Here’s a version with some helper functions split out, using Stream to eliminate intermediate lists, and binary matching to count the G and C characters:

defmodule Gc do
  def gc_content(dataset) do
    {key, gc_percent} =
      dataset
      |> parse_lines()
      |> Stream.map(fn {k, v} -> {k, gc_percent(v)} end)
      |> Enum.max_by(&elem(&1, 1))

    "#{key}\n#{gc_percent}"
  end

  @spec parse_lines(String.t()) :: Enumerable.t()
  def parse_lines(dataset) do
    dataset
    |> String.replace("\n", "")
    |> String.split(">", trim: true)
    |> Stream.map(&String.split_at(&1, 13))
  end

  @spec gc_percent(String.t()) :: float
  def gc_percent(val), do: Float.round(100 * gc_count(val) / String.length(val), 7)

  @spec gc_count(String.t(), integer) :: integer
  def gc_count(val, n \\ 0)
  def gc_count("", n), do: n
  def gc_count("G" <> rest, n), do: gc_count(rest, n + 1)
  def gc_count("C" <> rest, n), do: gc_count(rest, n + 1)
  def gc_count(<<_::utf8>> <> rest, n), do: gc_count(rest, n)
end