Reports of single base-pair substitutions that cause human genetic disease and that have been located and characterized in an unbiased fashion were collated; 32% of point mutations were CG----TG or CG----CA transitions consistent with a chemical model of mutation via methylation-mediated deamination. This represents a 12-fold higher frequency than that predicted from random expectation, confirming that CG dinucleotides are indeed hotspots of mutation causing human genetic disease. However, since CG also appears hypermutable irrespective of methylation-mediated deamination, a second mechanism may also be involved in generating CG mutations. The spectrum of point mutations occurring outwith CG dinucleotides is also non-random, at both the mono- and dinucleotide, levels. An intrinsic bias in clinical detection was excluded since frequencies of specific amino acid substitutions did not correlate with the 'chemical difference' between the amino acids exchanged. Instead, a strong correlation was observed with the mutational spectrum predicted from the experimentally measured mispairing frequencies of vertebrate DNA polymerases alpha and beta in vitro. This correlation appears to be independent of any difference in the efficiency of enzymatic proofreading/mismatch-repair mechanisms but is consistent with a physical model of mutation through nucleotide misincorporation as a result of transient misalignment of bases at the replication fork. This model is further supported by an observed correlation between dinucleotide mutability and stability, possibly because transient misalignment must be stabilized long enough for misincorporation to occur. Since point mutations in human genes causing genetic disease neither arise by random error nor are independent of their local sequence environment, predictive models may be considered. We present a computer model (MUTPRED) based upon empirical data; it is designed to predict the location of point mutations within gene coding regions causing human genetic disease. The mutational spectrum predicted for the human factor IX gene was shown to resemble closely the observed spectrum of point mutations causing haemophilia B. Further, the model was able to predict successfully the rank order of disease prevalence and/or mutation rates associated with various human autosomal dominant and sex-linked recessive conditions. Although still imperfect, this model nevertheless represents an initial attempt to relate the variable prevalence of human genetic disease to the mutability inherent in the nucleotide sequences of the underlying genes.