The information you provide is too limited to give a good answer. Maybe your geometry is not right. Best start from a similar transition state (TS) from litrature. From the topics I deduce, that you are modelling a TS with the CYP450 reactive center including a transition metal (iron) using DFT. This process involves single electron transfer so you should use unrestricted basis functions. With restricted basis functions (no triplet states involved) there will be no TS. I belive there are TS optimized with DFT in the literature. Best check there.

You must have a powerful machine. Things like this usually require a lot of guess-and-check jobs starting from slightly different geometries so you might try to find a simpler way to model your system. As frustrating as it can get, you still have to go with the qualitative description of your imaginary mode(s) that Gaussview shows you. You haven't found the right one yet, but keep trying! As Denis said, this will definitely require an unrestricted wavefunction. If simplifying your system is not an option then I hope you know what ONIOM layering is... If not you need to look that up.

If you are beyond this already, then the best advice I can give you is to use all your knowledge of orbital symmetries to make the best guesses that you can. For instance, I would expect a proton transfer to water to occur in an orbital approximately 109 degrees to both hydrogen atoms (in a lone pair orbital).

I imagine characterizing a TS like this will also require a lot of post-processing of your checkpoint file(s) when you DO find an optimized structure just to make sure you have found the correct electronic state. At the least, it will probably require a stability analysis and subsequent energy reevaluation at the level of theory that you are using. At the most it could require you to read in manually-defined orbital occupations with guess=permute and/or guess=alter. If you still do not know how to use these, just start with guess=mix. I shudder at the thought of this, but keep trucking on -- that's a cool project!

Thanks guys for these useful information

I know that I'm giving the typical advise, but a good initial geometry is essential. Do a frequency calculation on your initial geometry. Ideally, it should have just one imaginary frequency. That frequency must correspond to what your reaction should you, in your case OH transfer. If it doesn't forget about it as you'll be wasting your time.

As I said, the initial structure should ideally have only one imaginary frequency. If there's a second one but with a very low value, you could carry on as normal. Otherwise, you should seriously consider to remove it. A fast way to do it it to do a constraint optimisation. Freeze the modes you're interested in (distances between OH-metal and OH-substrate?) and optimise the rest. If you don't remove other imaginary frequencies it can happen that during your optimisation can have a value larger than the mode in which you're interested in. In that case, Gaussian will follow that mode and won't arrive to the TS you're interested in. Be aware that happens quite often, even when you have just one initial imaginary frequency. You can check in the output which mode is following in each step of the TS search and how many modes with negative eigenvalues are.

Besides, you might wanna use the option opt(calcall). It calculates the Hessian analytically for each optimisation step. Of course, it much more time demanding but produces a better quality step. That means i) optimisation is achieved with less steps ii) sometimes it makes the difference between finding or not a TS. Besides, reducing the step size in the upgrade of the coordinates (maxstep=20 or 10, Gaussian default is 30) might help in situations where you are close to the TS (low gradients) but still doesn't converge (large displacements).

Small imaginary frequencies are caused (especially for large system) by some rotors (methyl or other). I will try to do a "two" or many step process to avoid them:

1. You have to fixed the position (with somewhat larger extension of bond) OH moving, and optimise all other geom parameters.

2. Let free the moving OH parameters and redone the optimization. (in this cases you can try to decrease the step size of optimization as suggested by José Antonio Gámez in his earlier answer.

This is nice challenge, have good time!

Thanks a lot for these helpful answers Dr. José and Prof. Viskolcz 

Cory Camasta wrote: " For instance, I would expect a proton transfer to water to occur in an orbital approximately 109 degrees to both hydrogen atoms (in a lone pair orbital)." 

An sp3 lone pair orbital?

He also cautioned : "You must have a powerful machine."

Can't this project be done using molecular mechanics which can handle sp3 lone pairs better than Gaussian, and which would need less juice/time to do so?

You would, at least, arrive at an initial structure that you can then "optimize".

At last , I got by changing the geometries of the reactants to a better position