I think that you have made a little mistake; we need some precise definitions.
Consider first-order logic and a "standard" proof system for it; i.e.logical axioms + inference rules (a so-called Hilbert-style proof system; see Elliott Mendelson, Introduction to mathematical logic (4ed - 1997) or Herbert Enderton, A Mathematical Introduction to Logic (2nd - 2001); for natural deduction, thinghs are a little bit different [more rules, no logical axioms], but the basic concepts are the same; see Dirk van Dalen, Logic and Structure (5th ed - 2013) or Ian Chiswell & Wilfrid Hodges, Mathematical Logic (2007)).
Note. Mendelson uses two inference rules : modus ponens and geenralization, while Enderton uses only modus ponens.
We have to start from the definition of deduction form assumptions [see Enderton, page 111] :
A deduction of $\varphi$ from $\Gamma$ is a finite sequence $(\alpha_0, ... \alpha_n)$ of formulas such that $\alpha_n$ is $\varphi$ and for each $k \le n$, either
($a_1$) $\alpha_k$ is in $\Gamma$, or
($a_2$) $\alpha_k$ is in the set $\Lambda$ of logical axioms, or
($b$) $\alpha_k$ is obtained by modus ponens from two earlier formulas in the sequence; that is, for some $i$ and $j$ less than $k$, $\alpha_j$ is $\alpha_i \rightarrow \alpha_k$.
If such a deduction exists, we say that $\varphi$ is deducible from $\Gamma$, or that $\varphi$ is a theorem of $\Gamma$, and we write $\Gamma \vdash \varphi$.
Consider now, your example : "a proof of the theorem represented by $A \rightarrow C$ is to be able to deduce $C$ from $A$".
In this way, we obtain a deduction of $C$ from the assumption $A$ :
$A \vdash C$.
Of course, in the above deduction, we have used also the set $\Lambda$ of (logical) axioms.
Then we apply Deduction Theorem to get :
that if $C$ was deduced from $A$, then $A \rightarrow C$ is provable, i.e. $\vdash A \rightarrow C$.
By soundness, we conclude that $A \rightarrow C$ is a (first-order) valid.
See this post for more details.
Consider now a finite set of sentences $\Gamma = \{ \gamma_1,... , \gamma_n \}$. Then, as you said :
we can join all assumptions into one single formula $G = \gamma_1 \land \gamma_2 \land ... \land \gamma_n$, and simply consider a theorem as a conditional statement $G \rightarrow \varphi$.
For example, consider Group theory [see Mendelson, page 71].
$G$ is the conjunction of the (finite) list of first-order axiom of group theory.
Now, let $\varphi$ be a theorem whatever of group theory; from the consideartions above, we have :
$G \vdash \varphi$,
from which we conclude :
$\vdash G \rightarrow \varphi$.
The last one is a valid formula of first-order logic; that is it is true in every interpretation. This is consistent with mathematical practice: if we consider a structure that is a group, this will obviously satisfy group axioms, i.e. $G$.
Being $\varphi$ a tehorem of group theory, by soundness it will be a logical consequence of $G$, i.e. it will be true in every interpretation that satisfy $G$. In conclusion, $G \rightarrow \varphi$ will be true in every interpreattion, i.e.valid.
But this does not imply that $\varphi$ is valid, because it is not a theorem of first-order logic.
In conclusion,
$G \rightarrow \varphi$ is a logical theorem (of f-o logic) and thus it is valid.
Instead :
$\varphi$ is a theorem of group theory but not necessarily a logical one; thus, it is not valid.
Added - April,23
Consider Enderton's system (I use it for simplicity, because it has modus ponens as single rule of inference).
The axiom system for f-o logic [see page 112] is :
The logical axioms are then all generalizations of wffs of the following forms, where $x$ and $у$ are variables and $\alpha$ and $\beta$ are wffs:
Tautologies;
$\forall x \alpha \rightarrow \alpha(t/x)$, where $t$ is substitutable for $x$ in $\alpha$;
$\forall x (\alpha \rightarrow \beta) \rightarrow (\forall x \alpha \rightarrow \forall x \beta)$;
$\alpha \rightarrow \forall x \alpha$, where $x$ does not occur free in $\alpha$.
In addition, the equality axioms.
Consider the f-o language for number theory [see Enderton, page 70] : an individual constant : $0$, the successor (unary) function : $S$, the two (binary) functions "sum" : $*$ and "product" : $\times$.
With this language we build the following instance of axiom 2 :
$\forall x \lnot (S(x) = 0) \rightarrow \lnot (1 = 0)$;
where we have used as $\alpha$ the formula : $\lnot (S(x) = 0)$ and as $t$ the term : $S(0)$, i.e.$1$.
Being an instance of a first-order axiom, the formula is provable :
$\vdash \forall x \lnot (S(x) = 0) \rightarrow \lnot (1 = 0)$,
i.e.it is a theorem of f-o logic, because we have proved it using (only) f-o logical axioms.
Thus, by soundness is valid (i.e.true in every interpretation).
Consider now the f-o axiom system for number theory [see Enderton, page 203] ; the first one is :
$\forall x \lnot (S(x) = 0)$.
Thus, from the above instance of f-o axiom 2, we have :
$\forall x \lnot (S(x) = 0) \vdash \lnot (1 = 0)$.
In conclusion, $\lnot (1 = 0)$ is a theorem of number theory, because we have proved it (in f-o logic) from the (mathematical) axioms of f-o number theory.
This theorem is not valid; it is a logical consequence of the mathematical axioms of number theory; thus, it will be true in every structure which satisfy f-o number-theory axioms.
