otherwise
            k = k+1;
    end
end

%% correct some wrong inputs
if strcmpi(options.type, 'kernel')
    options.window = numel(options.pars);
end